On 1 May 2026, SISTRIX published 1,548,213 snapshots of AI answers taken over 17 weeks. ChatGPT Search replaced 74% of its cited domains every single week. Google’s AI Mode replaced 56%. For the median ChatGPT prompt, not one domain survived all 17 weeks.

That is a flaky test. Most AI visibility programmes are reporting it to the board as a result.

Here is the argument. Brand mention and citation are sold as one metric and they are two different instruments. A mention is the model saying your name in the answer. It tracks reputation, it is built over years by other people writing about you, and it moves slowly. A citation is the model linking to your page as a source. It tracks retrieval, it is decided fresh on every run, and it resets weekly. Most programmes are priced on the first and reported on the second, which means they are mispriced in both directions at once.

Two instruments, one dashboard

Start with the definitions, because the vendors blur them.

A mention happens inside the generated text. The model writes “the usual shortlist here is X, Y and Z” and your brand is one of the letters. Nothing is linked. Nothing is clicked. The buyer simply hears your name in the same sentence as the category.

A citation happens in the source list. The model retrieves pages to ground its answer and shows some of them as links. Your page is one of the links. The answer itself may never name you.

These sound like two views of the same thing. The data says otherwise. On 9 June 2026, Semrush published a study run with Kevin Indig’s Growth Memo across 3,981 domain appearances, 115 prompts and 14 countries, on ChatGPT, Gemini, Google AI Overviews and AI Mode. They split every appearance into three buckets.

61.7% were cited but never mentioned. The brand’s page was a source and the brand’s name was absent from the answer. Semrush calls these ghost citations.

25.1% were mentioned but never cited. The name was in the answer and no page of theirs was linked.

13.2% were both.

So in the combined set, only about one appearance in eight did the thing every AI visibility pitch deck implies is normal: name the brand and link to it. The other seven did one or the other. If your dashboard rolls those into a single “visibility” figure, it is adding a reputation number to a retrieval number and reporting the sum as if it meant something.

Why they move at different speeds

The mechanism is not mysterious once you look at where each one comes from.

A mention comes out of what the model already believes about a category, reinforced by what it reads at query time. That belief is formed by volume of third-party writing over a long period: reviews, comparisons, forum threads, analyst notes, press. Ahrefs studied 75,000 brands in its AI Overview analysis, published May 2025 and updated April 2026. The strongest correlate of being mentioned was branded web mentions across the web, at 0.664. Backlinks came in at 0.218. Domain Rating at 0.326. The number of pages on your own site, the thing most content budgets buy, sat at 0.17.

The distribution is steeper than the correlation. Brands in the top quarter for web mentions averaged 169 AI Overview mentions. The next quarter down averaged 14. The bottom half averaged between zero and three. That is not a gradient you climb in a quarter. It is a stock that compounds, the way a reputation does, and the brands at the top got there by being written about for years.

A citation comes out of retrieval. The engine fans out the question into sub-queries, fetches candidate pages, and picks a handful that best support the sentence it is about to write. Change the phrasing, the user’s location, the date, the model version or simply run it again, and the candidate set changes. AirOps tracked more than 45,000 citations and found only about 30% of brands stayed visible from one run to the next. Only one in five held visibility from the first run to the fifth. About 57% of those that dropped out came back within a couple of runs. That is not a brand gaining or losing ground. That is sampling noise.

SparkToro and Gumshoe ran the same experiment from the buyer’s side. Rand Fishkin and Patrick O’Donnell had 600 volunteers put 12 identical prompts through ChatGPT, Claude and Google’s AI, nearly 3,000 runs in total. The same list of recommended brands came back less than once in 100 runs. The same list in the same order came back about once in 1,000. And yet some brands appeared in 60% to 90% of answers for a given intent, regardless of where they ranked. Their conclusion was the right one: repeat presence means something, exact position does not.

Put those together and you have two instruments. One is a slow reputation gauge that should be read quarterly. The other is a fast retrieval sampler that should be read as a distribution, never as a point. Most dashboards read both weekly, on one line, in one colour.

Case one: the same brand, opposite numbers, on two engines

The Semrush data contains a result that should end the single-score argument on its own.

ChatGPT cited brands in 87% of their appearances and mentioned them in only 20.7%. Gemini did the reverse. It mentioned brands in 83.7% of appearances and cited them in 21.4%.

Now think about what a brand sees when it buys a tool that reports citation share. On ChatGPT it looks excellent. Its pages are being used as sources almost every time. On Gemini it looks like it barely exists. Its pages are rarely linked.

But the buyer reading Gemini is hearing the brand’s name four times out of five. The buyer reading ChatGPT mostly is not. The citation dashboard has the commercial reality upside down. The engine where the brand is being recommended aloud looks like the problem, and the engine where its content is being used anonymously to support somebody else’s answer looks like the win.

Flip the tool to report mentions and the picture inverts. Neither view is wrong. They are measuring different things, and a programme that picks one and calls it visibility has chosen which half of the truth to fund. Ask which half before you sign.

Case two: the weekly report that is mostly noise

The second case is in the volatility numbers, and it is the one that costs the most money.

The SISTRIX data covers Google AI Overviews, AI Mode and ChatGPT Search across six countries from 17 December 2025 to 8 April 2026, 82,619 prompts in all. The three platforms behave like three different climates.

AI Overviews are close to classic search. Weekly drift was around 5%, and 53% of prompts showed no change in sources across the whole 17 weeks. Answers carry about 11 domains and about 8 of them stay put.

AI Mode churns 56% of its domains a week, citing 14 to 16 per answer. It does keep a stable core of one to five domains for about 86.5% of prompts. Everything around that core is turnover.

ChatGPT Search churns 74% a week and cites only three or four domains per answer. Fewer slots, faster turnover.

BrightEdge, measuring citation share separately, found the same shape from another angle. Domains cited frequently moved about 0.7% a week. Domains cited sporadically swung 50% or more, a gap of roughly 70 times. Their threshold for stability was about 50 citations. Below it, weekly volatility ran around 50%. Above it, around 8%.

Here is what that does to a monthly report. A mid-sized B2B brand is almost never in the frequently cited group. It sits well below 50 citations on most engines. Its weekly citation share will swing by half on its own, with no change to anything the brand did. A report that says “citations up 38% this month” is, for that brand on that engine, mostly a report about the dice.

And the dice get acted on. Somebody rewrites a page because citations fell. Somebody gets credit because they rose. Somebody renews a retainer on a good month and cancels it on a bad one. Each of those decisions is a reaction to a flaky test.

The one stable line in the SISTRIX data is worth noting. For brand queries in AI Mode, the brand’s own domain appeared in all 17 weeks 43% of the time and in at least 80% of weeks 66% of the time. The domains co-cited alongside it changed about 70% a week. When somebody asks about you by name, your site is reliably there. When somebody asks about the category, which is where new demand comes from, the citation slots rotate constantly. The stable number is the one that matters least for growth.

Case three: the reputation stock nobody budgets for

The third case runs the other way. It is where the slow instrument gets judged on the fast one’s timetable and loses.

Look again at the Ahrefs quartiles. The bottom half of brands by web mentions get between zero and three AI Overview mentions on average. The top quarter get 169. The only way across that gap is to be written about, at volume, by people who are not you: customers on review sites, practitioners in forums, journalists, analysts, partners, comparison articles. Seer Interactive’s study of 10,000 questions run through GPT-4o found the same story through a different lens: page-one Google rankings correlated with brand mentions at about 0.65, while backlinks were weak or neutral. Mentions follow the brands the open web already treats as the answer.

That kind of work is PR, analyst relations, review generation and customer marketing. Its payback period is measured in quarters, often years. It compounds. It is also almost invisible on a weekly citation dashboard, because a new review or a trade press piece does not reliably turn into a link from ChatGPT next Tuesday. It turns, slowly, into the model saying your name.

So the reputation programme gets reported on a retrieval chart. Three months in, the chart is flat or noisy. The CFO asks what the money bought. Nobody can point at a line that moved, because the line they are looking at was never going to move from that work. The programme is cut just before the stock would have started to show.

Meanwhile the citation programme, the page restructuring and freshness and schema work, gets priced like reputation. Tools are sold per prompt tracked: Profound’s self-serve Starter plan, before it was withdrawn in September 2026, was 50 ChatGPT prompts for $99 a month. Agencies sell annual retainers on top. Both are priced as if they buy a durable asset. What they mostly buy, below the 50 citation threshold, is a better position in a lottery that redraws weekly. That work is real and worth doing. It is worth doing at the price of maintenance, not the price of brand.

That is the mispricing in both directions. The slow asset is killed on a fast clock. The fast activity is paid for on a slow contract.

How I read a number like this

Every reporting layer I have built for a private equity owner separates two kinds of number: stocks and flows. Cash on the balance sheet is a stock. Weekly receipts are a flow. You never put them on the same axis, because the scale and the noise are different, and somebody will eventually react to the wrong one.

AI visibility needs the same split. Mention rate is a stock. You measure it with a large, fixed prompt set, many runs per prompt, and you read it quarterly. Citation share is a flow. You measure it per engine, as a rolling distribution, and you read the band, not the point. The moment those two go on one line, you have built a dashboard that will tell you a reputation programme failed and a formatting tweak succeeded, in the same month, for reasons that have nothing to do with either.

Where this breaks

Three honest objections.

First, citations send traffic and mentions do not. A ghost citation still puts a clickable link in front of a buyer, and a mention without a link sends nobody anywhere. That is true, and for a business that lives on referral clicks the citation flow matters more than I have given it credit for. But for a considered B2B purchase the click is rarely the point. The buyer who hears your name in an answer and then searches for you directly will show up as branded search, not as an AI referral. The mention’s value lands somewhere the citation dashboard does not look.

Second, the two are not independent. AirOps found brands that earned both a citation and a mention were 40% more likely to resurface across runs than citation-only brands. The reputation stock clearly steadies the retrieval flow. I agree, and that is the strongest argument for funding the slow work. It is not an argument for reporting them as one number. Correlated instruments still need separate gauges, or you cannot tell which one moved.

Third, not every engine is weather. Google AI Overviews drifted about 5% a week in the SISTRIX data, with more than half of prompts unchanged across four months. On that surface a citation behaves much more like a classic ranking, and reading it monthly is fine. The weather problem is specific to AI Mode, ChatGPT and Perplexity. That is a real limit on my argument, and it is also the reason a single cross-engine score is worse than useless: it averages a stable climate with two unstable ones and hides which is which.

What I would do on Monday

1. Split the dashboard. Report mention rate and citation share as two numbers, per engine, never blended. If your tool cannot separate them, that is the first finding.

2. Ask your vendor for the denominator. How many prompts, how many runs per prompt, which engines, which countries. SparkToro’s data says a single run per prompt tells you almost nothing. If the answer is one run a week, treat every movement as noise until proven otherwise.

3. Put a band on the citation line. Plot an eight-week rolling share with its normal range shaded. Any week inside the band gets no commentary and no action. Only a break outside the band is news.

4. Move the mention budget to a longer clock. Give PR, review generation and analyst work a two to four quarter target on mention rate, owned by whoever owns brand, and stop reviewing it monthly against citations.

5. Count how many of your domains clear about 50 citations on each engine. That is your stable footprint. Everything below it is weather, and it should be priced as maintenance, not as an asset.

Back to the flaky test. Nobody on a serious engineering team pages a director because a known flaky test went red overnight. They quarantine it, fix the harness, and find a measurement they can trust. Most marketing teams are still paging the director. Quarantine the weekly citation number, measure the slow thing on its own clock, and the board pack will finally show which work is building the brand.

The name in the answer is the asset. The link is the weather report.

If you want a second pair of eyes on how your AI visibility is being measured and priced, send me your last three reports and put time in my diary.

That detail sits in the New York Times profile of 2 April 2026, which verified that his telehealth company, Medvi, did $401 million in sales in 2025 and $65 million in net profit. He started it from his house in Los Angeles in September 2024 with $20,000. His only employee is his brother.

Six weeks before that profile ran, on 20 February 2026, the FDA issued warning letter 721455. It was addressed to Medvi LLC.

The lasagna test checks whether the bot stays on topic. The other test is the one nobody runs: when something goes wrong, whose name is on the envelope.

The thesis

The one-person company did not remove the org chart. It rented it. AI made renting cheap enough that a founder can run a nine figure business through partners, contractors and agents, so headcount now measures payroll, not the organisation. Every consequence of the work still comes home to the brand. Headcount has become the least informative number in a small company, and the number that replaced it is one almost nobody writes down: the functions you answer for that somebody else performs.

What actually changed

Here is the Medvi org chart as the Times describes it. CareValidate and OpenLoop Health supply the doctors, the pharmacies, shipping, compliance and the customer service calls. Two engineers are on contract. A law firm replaced LegalZoom. An accounting firm replaced the AI accounting tools. Media agencies buy the ads. ChatGPT, Claude and Grok built the website, Midjourney and Runway make the images, ElevenLabs does the voice. The brother, hired in April, filters communications. Gallagher himself told the Times it is not technically a one-person company.

That is not a small organisation. It is a large one with a short payroll.

What AI changed is the coordination cost. Renting six functions used to mean six account managers, six contracts and a full time operations hire to hold them together. Gallagher built custom agents so his software systems talk to each other. The glue that used to need people became code, so a founder can rent almost everything and still run the machine from a laptop.

What AI did not change is where the consequences land. The FDA letter cited the website for presenting compounded drugs under the Medvi name, which implied Medvi compounds them, and for the line “Same active ingredient as Wegovy and Ozempic”. A partner pharmacy does the compounding. The letter still went to the brand. Regulators, customers and courts do not read your vendor list. They read your logo.

I have written code for 30 years and I now build software for private equity. When a small company tells me its headcount, I ask for a second list before I read anything else: every function a regulator, a customer or a court would hold the company responsible for, and the name of the person or firm that actually performs it. The gap between those two lists is the real size of the company.

Three cases

The supplier you share with your rivals. On 12 March 2026 STAT reviewed more than 70 telehealth companies the FDA had warned in the previous six months. At least 30% publicly listed affiliations with the same four nationwide medical groups: Beluga Health, OpenLoop, MD Integrations and Telegra. The rented org chart is not exclusive. It is the same org chart rented to your competitors, so its failures arrive for all of you at once.

The breach that was not yours. OpenLoop’s systems were breached on 7 and 8 January 2026. The federal breach portal now lists up to 716,000 individuals, and the Texas Attorney General separately lists 68,160 Texas residents. OpenLoop sells platform services to many healthcare brands, so the people affected sit across many client organisations. I am not claiming any particular brand’s patients were among them. The point is structural. If you rent the platform, the incident belongs to somebody else and the patient’s anger belongs to you.

The price of the people. Wix bought Base44 in June 2025 for $80 million in cash. It was six months old, bootstrapped, and celebrated as a solo founder’s company. It had eight employees, and $25 million of the $80 million was set aside as retention for them. Thirty one percent of the price of the most famous one-person exit was paid to keep the people. The acquirer did the arithmetic the headcount story skips.

Carta’s data says the pattern is spreading and that capital has noticed. Solo founders started 23.7% of new US startups in 2019 and 36.3% in the first half of 2025. In 2024 solo-led companies took 14.7% of the cash raised in priced equity rounds. More than a third of the companies. Under a sixth of the money.

Where this breaks

Renting is usually the right call. OpenLoop can carry more clinical compliance capacity than any founder could hire in year one, and Medvi’s 16.2% net margin against 5.5% at Hims & Hers says the rented model can beat the owned one on the P&L. I agree with all of that. The argument is not that you should hire. It is that you should stop reading a short payroll as a short list of exposures.

It also matters less where the liability surface is small. A developer tool sold to developers can rent nearly everything with little regulatory exposure. Healthcare, finance, food, anything with a licence or a patient: that is where the rented org chart bites.

Dario Amodei, asked in May 2025 when the first billion dollar company with one human employee would appear, answered “2026.” On revenue he may well be right. My objection is to the word employee. It was always a payroll term, and payroll was never where the risk lived.

What I would do on Monday

1. Write two columns: every function the business answers for, and who performs it. Mark each rented line.

2. For every rented line, find out which of your competitors rent the same supplier. A shared supplier is a shared failure point.

3. Put one named internal owner on every rented function, even if that owner is you, with hours budgeted each week to read what the partner is doing in your name.

4. Rewrite partner contracts so any incident and any regulator contact reach you within 24 hours, with the right to review every claim made under your brand before it runs.

5. If you are buying or funding one of these companies, price it the way Wix did. Ask what share of the price depends on people staying, and whose payroll those people are on.

The other test

Gallagher’s lasagna test was a good one. It caught a bot drifting off its job, and he fixed it.

Run the second test before you need it. Pick any line on your rented list, assume it fails tonight, and ask who gets the letter.

At Medvi, the letter came to Medvi.

Most of what I write about here started as a client problem. If you are deciding what to rent and what to own, grab thirty minutes with me.

The first: best expense management tools for a 200 person company. The second: is your product any good for a company our size.

The first prompt decided if you would ever be considered. The second arrived after that decision was made, and its answer was close to guaranteed the moment your name went into the box. Most AI visibility reports add both into one percentage, report it monthly, and the budget follows whichever half moves easiest. That is the second prompt.

I have been writing software for 30 years. Every metric I have watched go wrong went wrong the same way: somebody let the easy part of the denominator grow.

The thesis

Branded prompt visibility is not an asset you build. It is a receipt for a decision made earlier, in a prompt that did not contain your name. A blended AI visibility score counts the receipts as wins, so most programmes are optimising and reporting the half of the prompt list they had already won. Split the score in two and put almost all of the budget on the unbranded half.

The mechanism

A buyer types your name for one reason: you are already on the list. 6sense surveyed about 4,000 B2B buyers for its 2025 Buyer Experience Study. Buying groups had filled four of their five shortlist spots on day one, first contacted a seller 61% of the way through the process, and bought from the vendor they ranked first 77% of the time. The list forms before anyone is talking to you. In 2026 a growing share of it forms in an answer box.

Now look at what the engine has to do in each case. When the prompt names you, retrieval is trivial. The engine searches your name, fetches your own pages, and talks about you, because you are the subject of the question. When the prompt names a category, the engine has to find you first, and it finds you only if something in its retrieval path points at you.

A paper Benjamin Tannenbaum posted to arXiv on 19 September 2026 measured exactly this. He modelled 34,960 unbranded prompt observations across 75 monitored projects, run repeatedly on GPT and Gemini between June and September. When neither the brand nor its domain appeared anywhere in the live retrieval path, the brand was mentioned 2.8% of the time on GPT and 3.8% on Gemini. With its own domain cited, 49.0% and 58.4%. With its own domain cited and the engine running a follow up search on the brand name, 91.4% and 100%.

The second finding matters more. Visibility persists. A brand that was absent in the previous run and had no current exposure was mentioned in the next run 1.6% of the time on GPT. A brand that was present before and exposed now: 80.5%. The unbranded answer has incumbents, and absence compounds.

The arithmetic that hides it

Take a tracked list of 100 prompts. This is a hypothetical, but a common shape: 40 name the brand, 60 do not. The brand appears in 90% of the branded answers and 5% of the unbranded ones. Blended visibility is 39%.

Now improve either half. Lift branded from 90% to 98% and the blend goes to 42.2%. Double unbranded from 5% to 10% and the blend goes to 42%. The board pack shows the same 3 point gain both times. One change added nobody to a single shortlist. The other doubled the number of buyers who could find you without already knowing your name.

The branded half is also cheaper to move. Fix a wrong price on your own site, publish a comparison page, and the branded number jumps inside a fortnight. Getting into a category answer means getting into pages you do not own. A team paid on the blend will pick the first job every time, and on the incentives it was given, it is right to.

Three data sets, one shape

The NeuroRank AI Visibility Index 2026, reported on 6 October, tested 122 brands across nine sectors on ChatGPT, Gemini, Claude and Perplexity, using 7,580 prompts. Results where the brand was named in the prompt reached high inclusion 56.2% of the time. Results without the name: 17.8%. Of 82 brands assessed both ways, 51 cleared 50% when named. Three cleared it unnamed. No sector reached 30% on unnamed prompts. The authors say this is not a controlled test of adding a name. It does not need to be. It is the gap a blended number averages away.

BuzzStream’s mentions and citations study, published in June, scored 12,000 responses across Google AI Mode, AI Overviews, Gemini and GPT-5.4-mini for about 200 brands. On single brand prompts such as “Is Asana good”, a brand mention came with a citation 39.0% of the time. On list and category prompts such as “best project management tools”, 7.2%. Even when a brand’s own page was cited in a category answer, the brand was named in the text only 36.5% of the time, against 79.1% when the prompt was about the brand. In the category answer, being a source is not the same as being recommended.

Tannenbaum’s persistence result is the third. Read together, the three say one thing. The branded half of the list is close to saturated for anyone with a working website. The unbranded half is where brands are absent, where absence sticks, and where most of them are losing.

My rule follows from that. When an AI visibility number reaches a board pack, I want the prompt list before I read the number, and I want every prompt tagged by whether a brand name is in it.

Where this breaks

Branded answers are not free. “Is it any good” is the last prompt before a buyer books a demo, and an engine that quotes last year’s pricing or a dead integration loses the deal at the final step. That is real work. It is also a correctness job, not a visibility job, and it should be priced as maintenance: a fixed quarterly line to find and fix what the engines get wrong about you, not a growth budget.

The thesis also weakens for category leaders. If most of your buyers arrive already knowing your name, the branded half carries more of the demand and the split matters less. Most companies paying for an AI visibility programme are not in that position, which is usually why they are paying.

And prompt lists are synthetic. Nobody outside the engines knows real prompt volumes, so the 40 and 60 in my example are a choice. That is the point. Somebody chose the mix in your report too, and the mix sets the number.

What I would do on Monday

1. Export the tracked prompt list from whatever tool you use and tag each prompt three ways: names your brand, names a competitor, names nobody.

2. Report two numbers from now on. Branded accuracy: is what the engine says about you correct. Unbranded inclusion: do you appear when nobody asked for you. Retire the blend.

3. If fewer than half the prompts name nobody, rebuild the list from sales call notes, using the first question buyers say they asked before they knew any vendor.

4. Give the branded half a fixed maintenance budget. Point everything else at the pages already sitting in the retrieval path for your category prompts.

5. Track unbranded inclusion run over run, per engine. Absence persists, so the first movement is the one that counts.

Two prompts, three weeks apart. The second one was always going to mention you, because the buyer typed your name. The first one decided whether the buyer ever would.

Score the first one.

If your AI visibility report still carries one number and nobody has split the prompt list behind it, pick a slot and bring the export.

Run that through a customer that held the same number of seats in July as it did in June. The vendor books expansion. Net revenue retention goes up. Nobody at the customer used more of anything.

I build software for private equity, and the first file I ask for in a software data room is the renewal ledger, not the retention slide. I read that Microsoft page the same way. It is the cleanest public example I know of a number every software deal underwrites as one thing when it is two.

The thesis

Expansion revenue is two products. Volume led expansion, more seats, more usage, more modules adopted, compounds with the customer’s business. Price led expansion, the same seats at a higher unit price, has a ceiling and a procurement team with a memory. Every diligence pack I have read reports them as one line inside net revenue retention. They deserve different multiples, and in 2026 they are moving in opposite directions.

The mechanism

Net revenue retention is a ratio. What a cohort of customers pays you now, divided by what the same cohort paid a year ago. Churn and contraction pull it down. Expansion pushes it up. The figure says nothing about why the expansion happened.

Split a single renewal and every dollar comes from somewhere specific. Either the customer is paying for more units, or it is paying more for each unit. The first is a claim on the customer’s growth. The second is a claim on the customer’s patience.

Volume compounds because it rides on something outside the vendor: the customer hiring, processing more transactions, rolling the product into another division. Price does not compound that way. A 6% uplift is a one off step, and the next one has to be negotiated again against a buyer who now has the last invoice on file. That memory shows up at renewal as a discount request, a competitive tender, or a quiet seat reduction to hold the total flat.

AI is what makes this urgent. Seat priced software has always been priced on the customer’s headcount, and AI is going after headcount first in the roles that hold the most seats: support agents, frontline staff, junior operations. The volume engine in a seat priced business can now run backwards while the price engine runs forwards. Net revenue retention nets the two into one comfortable number.

Three cases on the public record

Microsoft. E3 up 8.3%, F1 up 33%, both from 1 July 2026. A customer holding seats flat produces 8.3% expansion on E3 alone. The largest percentage rise lands on the cheapest and most numerous seat, the one issued to people on a shop floor or a ward. Any business whose revenue tracks that seat base will show its best renewal cohort in years, and almost none of the improvement will be volume.

Salesforce. On 17 June 2025 Salesforce announced an average 6% list price rise on Enterprise and Unlimited editions of Sales Cloud, Service Cloud, Field Service and some industry clouds, effective 1 August, citing “significant ongoing innovation and customer value”. By early September Marc Benioff had told The Logan Bartlett Show he had cut his own support organisation “from 9,000 heads to about 5,000, because I need less heads.”

Put those two numbers on one account. A customer that makes the same cut on a per agent support product keeps 56% of its seats at 106% of the price. That account now pays 59% of what it paid before. Price can paper over a small seat decline. It cannot cover a 44% one, and the support desk is where AI is landing first.

Workday. Workday is the honest version. CFO Zane Rowe reported gross revenue retention of 97% in November 2025 and again in August 2026, and said in August that net expansion from existing customers drove about 60% of subscription revenue growth. In November an analyst asked how Workday grows if its customers lay people off. Rowe said customer headcount growth was a net positive, “but it’s not a significant amount of our growth.” Carl Eschenbach, then chief executive, said the company was “focused not on just seats, but actually revenue per seat”, and that contracts carry “floors and minimums” on how far a customer can bring headcount down.

Read those together. The expansion is not coming from more employees, so it is coming from modules, agents or price per head, and the floor exists because the vendor expects the head count may fall. That is a listed company telling you on a public call which of the two products it is selling. Most private targets will not tell you unless you ask for the rows.

Where this breaks

Three objections are fair.

First, pricing power is a moat signal. A vendor that raises prices 8% and keeps 97% of its revenue has a product customers cannot easily leave. Agreed. Price led expansion that survives three renewal cycles with gross retention intact is evidence, and I would pay for it. What I will not pay for is one cycle of it extrapolated across a five year hold.

Second, some price rises carry real new value. Microsoft’s announcement pointed to added AI and security features, and Salesforce cited innovation. Where the ledger can show that as a new module, it is volume and should be booked as volume. Where it cannot be separated, I treat it as price, because the buyer will.

Third, volume is not pure either. Seat growth tied to customer hiring is a bet on employment in your customer base, and in a downturn it reverses faster than price does. True, and that is the reason to split them. They fail in different conditions, so anyone holding a book of these businesses needs to know the mix in each.

What I would do on Monday

1. Ask for the renewal ledger at contract line level: units and unit price for every customer at each of the last three renewals. Not the retention slide. The rows.

2. Split every expansion dollar into new units at the old price and the price change on all units. Report a volume NRR and a price NRR beside the blended one.

3. Date every list price event and check gross retention and discount depth in the 12 months after each. That is where procurement memory shows up.

4. Map the seat base by role. Any share sitting in support, frontline or junior operations gets a volume haircut in the model, because those are the seats AI reaches first.

5. Value price led expansion as a step, not a rate. Hold it flat after the next renewal unless the ledger proves it repeats.

The two columns

Microsoft’s licensing page has two columns and no explanation. That is fine for a price list. It is not fine for a retention number, and most retention numbers I am handed are built the same way: a before, an after, and nothing about why it moved.

The why is in the rows. Ask for them.

Diligence on a software asset usually comes down to a renewal ledger nobody has split yet. If one is open on your desk, find an hour with me.

He was not talking about the code. The code survives. It sits in the repository, it compiles, it passes its tests. What dies is anyone’s ability to say why it is the way it is, and therefore what is safe to change.

Forty one years later, the code is the cheap part. The theory is not.

The thesis

When AI collapsed the cost of writing code, the handover became the only deliverable in an engagement that got more expensive. What a client is actually buying now is the reason each decision was made. Almost nobody prices that separately from the code, and on any system with money running through it, it should be the larger line.

What changed

Three numbers tell the story of the build. Google said about 25% of its new code was AI generated in October 2024. By late 2025 it was reported at 50%. In April 2026 TechSpot reported Google putting the figure at roughly 75%, with engineers reviewing the output. Whatever your own number is, the direction is the same. The marginal cost of producing code is heading towards the cost of reviewing it.

That changes what a handover document is worth, because a handover has two halves and AI only touched one of them.

The first half is what the system does. Which services exist, what calls what, where the data lands. That half is now derivable. Point a model at the repository and you get an accurate tour in minutes. So a handover that describes the system is worth roughly what it costs to regenerate, which is close to nothing.

There is evidence for this. A team at ETH Zurich evaluated repository context files, the AGENTS.md and CLAUDE.md files developers now write for coding agents, in a paper accepted at ICLR 2026. Their finding: context files, whether model generated or committed by developers, generally did not improve task success, and they raised inference cost by over 20% on average. Repository overviews in particular did not help. I read that as a clean result. A description of the repository is redundant with the repository.

The second half is why. Why this queue and not a cron job. Why the pricing table is duplicated in two places. Which alternative was tried in March and abandoned, and what broke. Which constraint came from the client’s finance team and not from engineering at all. None of that is in the code, and no model can infer a conversation that was never written down.

Naur put it precisely. The programmer who holds the theory “can explain why each part of the program is what it is” and “is able to respond constructively to any demand for a modification of the program”. That is the asset. It is also the only asset in the engagement whose replacement cost went up this year.

It went up because volume went up faster than understanding. GitClear analysed 211 million changed lines and found moved lines, its proxy for refactoring, fell 39.9% in 2024. Blocks of five or more duplicated lines rose eightfold, and copy pasted lines outnumbered moved lines for the first time. More code per person who understands it means the theory is spread thinner every quarter.

I have written code for 30 years. In every system I have been handed in that time, the code told me what. The expensive week was always the one spent finding the person who could tell me why, and the worst weeks were when that person had already gone.

Three cases, all public

Naur’s own compiler. The 1985 paper rests on a real case. A second group took over a working compiler to extend it, with “full documentation, including annotated program texts and much additional written design discussion”, plus advice from the team that built it. In several major cases their extensions were patches that ignored facilities the compiler already had, facilities the documentation discussed at length. The original team spotted it immediately. Ten years later, under other maintainers, the original structure was “still visible, but made entirely ineffective by amorphous additions.” That was a well documented system. The documentation described it. It did not explain it.

The truck factor. In a study of 133 popular GitHub applications, Avelino and colleagues found 46% had a truck factor of one and a further 28% had a truck factor of two. Nearly three quarters of widely used open source applications would lose most of their theory with two departures. Those are public projects with public commit history and public issue threads, which is the best case. A private codebase built by a contractor on a fixed term is worse.

The rent on a missing why. The 2024 Stack Overflow survey found over 60% of developers spend at least 30 minutes a day looking for answers, about one in four spend an hour or more, and 30% say knowledge silos hit their productivity 10 or more times a week. That is the cost of an unwritten theory inside companies where the people who hold it are still employed. After the handover, the people are gone and the rent becomes a rewrite.

The fix has existed for fifteen years. Michael Nygard proposed the architecture decision record on 15 November 2011, five headings: title, context, decision, status, consequences. His reason was exact. Someone new to a project, without the rationale, has two choices: “Blindly accept the decision” or “Blindly change it.” AI made the second option fast and cheap. That is the problem, not the solution.

Where this breaks

The strongest objection is that models will recover the why from the exhaust: commit messages, tickets, pull request threads, chat history. Partly true. Where a decision was argued in writing, retrieval finds it, and that is a real gain. But the expensive decisions are usually the ones made in a meeting, and a rejected alternative leaves no trace in the code that replaced it. Retrieval recovers what was written. It cannot recover what was not.

Second, Naur himself argued the theory cannot be fully written down. Granted. A decision record is a lossy copy of what the team knew. It does not replace the team. It shortens the revival, and on a system with live revenue paths, the length of the revival is the cost.

Third, a growing share of code is disposable. Prototypes, internal tools, a script due to be replaced next quarter. If rewriting is close to free, rewrite it and skip the record. That is a fair call, and I make it myself. The thesis bites on the other kind of system: the one with customers, integrations and money flowing through it, where a blind rewrite is the most expensive option on the table.

What I would do on Monday

1. Split the next statement of work into two priced lines: the build, and the decision record. If the second line is zero, you are buying code you will not be able to change.

2. Require one decision record per non obvious choice, on Nygard’s five headings plus one more: what we chose not to do, and why.

3. Test the handover before the team leaves. Give an engineer who has never seen the system the records and one real change request. Time how long it takes them to say whether the change is safe.

4. Stop paying anyone to write repository overviews. A model produces those on demand, and the ETH Zurich result says they do not help much anyway.

5. For every path that touches revenue, count the people who can explain why it works the way it does. Where the answer is one, you have found your exposure.

The close

Naur wrote one more sentence in that paper. “Revival of a program is the rebuilding of its theory by a new programmer team.”

The code will still compile the morning after the team leaves. It always has. The question was never whether the program runs. It is whether anyone left can tell you why.

Most of what I write about here started as a client problem. If yours is a system nobody left in the building can explain, put it on my calendar.

In March 2026, the Financial Times reported that Amazon had called its engineers into a mandatory session after a run of outages, including a six hour failure of its main retail site. The briefing pointed to a “trend of incidents” with a “high blast radius” and to “Gen-AI assisted changes”. Dave Treadwell, a senior vice president, wrote that site availability “has not been good recently”.

Same company. Two ledgers. They never appear on the same page.

The thesis

AI productivity gains and AI delivery risk do not land in the same part of the codebase. The gains concentrate in new, simple code where nothing depends on the change. The exposure concentrates in old, complex code where everything does. The standard AI business case nets the two off as one blended number per engineer, as if they were the same system. They are not, and in the column where most of the money sits, the net can be negative.

Where the gain lands, and where the risk lands

Yegor Denisov-Blanch’s team at Stanford has productivity data on more than 100,000 developers across more than 600 companies. In the 136 teams where they split the work by maturity and complexity, the pattern is blunt. New code, low complexity: 30 to 40% gain. New code, high complexity: 10 to 15%. Existing code, low complexity: 15 to 20%. Existing code, high complexity: 0 to 10%, with some teams slower. They also found that once rework is subtracted, a 30 to 40% jump in apparent output nets to roughly 15 to 20%.

Now look at where the money is. GAO reported in July 2025 that the US federal government spends more than $100 billion a year on IT and that about 80% of it goes to operating and maintaining existing systems. Eleven of the most critical are between 23 and 60 years old. Commercial ratios vary. The direction does not: an established company spends most of its engineering budget changing systems that already make money.

So the gain is largest where the spend is smallest, and smallest where the spend is largest. That alone breaks the blended number.

The risk runs the other way. A new service with no users fails quietly. A bad change to a twenty year old billing path fails in front of customers, and it fails through dependencies nobody wrote down. Old code carries implicit contracts: the field that must never be null because a downstream report divides by it, the batch job that assumes it runs after midnight. No model sees those contracts, because they are not in the repository. They are in the heads of the people who were there.

I have written code for 30 years. In old systems the costly failure is rarely bad logic. It is a correct change to a system whose real behaviour was never written down. That is the failure an assistant is worst placed to avoid, and it sits in exactly the code where it saves the least time.

The blended number is not an average of two outcomes. It is a gain from one system used to approve work in another.

Three readings

Amazon, both columns. The Java upgrade is the clean version of the gain: mechanical change, verified by tests, on code where correct is easy to define. The March incidents are the other column. Amazon disputes that AI caused them, and in the earlier case of a 13 hour outage of AWS Cost Explorer in one China region in December 2025, where its Kiro agent chose to “delete and recreate the environment”, it called the cause user error, specifically misconfigured access controls, not AI. Take that at face value and the point survives. The agent ran with the operator’s permissions on a live system, and no second person had to approve it. The control that failed was a control for old systems with live dependents. The fix Amazon reportedly reached for was senior sign off on AI assisted changes. That is a cost, and it sits in the legacy column only.

METR, the expert on the mature repository. In July 2025 METR ran a randomised trial with 16 experienced developers on 246 real issues in their own projects, repositories averaging more than 22,000 stars and over a million lines of code. Before starting, they expected AI to make them 24% faster. Afterwards, they believed it had made them 20% faster. It made them 19% slower. The people best placed to judge a mature codebase misjudged the direction of the effect. A business case built from self reported time savings is built on that misjudgement.

The pilot. An AI business case usually carries one productivity figure, hours saved per engineer per week, taken from a pilot. Pilots get run where a team can show progress in six weeks, which means new internal tools and greenfield services. The figure is then multiplied across the whole engineering headcount, most of whom spend their week on the old systems. When I read one of these cases for a capital partner, the first thing I ask is which repositories the pilot ran on. If the answer is new ones, the pilot measured the 30 to 40% column and the rollout spends in the 0 to 10% column.

Where this breaks

The strongest objection is the first case. Amazon’s Java upgrade was old code, at scale, and the gain was enormous. That is true, and it forces a sharper version of the thesis. The line does not run between new and old. It runs between change whose correctness a machine can check and change whose correctness lives in someone’s head. A framework upgrade with a strong test suite is old code in the first category. A pricing rule in a legacy billing engine is in the second. Age is the proxy most business cases can measure today. Verifiability is the real variable.

Second, the data dates fast. METR itself now says its 2025 results are out of date, and tools improved sharply over late 2025. Grant it. A better model closes the context gap only on contracts that are written down somewhere it can read. The implicit ones stay implicit.

Third, some companies have very little legacy. A three year old SaaS business is mostly the first column, and for it a blended number is close to honest. The thesis bites hardest on established companies and on anything private equity has bought for its cash flow, which is to say the companies writing the largest AI business cases.

What I would do on Monday

1. Split the productivity line into two columns: new and existing systems. Use 30 to 40% and 0 to 10% as the opening assumptions until your own data replaces them.

2. Ask where the pilot ran. If it ran on a greenfield service, its number applies only to greenfield work, and the business case should say so in writing.

3. Mark every AI assisted change in version control with a commit trailer, so every incident review can count them. Amazon had to ask the question in a meeting. Your post incident template should answer it automatically.

4. Put a priced risk line in the existing systems column: change fail rate on those repositories before and after rollout, multiplied by your cost per incident hour.

5. Point the agents at verifiable legacy work on purpose: upgrades, migrations, dependency bumps behind a strong test suite. Keep behavioural change on revenue paths under named senior review, and budget that review time as a cost of the programme.

Two ledgers

Jassy’s post and Treadwell’s email were written about the same engineering organisation, nineteen months apart. One counted hours saved. The other counted hours down. Neither mentioned the other, because they were written for different audiences about different systems.

Your business case is written for one audience. It should carry both.

I read AI business cases from the operator’s side and the investor’s side. If yours carries one blended productivity number, open my calendar.

Read the apology as a margin document, not a communications failure. The newest Claude model cost more to run than the plan could carry, so the company handed the difference to its heaviest users.

Every model was supposed to be getting cheaper that year. It was. The margin still went the other way.

The thesis

When the model gets cheaper, the saving does not land in your gross margin. It lands in your customer’s price or in more tokens per task. A margin bridge built on the token deflation curve is a forecast of somebody else’s pricing decision, and investors are underwriting it as if it were an operating plan.

Founders raising this year are now asked the question directly: what happens to your gross margin when the model gets cheaper. The common answer is a chart. The honest answer, for most of them, is nothing.

Where the saving goes

The chart is real. Epoch AI measured it in March 2025: the price of reaching a fixed level of performance fell between 9x and 900x per year depending on the benchmark, and the price of GPT-4 level performance on PhD science questions fell about 40x a year. Gartner said in March 2026 that inference on a one trillion parameter model will cost providers more than 90% less in 2030 than in 2025.

The trouble is the word fixed. Nobody sells fixed performance. The saving leaks out through three holes, and I have not seen a pitch deck that plugs all three.

Leak one: the cost is common, so it gets competed away. Every competitor buys the same model at the same price on the same morning. A cost that every rival shares is not an advantage, it is a floor, and floors get passed through. Nicholas Charriere, founder of Mocha, told TechCrunch in August 2025 that variable costs across the code generation sector sit within 10% to 15% of each other, and that margins on those products were “either neutral or negative”.

Leak two: customers buy the frontier, and the frontier does not get cheaper. The price that falls is the price of last year’s capability. The model your customer asks for by name is this year’s. Tom’s Hardware, working from Artificial Analysis data in September 2026, put the newest flagship at $3.69 per benchmark task against $0.58 for a mid-tier model scoring about 10% lower. That is more than six times the cost for the model the buyer wants on the feature page.

Leak three: agents multiply the tokens. Gartner’s own release, the same one with the 90% forecast, says agentic models need 5 to 30 times more tokens per task than a chatbot, that provider savings will not be fully passed to enterprise customers, and that consumption will rise faster than unit costs fall. In the same document that promises the deflation, Gartner tells you total inference spend goes up.

All three leaks end in the same place: a flat margin. The only route to a wider one is something the company controls, and it is an engineering capability, not a market trend.

What it looks like in the numbers

The apology. Cursor reached $500 million in annual recurring revenue in June 2025, the same month as the pricing change. That is a company with scale, pricing power and a product people love, and it still could not hold a flat price when its preferred model got more expensive per task. Scale did not protect the margin. Repricing did.

The more autonomous agent. Sacra’s research on Replit puts its gross margin swinging between 36% and negative 14% during 2025, with the low in April, after a more autonomous agent shipped. The model price did not jump. The product started doing more work per request, and every extra step was billed to the company before it was billed to anyone else. The margin was back at 23% by July, after Replit changed how it charged for agent work. Again: repricing, not deflation, did the repair.

The survey. ICONIQ’s January 2026 State of AI, about 300 software executives building AI products, shows average AI product gross margin of 41% in 2024, 45% in 2025 and 52% projected for 2026. Over the same two years the price of a given capability fell by one or two orders of magnitude. An 11 point margin gain against a 90% plus input cost decline is not a pass through. It is a leak. The same survey shows consumption pricing at 35% of companies, up from 19%, and outcome pricing at 18%, up from 2%. Consumption pricing is how firms pass rising costs to customers. It is also, mechanically, how they pass falling costs to customers. You cannot have one without the other.

I build software for private capital and I pay model invoices every month. The line on those invoices that matters has never been the price per token. It is tokens per finished job, and that line only moves when somebody changes the product. The price list moves on its own. The margin does not. When I read a margin bridge, the first thing I look for is the line that assumes model prices fall, and then the name of the person who owns it. There is usually no name.

Where this breaks

Three objections, and the third is the strongest.

First, margins did rise. 41% to 52% in two years is real money, and Bessemer’s 2025 State of AI put its steadier cohort of AI companies at around 60%. True, and it supports the thesis rather than refuting it. An input cost that fell by orders of magnitude bought 11 points of margin. The rest of the saving went somewhere, and it was not to the shareholders.

Second, some products sit on capability that stopped moving. A document classifier that was good enough on a 2024 model does not need the 2026 flagship, and its owner keeps the whole saving. Correct. If the buyer cannot tell the difference between models, deflation flows to the seller. That describes a narrower set of companies than the decks claim, and it is a set whose moat is unlikely to be the AI.

Third, routing is the answer, and well run companies already do it. Send each task to the cheapest model that passes the evaluation, keep the frontier for the hard ten percent, and the margin expands while the price holds. I agree completely. That is the one route to a wider margin. But notice what it is: an engineering programme with an evaluation suite, a routing layer and an owner. It is not the deflation curve. A company that has built it should be valued for it. A company that points at the curve and says routing is on the roadmap is asking you to pay today for an asset it has not built.

What I would do on Monday

If you are underwriting an AI software company this quarter:

1. Ask for gross margin by month for the last 12 months with every model price change marked on the chart. If the margin did not move when Anthropic cut its flagship by two thirds on 24 November 2025, the deflation thesis has already been tested on the company’s own data, and it failed.

2. Ask for tokens per completed outcome, not cost per token. If the company cannot produce that number, it does not know where its margin goes.

3. Ask what share of revenue is consumption priced. That share passes every future price cut straight to the customer, by contract.

4. Ask what share of requests is served by something other than the flagship model, and to see the evaluation that permits it. No evaluation means no routing, whatever the roadmap says.

5. Strike the deflation line from the margin bridge in the model. Replace it with an engineering line that has a named owner and a date.

The apology

Truell’s apology was about how a change was communicated. The arithmetic underneath it was about who keeps the saving when the model moves, and in June 2025 the answer was nobody inside the company.

That has not changed. The model got cheaper. The customer noticed first.

I read AI margin bridges from both sides of the table, the one that builds the product and the one that funds it. If you are underwriting one this quarter, grab thirty minutes.

Last month I wrote an AGENTS.md. No human has read it. It is loaded in full at the start of every agent session, and every one of those loads is billed to me.

Thirty years of writing software and I have never had a document charge rent before.

The file is a purchase

The agent context file is not documentation. It is a line item with a per use price, the price is measurable, and almost nobody has measured it. The evidence we have says you are paying a certain cost for an uncertain gain, and you are paying it on every run for the life of the repository.

That is a procurement decision. It is currently being made by whoever had a spare hour.

Why this one is different

A README is pull. Somebody wants it, somebody opens it, the cost of the read is their attention. A context file is push. It is prepended to the working context every time an agent starts, every time a session compacts, every time a subagent spawns. The cost does not scale with how many people find it useful. It scales with how many times you run.

So the economics invert. A long README written once and read rarely is cheap. A long context file written once and read constantly is expensive, and it gets more expensive the more successful your agent programme becomes.

In February 2026 a group out of ETH Zurich put a number on it. Thibaud Gloaguen and colleagues built a benchmark they call CTXbench: 138 task instances drawn from 12 Python repositories that already had a developer committed AGENTS.md or CLAUDE.md at the root. They screened 5,694 pull requests to get there. They ran four agent configurations over it: Claude Code on Sonnet 4.5, Codex on GPT-5.2, Codex on GPT-5.1 mini, and Qwen Code on Qwen3-30b-coder served locally. Then they ran each task three ways: no context file, a model generated one, and the real one the developers had written.

The results are not what the file is sold on.

A model generated context file moved task success by about minus 0.5% on SWE-bench Lite and minus 2% on CTXbench. Neither result was significant. A developer written file moved it by plus 2.4% on average, p equal to 0.21, also not significant. The one comparison that cleared the bar was developer written against model generated, where the human file won by a margin the paper puts at 7%. Their own summary sentence is flat: “providing context files does not generally improve task success rates.”

The cost side was not ambiguous at all. Model generated files raised inference cost by roughly 20% on SWE-bench and 23% on CTXbench. Developer written files raised it by at most 19% and added 3.34 steps per task on average. The agents were not ignoring the files. The paper is explicit that “instructions in context files are well followed.” They read the instructions, then went and did more work: more tests, more greps, more files opened, more quality checks, much of it beside the point of the task in hand.

And the single most common section in these files, the repository overview, is the one the paper singles out as useless: “repository overviews, although popular and recommended by model providers, are not helpful.” Overviews did not meaningfully reduce the number of steps an agent took before it touched the file it actually needed to change.

What that costs in money

Gartner Peer Insights found in June 2026 that 23% of technology leaders are spending $200 to $500 per developer per month on tokens for coding agents, and 6% are spending over $2,000. DX puts enterprise Claude Code deployments at about $13 per active developer day, or $150 to $250 a month. Gartner’s Nitish Tyagi expects per developer token cost to go from $20 to $200 and in some cases to $2,000, and Gartner forecasts that by 2028 AI coding cost will exceed the average developer salary.

Take the conservative end. Call it $300 a month per developer. A 19% context tax is $57 per developer per month. Across forty engineers that is $27,360 a year, recurring, for a gain the measurement could not distinguish from zero.

That is my arithmetic, not Gartner’s. Run it on your own numbers. The point is that the number exists and it is not small, and I have not once seen it on a budget line or in an AI programme business case.

Three things the field data says

First, almost nobody has one. Research presented at the 23rd Mining Software Repositories conference in 2026 scanned 10,000 active repositories and found roughly 5% carried any AI context file at all. 466 had an AGENTS.md. So the practice everyone is writing about is a minority practice being adopted at speed, before anybody measured it.

Second, the files are written once and abandoned. In the deep analysis of that sample, 50% of the files had not been modified after the commit that created them. Another 23% were modified exactly once. A context file is a set of imperative instructions that an agent follows well. An abandoned one is a set of wrong instructions that an agent follows well. That is not a neutral state, it is worse than having no file, and the failure is silent because the agent never tells you it obeyed something false.

Third, the maintenance is real work and it compounds. An experience report published in February 2026 documents one developer building a 108,256 line C# simulation with Claude Code over 283 sessions and 70 days. The supporting context layer ended up at 54 files and about 26,200 lines, which is 24.2% of the line count of the product itself. Maintenance ran one to two hours a week. The author names the main failure mode without hedging: stale specifications produced conflicting code at least twice.

A quarter of your codebase, in prose, that nobody compiles and no test covers.

Where this breaks

The honest objections are good ones.

The ETH sample is 12 repositories and 138 tasks, Python only, and the repositories are small and recent by construction, because they had to already have a context file to qualify. A twenty year old monorepo with genuinely non obvious conventions is exactly the case where a file should earn its keep, and it is exactly the case the benchmark could not include. I would not call this settled.

A 2.4% lift that fails a significance test is not the same as no lift. It is an underpowered result. If a single failed agent run on a critical path costs you a day of senior time, paying 19% more per task to shift the odds is a trade most operators would take without blinking, and the benchmark has no way to price that asymmetry.

And the strongest objection is the one the paper cannot measure at all. Writing the file forces someone to write down the knowledge that was only in three people’s heads. That is worth doing whether or not an agent benefits, and the humans who join next year will benefit from it more than the model does.

I accept all three. None of them makes the file free. They change the answer to “is this worth buying”, which is a question you can only ask once you admit you are buying something.

What I would do on Monday

1. Measure your own tax before you argue about mine. Take ten representative tasks, run them twice with the same agent and model, once with the context file and once with it removed, and compare token cost per completed task. An afternoon of work and you will never have to take a blog post’s word for it again, including this one.

2. Delete the repository overview. It is the most common section in the corpus and the one section the research specifically found does not help. The agent can read your directory tree. It is faster at it than your prose is.

3. Put a hard ceiling on the file and make every new line compete for space. The observed average for AGENTS.md is 142 lines. CLAUDE.md averages 287 and Copilot instruction files average 310. Length here is a cost, not thoroughness.

4. Put an owner and a review date in the first three lines. Half of all these files are never touched again after the day they are born, and the instruction that was true in March is being executed with full confidence in October.

5. Move anything checkable out of prose. If the rule is “always run the linter before committing”, that belongs in a pre commit hook where it is enforced, not in a paragraph where it is re read and re billed a thousand times and still only sometimes obeyed.

The general rule I now apply: a line earns its place in the file only if the agent cannot infer it from the repository and it will still be true in ninety days. Everything else is a script, a test, or nothing.

The binder

That printed README is probably still in a cupboard somewhere. It was a bad document by any modern standard. It was out of date within a year and nobody noticed, because nobody opened it.

The new one gets opened every few minutes, forever, by something that believes every word.

Write it like it costs money. It does.

Most of what I write here started as somebody’s build cost going somewhere nobody forecast. If that is your week, grab a time.

A hundred and ten million dollars is a rounding error on Gartner’s balance sheet. It is also the price of the property that now decides which three vendors get named when a buyer asks an AI engine what to buy. I have never once seen that property appear on a marketing plan as an asset with an owner.

The thesis

Your content budget is pointed at the property that only pays after a buyer already knows your name. The property that produces buyers who do not know your name is owned by somebody else, and this year it started consolidating. This is not an authority problem to be solved with more publishing. It is an ownership problem.

Two populations of prompt, and almost nobody separates them

Here is the thing that makes the AEO debate go in circles. Two credible studies published within a year of each other say opposite things, and both are right.

Profound classified 11.84 billion citations between 16 April and 16 July 2026, across eight engines: ChatGPT, Claude, Google AI Mode, Google AI Overviews, Gemini, Grok, Microsoft Copilot and Perplexity. It mapped 3.02 million domains covering 98.3 percent of citation volume. The headline: about 57 percent of citations go to brand sites. In 24 of 29 industries, the brand’s own domain is the largest bucket for the median company. Cybersecurity runs at 74 percent brand. Marketing runs at 72 percent.

AirOps looked at 21,311 brand mentions across more than 500 commercial-intent queries on GPT-5, Claude Sonnet 4.5 and Perplexity Sonar. Queries of the form “best HR software” and “top CRM platforms”. The headline: 85 percent of brand mentions came from external domains. Only 13.2 percent came from the brand’s own site. A brand was 6.5 times more likely to be mentioned through somebody else’s page than its own.

Fifty-seven percent versus thirteen percent. The gap is not a methodology dispute. It is the prompt.

Profound measures the prompt sets its customers track, and companies that buy a visibility tool track their own name. When your brand is in the question, the engine goes to your documentation, your pricing page, your changelog, and it should. AirOps measured category questions with no brand name in them. When your name is absent from the question, your pages are not candidates for the answer. They are candidates for the follow-up.

So the owned-media spend is not wasted. It is conditional. It pays on the prompts that already contain the thing you were trying to buy.

The unbranded prompt is where the new revenue is

G2 surveyed 1,076 B2B software buyers and decision-makers in March 2026 and interviewed 39 marketers alongside it. 51 percent now start software research with an AI chatbot more often than with Google, up from 29 percent in April 2025. Within seven months, the share calling chatbot research more productive than search went from 36 percent to 53 percent.

Two numbers from that survey matter more than the adoption curve. 69 percent chose a different vendor than the one they planned to choose, on chatbot guidance. 33 percent bought from a vendor they were not previously familiar with.

Read that second figure as a prompt type. A buyer who was not familiar with you did not type your name. One in three purchases now begins in a question your own pages cannot be cited in. That is the population your content budget does not reach, and it is the only population that grows your customer list rather than confirming it.

The Content Marketing Institute surveyed 1,015 B2B marketers between 24 June and 14 August 2025 on where 2026 investment goes. Owned media came in at 32 percent of top-three selections. Social and earned media came in at 20 percent. The budget is weighted 1.6 to 1 towards the property that works on the branded half.

Three readings of the same data

Software has the thinnest earned base in the market. Profound puts the median earned share for SaaS and software at 11.4 percent. Set that against government and nonprofit at 15.9 percent brand share, which is to say almost everything cited about them sits somewhere else. Software companies publish the most and are the most dependent on their own domain, in the category where a third of buyers just bought from somebody they had not heard of. Inside a single industry, the spread between the 75th and 25th percentile company on earned share reaches 43 percentage points. That is not a market average problem. Some companies in your category have already made this trade.

The shape of the third-party page is specific, and it is not a blog post. Nearly 90 percent of those third-party mentions sat in listicles, comparisons or reviews. About 80 percent of the brands cited in them appeared as one of the first three companies discussed. Position inside somebody else’s page is the unit. Not the page you wrote. Not the word count.

The engines do not agree, so neither should your plan. ChatGPT runs 47 percent brand and 30 percent earned, the highest earned share of the set. Gemini runs 69 percent brand. Google AI Overviews sits at 17.0 percent earned and 15.3 percent social, citing social sources at roughly 1.3 times ChatGPT’s rate and 4 times Microsoft Copilot’s, where social is about 1 citation in 29. On first-party share, GPT-5 ran 4 to 11 percent while Claude and Perplexity ran 13 to 21 percent. And 68 percent of brands appeared on only one platform at all. A single content strategy, measured as a single score, averages four different games into a number that describes none of them.

Then the ownership fact lands on top. 45 percent of those 1,076 buyers said a citation from a software review site is the most confidence-inspiring signal in an AI answer, the highest of any source. Review sites are also the only source besides the chatbot itself whose influence rises deeper in the funnel, from 40 percent at discovery to 47 percent at retention. That is the signal buyers trust most. As of 5 February it has one fewer landlord than it had in January.

Where this breaks

Four honest limits, and the first two run against me.

Profound’s 57 percent is real and it is the largest sample here by three orders of magnitude. If your category’s demand genuinely arrives through branded prompts, your owned-content budget is correctly placed and this post is an expensive distraction. Cybersecurity at 74 percent brand looks exactly like that category.

AirOps sells the platform its report recommends, the report gives no data collection window, and it ran on GPT-5, Claude Sonnet 4.5 and Perplexity Sonar. It did not touch Google’s surfaces, which carry far more query volume than all of those combined. The 85 percent is directional, not a constant.

G2 asked buyers who use G2 whether they trust review sites, and G2 bought three more review sites a month before fielding it. I am using the number anyway, because the direction matches citation data collected by people with no review marketplace to sell. Weigh it accordingly.

And the limit I cannot close. Everything above shows where the citation slot sits. None of it shows that being named in a comparison page causes a purchase. Citation share is an availability measure. Moving budget on it is a bet on a correlation, and anyone who tells you the attribution is solved is selling something.

What I would do on Monday

1. Split your tracked prompt list in two: prompts that contain your brand name, prompts that do not. Report the halves separately and never again as one score. If your tool will not do this, export and do it in a spreadsheet.

2. For the unbranded half only, pull every cited URL from the last 30 days and add one column: owned, or not owned. That column is the whole diagnosis and it takes an afternoon.

3. Take the top 20 not-owned URLs and write the owner and the access mechanism next to each. Some are a review profile you control. Some are an editor who takes a pitch. Some are a forum thread nobody controls and you need to stop pretending otherwise.

4. Treat your review profiles as a product surface, not a vanity page. Category placement, review recency, response coverage, feature tagging. You are being read by a retrieval system that counts position, and 80 percent of cited brands are in the first three named.

5. Invert one slice of the 1.6 to 1 split for a single quarter, and judge it on unbranded citation share rather than sessions. If it does not move in 90 days, you have an answer worth having.

Close

The filing was public on 12 February. Anybody could read it. A $110.0 million transaction changed who owns the pages that introduce vendors to buyers who have never heard of them, and the teams whose pipeline depends on those pages were not in the room and did not notice.

You cannot buy the property now. You can at least find out whose it is.

Most AEO programmes I get asked to review are measuring one number that averages the half they can influence with the half they cannot. If you want your prompt set split into branded and unbranded, and the cited URLs marked owned or not owned, my calendar is here.

That midpoint is what an AI visibility dashboard prints. One number, a green arrow, a month on month delta. It is not a bad dashboard. It is an index with no published methodology, sold as a KPI.

The thesis

A single AI visibility score is an index. Every index that anyone allocates real money against publishes three things: its constituents, its weights, and its rebalancing rule. AI visibility vendors publish none of the three. The four engines they blend share 3.8 percent of their sources, so almost the entire value of the composite is determined by a weighting the vendor will not show you. More data does not fix this. It makes the number smoother and no more meaningful.

What the engines actually agree on

Two large studies landed in 2026 and they agree with each other, which is more than the engines do.

Writesonic analysed 161,286 prompts across ChatGPT, Gemini, Perplexity and Google AI Overviews in May and June 2026, with 70,879 of those returning citations from all four. Sources cited by all four engines: 3.8 percent. Domains appearing on exactly one engine: 72 to 73 percent. The pairwise Jaccard similarity between any two engines tops out at 0.237, for Perplexity and AI Overviews. ChatGPT against Gemini is 0.119. As the authors put it, no pair clears 0.25, so being cited on one engine does not predict being cited on another.

Wellows ran a larger set over January to June 2026: 22.7 million citations, 1.15 million questions, five engines, 531,889 questions where all five returned citations. Sources appearing on exactly one engine: 79.6 percent. Sources appearing on all five: 0.31 percent. Perplexity never touches 89.1 percent of the pages ChatGPT cites on the same question. ChatGPT is the most isolated of the five, with 76.3 percent of its citations untouched by any other engine.

Now put that next to Ahrefs, who looked at 863,000 keywords and 4 million AI Overview URLs and found that 38 percent of AI Overview citations come from pages ranking in Google’s top 10, down from 76 percent in July 2025. Read those two facts together. The strongest agreement between any two answer engines is 23.7 percent. Google’s own AI agrees with Google’s own top 10 on 38 percent. The engines disagree with each other more than the most Google-shaped engine disagrees with Google. A composite is being built out of constituents that are close to disjoint.

The weighting nobody prints

Here is the part that turns a measurement problem into a capital allocation problem.

BrightEdge put ChatGPT at 95.1 percent of all AI referral traffic in August 2026. Gemini 2.4 percent, Perplexity 1.1 percent, Claude 0.8 percent. ChatGPT is also the engine that agrees least with everything else.

So an equal weighted score across four engines gives ChatGPT 25 percent of the number. Three quarters of what a board sees is describing engines that between them send 4.9 percent of the clicks, and those engines cite a near disjoint set of sources, which means the three quarters is not even a noisy version of the one quarter. It is a different measurement of a different thing.

An audit of nine visibility platforms published on 10 September 2026 found exactly what you would expect. Denominators vary: the same underlying data gives 50 percent on an all response denominator and 83.3 percent on a brand only denominator, a 33 point swing with no change in the world. Run frequency ranges from once a day to 100 runs per prompt. Some vendors pool engines, some average them. Of nine, one discloses a weighting at all, and that one weights by Google search volume, which is a Google proxy applied to engines that are not Google.

Then add sampling. Kevin Indig’s work on prompt tracking puts within engine variance at 10 to 34 percent on identical prompts, with AirOps reaching 815,000 prompt page pairs to establish the baseline. After three runs in ChatGPT, 2.2 percent of citations remain. High versus low reasoning settings move the citation rate 18 points. Five runs per prompt per platform per week is the floor for a confidence interval worth printing.

Stack it up. Four constituents that share 3.8 percent of their contents, each measured with a sampling band of 10 to 34 percent, combined with an undisclosed weight and an undisclosed denominator, reported to one decimal place with a month on month arrow.

Where this breaks

Four places, and two of them cut at my own argument.

The brand layer is not the page layer, and this is the strongest objection. Wellows found engines name the same companies 30.3 percent of the time while citing the same exact page only 6.8 percent of the time. Agreement at the entity level is roughly four times agreement at the document level. If your score counts brand mentions rather than citations, the disjointness I have just spent this section on shrinks by a factor of four. It does not vanish, 30.3 percent is still low, but the argument is weaker there and I am not going to pretend otherwise.

Referral traffic is the wrong denominator for reading. BrightEdge counts clicks. A large share of AI answers produce no click at all, which is the whole zero click argument, so 95.1 percent of referrals is not 95.1 percent of impressions. If Gemini answers ten times as often as it refers, my weighting arithmetic overstates ChatGPT. I have not found a credible cross engine impression share for 2026 and I would not build the number without one.

A consistent bad index still has a usable first derivative. If the vendor never changes the method, the level is meaningless but the trend may be real. That is a fair defence and it holds right up until the vendor ships a model update, adds an engine, or silently changes the denominator, none of which they announce.

And against my own side: an operator who has never written down the prompt set has no standing to complain about the vendor’s weighting. Most of the teams I see have not specified the 30 questions their buyers actually ask. Until you have, the vendor’s defaults are your fault, not theirs.

What I would do on Monday

1. Kill the composite in the board pack. Report per engine or report nothing. One row per engine, with the run count next to it.

2. Pick the one engine that matters for your buyer and say out loud why. For most B2B in 2026 that is ChatGPT on referral evidence. Optimise against that one, measure the others quarterly for drift.

3. Ask the vendor three questions in writing before renewal: what is the denominator, how many runs per prompt per week, and how are engines weighted. If any answer is missing, the number is not a KPI.

4. Write the prompt set yourself. Thirty questions your buyers ask in their own words, versioned in a file, changed deliberately and dated. That file is the asset, not the dashboard.

5. Report every figure with a band. 78 plus or minus 6 on one named engine is an honest number. A single 56 with a green arrow is not.

The close

I have spent thirty years building software and a good part of the last few years reading index methodology documents, because when you work against listed company benchmarks you learn quickly that the rulebook matters more than the level. S&P publishes its constituents, its float adjustments and its rebalancing dates. The ASX benchmarks publish theirs. Anybody can reconstruct the number from the method. That transparency is not a courtesy, it is the only reason an index is allowed to be a decision input.

I have never once been handed the prompt list behind an AI visibility number without asking for it twice.

Seventy eight and thirty four are both real. Fifty six is not. The arrow on top of it is the only part anyone ever reads.

Thirty years of building software has mostly taught me to distrust a number I cannot take apart. If an AI visibility score is about to go into your board pack, book a call and we will decompose it first.

I deleted some of those boards. So did most platform teams I worked with between 2019 and 2023. It was the right call and I would make it again.

We are about to put all of it back. Not because 2019 was wrong. Because the reason it was right has stopped holding.

The thesis

The approval layer was safe to delete because accountability lived inside the actor. A named engineer with a manager and a career pushed the change. An agent has no career. Take the approver out of an agent workflow and nothing is holding the risk.

So every company scaling agents is about to rebuild the approval middleware it spent a decade deleting. It will cost real money. Almost nobody has a budget line for it, because the same business case that created the need deleted the line.

Two findings about the same budget

McKinsey published its 2026 State of AI on 25 August, from 1,719 respondents across 97 nations surveyed between 4 May and 8 June. Forty percent of organisations above a billion dollars in revenue now report scaling AI agents, up from 27 percent a year earlier. Thirty-one percent of large enterprises are scaling coding agents specifically.

Then the line that should stop a CFO. Thirty-two percent of organisations report deciding against buying one or more software products or features, because they could build it in house with coding agents instead.

Hold that against Gartner. On 26 May 2026 Gartner predicted that by 2027, 40 percent of enterprises will demote or decommission autonomous AI agents because of governance gaps found only after a production incident. Shiva Varma’s framing is four autonomy tiers: Observe, Advise, Act with Approval, Act Autonomously.

Look at what three of those four tiers actually require. Observe needs logging and retention. Advise needs a queue and a named reviewer. Act with Approval needs policy, routing, pending state, evidence and reversal. None of that is a model capability. It is middleware, and it is the same five things a 2008 BPM suite sold.

The software you stopped buying was the workflow layer: approvals, routing, exception handling, audit trail. You cancelled it because the tool that creates the need made building look cheap.

The two denominators

TheAgentCompany benchmark, built at Carnegie Mellon and published in the NeurIPS 2025 datasets track, stood up a simulated software firm on GitLab, OwnCloud, Plane and RocketChat and gave agents 175 real tasks. The best model tested, Gemini 2.5 Pro, finished 30.3 percent of them outright and scored 39.3 percent with partial credit. Claude 3.7 Sonnet finished 26.3 percent. GPT-4o finished 8.6 percent.

Read that as an operating cost rather than a capability score.

Verification cost scales with attempted work. Value scales with completed work. At roughly 30 percent completion the checking layer handles about 3.3 items for every one that ships. Every agent business case I have been shown sizes human review against the completions. Not one has sized it against the attempts. That is a factor of three error, and it widens as you deploy more agents, because you deploy them into the categories where they fail most.

Where the failures cluster

TheAgentCompany found agents strongest on software engineering and weakest on administrative, finance and communication work. RocketChat tasks, which means talking to a colleague, scored worst across every model tested. OwnCloud tasks, which means office documents, were close behind.

Now put Gartner’s CIO poll from 19 May 2025 next to it. Of 125 respondents, 52 percent said their agent deployments target internal administration, and they named IT, HR and accounting. 23 percent named customer-facing work.

The enterprise is aiming agents at the exact category the benchmark says they are worst at. It is also the category that already carries statutory controls: segregation of duties, approval thresholds, a signed audit trail.

Engineering is where agents perform best and where verification is cheapest, because code review already exists as a funded institution with tooling, a norm and a named owner per repository. Accounting has controls too. They are procedural, held in people and spreadsheets, and not expressed as an engineering artefact an agent can be routed through.

That mismatch explains Gartner’s demote-or-decommission number better than “governance gaps” does. The gap is not a missing policy document. It is a missing product.

There is a third turn. Build versus buy inverts here. You build what differentiates and you buy what regulates. A verification layer earns its keep by being boring, complete and auditable, not by fitting your business. Coding agents made everything look buildable, which pushes teams to build the one category where buildability was never the right test. Twelve months later you own a half finished approvals queue with no single sign-on, no retention policy and no evidence export, maintained by whoever drew the short straw.

Where this breaks

Four places, and two of them are aimed at my own side.

The strongest objection is that the layer arrives as a product before your budget cycle does. Gartner expects guardian agent technologies to take 10 to 15 percent of agentic AI markets by 2030, and 70 percent of AI applications to be multi-agent by 2028. If verification ships as a feature of tools you already license, the budget problem dissolves and this argument expires quietly. I think it is half right. The plumbing will commoditise. The policy about who may approve a $40,000 credit note is yours and nobody can ship that for you. But I would not bet a build against the plumbing.

Second, the benchmark is a simulation. TheAgentCompany is a fictional firm with synthetic colleagues, and 30.3 percent on 175 tasks is not 30.3 percent inside your finance function. I am using the ratio as a shape, not a forecast.

Third, and this is the one that cuts hardest, Gartner’s warning about uniform governance runs directly against me. Putting an approval gate on every agent action is the failure mode on the other side. An agent that turns a meeting into a draft nobody sends does not need a control, and if you route it through one you have rebuilt the change advisory board and earned the 2.6x. Proportionality is the entire skill and I have no clean rule for it.

Fourth, scope. If your agents only touch reversible, low value, internally visible actions, none of this applies. Build nothing. The argument starts where an action has a counterparty.

What I would do on Monday

1. Count attempts, not completions. Take one agent already in production. Pull the number of outputs it produced last month against the number a human accepted unchanged. That ratio is the real throughput requirement for your verification layer. If nobody can produce the number, that is the finding.

2. Classify every agent you run into the four tiers, on one page, this week. It takes two hours and it almost always demotes something. Anything sitting in Act Autonomously that touches money, a customer or a record of account moves to Act with Approval until a person signs for it.

3. Name an owner for the approve step, with authority to say no and time in their week to use it. In most plans this seat does not exist, because the headcount it would have come from was removed in the same business case that bought the agent.

4. Get a price on buying the layer before you build it. If your workflow or service management vendor already sells approvals, routing and an evidence trail, the honest comparison is their licence against your twelve months plus your maintenance.

5. Write the reversal procedure before the first autonomous action, not after the first incident. What undoes it, who can trigger the undo, and how long you have to do it.

Close

The 2019 finding was never that approval is waste. It was that approval by someone with no stake and no context is waste. We collapsed that into “approvals are slow” because the short version was easier to sell, and for ten years the shortcut cost nothing, because the actor pushing the change had a name, a manager and a career.

The agent has none of those. Somebody still has to hold the risk, and it will not be the model.

The approve step is the part of an agent program nobody volunteers to own, which is why it usually gets discovered after the incident instead of before it. If you want the verification layer in your agent plan counted and staffed before deployment rather than after, my calendar is here.

Most companies should not build agents this year. Not because the models cannot do the work. Because the number that decides an agent programme is the cost of finding out when it was wrong, and almost nobody prices that before they start. Get it wrong and the failure is not a pilot that dies quietly. It is a pilot that works, ships, and puts a permanent checking tax on every team downstream.

The number the demo does not show you

TheAgentCompany is the least forgiving public benchmark I know of. It drops a model inside a simulated firm with its own code host, its own chat and its own colleagues to ask, then gives it 175 tasks across software engineering, project management, data science, administration, human resources and finance. Real work, in sequence, with the context scattered the way it is in your company.

Gemini 2.5 Pro, the best of them, scored 30.3 percent full autonomous completion. With partial credit for getting most of the way there, the same model reaches 39.3 percent. It spends an average of 27.2 model calls and $4.20 per task attempted.

Now do the division nobody in the vendor deck does. $4.20 is the price of an attempt, not the price of an outcome. At 30.3 percent, a completed task costs $13.86 and burns roughly 90 model calls. And that figure is still wrong, because it counts none of the human time spent reading the other two attempts out of three to work out which bucket they landed in.

That is the argument in one line. The agent is cheap. Knowing whether it worked is not.

Reliability is not on the same curve as capability

A Princeton group published a preprint on 24 August 2026, Towards a Science of AI Agent Reliability. Rabanser, Kapoor, Kirgis, Liu, Utpala and Narayanan ran 14 models across two benchmarks and made one distinction that should reorganise how you buy this.

Pass at k is the probability the agent succeeds at least once in k attempts. That is the demo. Pass to the power of k is the probability it succeeds every time. That is your Tuesday. Benchmarks report the first, which is best case capability with a friendly operator in the chair.

Their headline finding is the one for your board. Across 18 months of model releases, accuracy climbed steadily while consistency, robustness and calibration did not keep pace. Capability and reliability are two curves and the industry reports one.

This matters for the wait or build decision, and most people have it backwards. Wait a year and what improves is capability. What you needed was calibration: knowing which runs to check. That is not obviously arriving on schedule.

The arithmetic that decides it

Take a step the agent gets right 95 percent of the time. Excellent by current standards. Chain five of them and the workflow completes 77 percent of the time. Chain ten and you are at 60. Chain twenty and you are at 36. To reach 95 percent end to end across twenty steps, every individual step needs to be right 99.74 percent of the time. Nothing in production is at 99.74 percent.

So the question is never whether the model can do the step. It is how many steps run before a human looks. That is a process design question, answered by how your work is arranged rather than by which model you licensed, which is why it belongs to whoever owns the operating model and not to the platform team that is excited about it.

Thirty years in, the step count is the first number I ask for. Before the model, before the vendor, before the budget.

Who is already doing it, and why

About 31 percent of enterprises now run at least one agent in production, on mid 2026 numbers from S&P Global Market Intelligence and McKinsey. Banking and insurance sit near 47 percent, well ahead of everyone else.

The usual reading is that regulated industries have more money and better engineers. Wrong. They already had what an agent programme requires, built for regulators decades before anyone said the word agent. A pending transaction that settles later. Four eyes above a threshold. An audit trail, so a wrong action can be traced and undone. Checkpoints and reversibility, bought in advance for a different reason.

Everyone else is sold the agent first and asked to build the checkpoints afterwards. Wrong order, and the order almost every programme is running in.

Gartner expects over 40 percent of agentic AI projects cancelled by the end of 2027, and names three causes: escalating costs, unclear business value, inadequate risk controls. All three are verification problems wearing different hats. Escalating cost is cost per completed run discovered late. Unclear value is nobody having defined correct. Inadequate risk control is an irreversible action surface.

Gartner also estimates that of the thousands of vendors marketing agentic products, roughly 130 are the real thing. Read that as a buyer. The base rate of any given vendor being genuine is under five percent, so honest diligence on a shortlist costs more than the pilot it was meant to de risk. In that market the correct default is not to buy.

The supporting numbers rhyme. IBM’s 2025 CEO study found 25 percent of AI initiatives delivered the expected return. Median time to value runs about 5.1 months on BCG and Forrester figures. Data quality is named the biggest blocker by 52 percent, which is another way of saying nobody wrote down what correct looks like.

The four questions

Here is the test. Four questions, in order. A no stops the programme there, not six months and two hundred thousand dollars later.

One. Can you write down what correct looks like, in a file, for thirty real past cases, before anyone builds? Not a description of correct. The thirty actual inputs and their actual right answers. If your team cannot produce that file in a week, you have an aspiration rather than a specification, and no model fixes that.

Two. Is a wrong action reversible? If the agent can send the money, email the customer or close the account and there is no undo, it proposes and a person commits. That is not timidity. It is what banking did.

Three. How many steps run between human checkpoints? Count them on a whiteboard. Over ten and the arithmetic says you are buying a system that finishes 60 percent of the time at best, whoever’s model is inside it.

Four. Do you know your cost per completed run? Not cost per call. Cost per call divided by success rate, plus the loaded cost of the minutes a person spends checking.

Clear all four and build it this quarter. That set is much smaller than the market implies and much larger than zero.

Where this breaks

Four honest problems with what I just argued, two of them aimed at me.

The benchmark is a benchmark. 175 tasks in a simulated company is not your company, and narrowly scoped deployments routinely beat general benchmark numbers by a wide margin. If your agent does one classification on one document type, 30.3 percent is not your number and I should not pretend it is.

Worse for my case: that benchmark’s frontier is moving, and quoting any fixed completion rate to argue against building is the same stale evidence move I complain about when vendors do it in reverse. The honest version of my position is that the ratio between capability and verification cost has not changed much, not that the capability number is fixed.

Gartner’s 40 percent is a forecast, not a measurement. Forecasts from firms that sell advice on avoiding the outcome they forecast deserve a discount.

And the strongest objection is one I cannot fully answer. The organisations at 47 percent got there by starting before the economics were legible. Learning by doing is real, capability compounds inside teams, and a rule that says wait until you can price it would have told them to sit out. Part of the cost of my position is a year of institutional learning the people who ignored me will have and you will not. I think the trade is right for most companies. I do not think it is free.

Where the thesis does not apply at all: high volume, low stakes, fully reversible internal work. Ticket triage, drafting, classification, retrieval, summarising a call. Build those today.

What I would do on Monday

1. Pick the workflow everyone keeps pointing at. Count the steps between human checkpoints and write the number on the wall. If it is over ten, this quarter’s project is shortening the workflow, not buying an agent.

2. Ask the team for the file: thirty historical cases with their correct outputs. Give them a week. What comes back tells you more about your readiness than any vendor evaluation.

3. Recalculate the business case as cost per completed run. Vendor per call price, divided by their own quoted success rate, plus the checking minutes at a loaded rate. Most cases do not survive this, which is the point of doing it first.

4. List every action the agent could take and mark each reversible or not. Everything in the not column becomes propose and commit before a line is written.

5. Ask any vendor for pass to the power of five rather than pass at one. Their reaction is worth more than the number.

Close

The demo always ends after the first run. That is the tell. It is not dishonest, it is just how demos work: everyone shows their best attempt.

Your company does not get the best attempt. It gets the average attempt, every day, forever, with someone downstream quietly checking the output because they do not trust it yet and nobody has told them when they can stop.

Ask to see the second run.

The expensive part of software has never been the build. If you want someone to look at your agent roadmap and tell you which two of the twelve are worth doing, my calendar is here.

In June 2026 a team ran 2,398 queries from 56 brand accounts across ChatGPT, Google AI Mode, Gemini and Perplexity. Same questions, every day, seven consecutive days. 530,875 citations. 181,225 distinct URLs. 55,387 domains. Ask an engine the identical question twice, twenty four hours apart, and 69 percent of the sources it hands back have changed.

That screenshot is not a position. It is a photograph of one roll of the dice.

The thesis

A citation is not an asset, because it does not persist. What compounds in AI search is your share of the pool the engine samples from, not any individual page it happened to pull on a Tuesday. The pool is third party text you mostly do not own and cannot edit. That is the moat, and it is a moat precisely because you cannot buy it in a quarter.

This cuts against what the category is currently selling. The 2026 consensus is to restructure your own pages for extraction, add the schema, write answer shaped paragraphs, then watch a citation counter go up. I think the counter is mostly noise, and the page work is the smaller half of the job.

What the churn actually looks like

The 69 percent is the pooled figure. Split it by engine and the picture sharpens.

Gemini turns over 88.3 percent of its cited sources overnight. ChatGPT, 79.2 percent. Google AI Mode, 75.9 percent. Perplexity is the outlier at 44.4 percent. Follow the same sources across the full week and the survival rates are worse: of everything cited on day one, 22.3 percent was still being cited by Gemini on day seven, 33.4 percent by ChatGPT, 33.7 percent by AI Mode, 48.6 percent by Perplexity. Sources that appeared on all seven days ranged from 0.4 percent to 11.1 percent depending on the engine.

Now put that next to a second number. Semrush analysed 126 million United States AI prompts between January and April 2026 and found ChatGPT cites an average of 15 sources per response. Gemini cites 3.

Do the arithmetic on ChatGPT. Fifteen slots, 79.2 percent overnight churn. Roughly twelve of tomorrow’s fifteen seats are filled by something that was not there today. Three carry over. If you were cited yesterday, your prior on being cited tomorrow is not a plan, it is a coin.

The number that contradicts all of that

Here is the part nobody puts next to the volatility data.

The same Semrush index found that in News and Media, the top three brands hold 82.9 percent of all category visibility. In Consumer Electronics, the top three hold 76.9 percent. And 36 brands held top one hundred visibility across all four platforms, every single month, for four straight months. Semrush calls them the Universal 36.

So both of these are true at once. At the URL layer, 69 to 88 percent of everything is replaced overnight. At the entity layer, almost nothing moves for a third of a year.

That gap is the whole argument. The engines are not unstable about who matters. They are unstable about which document proves it. Every night they re-sample the corpus, pull a different set of pages, and arrive at substantially the same set of names. Optimising the document is optimising the part that gets thrown away. Occupying more of the corpus is optimising the part that does not.

Where the stable part comes from

If the pool is what matters, the question is what the pool is made of.

Omniscient Digital published an analysis on 8 January 2026 covering 23,387 unique citations across 240 branded queries and five engines. When a brand is named in the question, 48 percent of citations came from earned media: 16 percent editorial, 11 percent forums and social, 11 percent review sites, 10 percent directories and reference. Commercial content took 30 percent. Content the brand actually owns took 23 percent.

And it moves hard by intent. On customer review questions, 82 percent of citations were earned media. On functionality and integration questions, owned content took 50 percent. So your own site does win, in the one place a buyer has already decided to consider you and is checking whether the thing plugs into their stack. In the place where they are deciding whether to consider you at all, you are outnumbered better than two to one by editorial, forums and review sites.

Seer Interactive approached it from the other end: 300,000 keywords across finance and software, nearly 600,000 related questions generated, 10,000 of them run through the GPT-4o API against measured rankings. Google position correlated with LLM brand mentions at roughly 0.65. Bing came in lower, around 0.5 to 0.6. Backlinks, which the entire optimisation industry has priced for twenty years, were weak to neutral.

Read those two together and the mechanism is unglamorous. The model is not forming an independent judgement of your authority. It is inheriting Google’s judgement of your relevance and then sampling a body of text that is mostly other people talking about you.

Where this breaks

The volatility study is a vendor study. Seven days, one month, 56 self selected brand accounts, published by a company that sells mention tracking. A firm in that position has an interest in churn looking frightening. In its favour, the finding is inconvenient for its own product: if 69 percent of sources change overnight, a weekly citation dashboard is largely measuring the weather.

The concentration figures cut against me harder. News and Media at 82.9 percent and Consumer Electronics at 76.9 percent are categories where brand recall was already everything. Look at the categories most readers of this actually sit in and the top three hold 41.4 percent in Finance and 42.2 percent in Industrial. My entity stability claim is strongest exactly where it matters least.

Then there is the composition of the Universal 36. YouTube, Google, Reddit, Amazon, Facebook, Apple, Walmart, Disney, Nintendo. None of those built a position in AI answers through an earned media programme. They are household names and platforms, and reading them as proof that a moat is buildable is survivorship reasoning of the worst kind. I am using them as evidence that entity level stability exists, not that you can purchase it.

And the honest limit on causation: Seer found Google rank correlates at 0.65 and backlinks near zero. That is at least as consistent with “just rank well and stop theorising” as it is with my argument. I cannot show you that corpus share causes mention share. I can show you that the thing you are currently being sold, presence in a given answer on a given day, has a documented shelf life of about a day.

Finally, three years is a long claim about a retrieval stack that did not exist three years ago. If the engines move to licensed corpora or paid placement, the earned pool stops being the input and this argument expires with it.

What I would do on Monday

1. Stop reporting citation count. Report share of pool: across a 30 day window on a fixed question set, what percentage of all cited domains were yours. A count without a denominator, at 69 percent churn, is a lottery result.

2. Run the same twenty buyer questions every day for fourteen days and log every domain cited. Then count the domains that appear on ten days or more. That short list, not your keyword competitors, is who you are actually competing with.

3. Split spend by intent. Owned content takes 50 percent of citations on functionality questions. On review questions earned media takes 82 percent and everything else splits the remaining 18. Put the product detail on your own site and stop expecting it to win a comparison.

4. Count your earned corpus honestly. How many pages published in the last eighteen months describe what you do in somebody else’s voice. If the answer is under twenty, you do not have a visibility problem, you have a coverage problem.

5. Check your Google position on those same twenty questions before you buy anything. The correlation is 0.65 and it is the cheapest input on this list.

I have been writing software for thirty years and I have never seen a measurement culture form this fast around a number this unstable. When a team shows me a month on month citation chart, I now ask one question before I look at it: how many times was each question asked. The answer is almost always once, on the morning the report was generated.

That is not a trend line. That is a screenshot with a chart drawn around it.

I spend a lot of my week inside companies trying to work out why a number went up. If someone has just shown you an AI visibility chart and nobody in the room can tell you the sample size, book a call before you buy the tool.

The room goes quiet. Someone names a person who is not in the meeting. Someone else names a different person. Nobody names a date. We have just spent forty minutes agreeing a list of deliverables that everyone nodded at, and the first question about how that list gets decided empties the room.

That silence is the project. The deliverable list is not.

The argument

A scope document that lists what you will build is describing the output. Output is not what moves. Decisions move. So the document should list the decisions nobody has made yet, put one name against each, put a date on each, and state what the team does if the date passes. One page. Four sections. Not one deliverable in it.

That is the opposite of the standard advice. Every statement of work template, every work breakdown structure, every scope creep guide published this year tells you the same thing: describe the deliverables in more detail, write tighter acceptance criteria, and bolt a change control process on the side. I think that advice is aimed at the wrong failure.

The numbers say the failure moved

PMI published Pulse of the Profession on 22 May 2026. Ninety seven percent of project professionals managed at least one complex project in the prior year. Thirty one percent of those complex projects failed to achieve their originally intended scope of benefits, against 12 percent in the 2024 report. That is two and a half times in two years.

Now put that next to delivery performance, which barely moved at all. Arcidiacono’s 2025 update in the PM World Journal, published January 2026, compares two multi year datasets on IT project outcomes. The 2015 to 2017 window: 29 percent successful, 52 percent challenged, 19 percent failed. The 2020 to 2024 window: 31 percent successful, 50 percent challenged, 19 percent failed. Nine years, two points of movement, and the failure column did not move by a single point.

So delivery discipline is flat and benefit capture fell off a cliff. Whatever broke, it did not break in the part of the project that a deliverable list controls.

The PMI report says where it broke. Under value and alignment gaps, which hit 61 percent of complex projects, the largest single item is stakeholder decision delays at 34 percent. Budget overruns are 12 percent. Profit loss is 9 percent. Decision delay is bigger than both of those put together and then some. Every scope artefact on the market is built to police budget and deliverables. Almost none of them is built to police decision latency, which is the variable doing the damage.

There is a second number in the same report that I keep coming back to. Forty five percent of project professionals name multiple stakeholders across functions and geographies as a driver of complexity. Twenty seven percent of executives name the same thing. An eighteen point gap between the people who sign the scope document and the people who have to live inside it, on exactly the variable that produces decision delay. The document is written by the party that cannot see the constraint.

Speed is not the enemy of quality

The usual objection to all of this is that forcing decisions onto dates produces bad decisions. McKinsey surveyed more than 1,200 managers for its 2019 work on decision making. Respondents who reported that decisions were made fast were 1.98 times more likely to report that those decisions were high quality. Not a tradeoff. A correlation running the other way.

The same survey found 61 percent of managers saying at least half the time spent making decisions is ineffective, and only 34 percent saying their organisation made cross cutting decisions that were both good and timely. Cross cutting is the category every project sits in.

Read those two findings together and the change control process looks different. A change control board exists to slow a decision down so it can be examined. The evidence says slow decisions are not better decisions, they are just later ones, and the lateness is the thing killing benefit capture. The safeguard is the failure mode wearing a lanyard.

The one page

Four sections. It fits on a page because if it does not fit on a page nobody reads it in week six, which is the only week it matters.

One. The open decisions. Every decision not yet made that would change what gets built. One line each, written as a question with at least two real answers. Not “confirm the data model”. Instead: “do we write to the existing customer table or stand up a new one”. If it has only one answer, it is not a decision, it is a task, and it belongs somewhere else.

Two. One name. A person, not a function, not a committee, not “the steering group”. If two names appear, the decision is actually two decisions and you have not finished writing it. This section is where most kickoffs fall apart, because naming an owner in writing is a commitment and the room can feel it.

Three. The date and the default. The date the answer is needed, and the sentence describing what the team will build if that date passes without an answer. The default is the whole trick. It converts a blocked decision into a made one. It also removes the incentive to stall, because stalling now has a named consequence that the owner has already read.

Four. What is already out. The things explicitly not in scope, written as decisions that have been made, with the person who made them. This is the section that survives a change of sponsor, which is when every undocumented exclusion comes back.

There is a loop hiding inside the first section and it is the part that costs money. An unnamed decision does not stop the team. The team assumes an answer, builds on the assumption, and the assumption surfaces at a review where it is expensive to unwind. Then the same decision goes back into the queue, still unnamed. I have watched the same decision arrive at three consecutive steering meetings, get discussed each time, and get owned none of them.

Where this breaks

Four honest limits.

The change control evidence cuts against me. PMI’s own research says projects without formal change management are 35 percent more likely to exceed cost or miss deadlines. That is a real finding and I am arguing for less process at the point of change, not more. My position is narrower than it sounds: the change process is fine for changes, and useless for decisions that were never made in the first place. Those are different objects and most documents treat them as one.

The PMI decision delay figure is self reported by project professionals, who have an obvious interest in locating the blame one level above themselves. I read a survey of practitioners as evidence about executives. That is a weakness and I am not going to pretend otherwise.

The default mechanism assumes the team can actually proceed on a default. In regulated work, in anything touching patient data or client money, there are decisions where the only lawful default is stop. For those the section still helps, because it makes the stop visible and dated before it happens, but the mechanism is doing less work.

And the honest one about me. A one page document is easy to write and easy to like. It does not survive contact with a sponsor who will not put their name on anything, and there are organisations where the refusal to name an owner is the culture rather than an oversight. In those places this document does not fix the project. It just tells you early, in writing, that the project is not going to work, which is worth something but is not what I sold.

What I would do on Monday

1. Take the scope document for your largest running project. Count the deliverables. Then count the decisions that are named, dated and owned by one person. If the second number is zero, you have a wish list, not a scope.

2. Write the open decisions down as questions with two real answers each. Most teams find between six and twelve on a project of any size. Anything with one answer moves to the task list.

3. Put one name and one date against every line, out loud, in a meeting, with the names in the room. Do not accept a function. Do not accept two names.

4. Write the default for each. The sentence starts “if we have no answer by this date, we will build”. Circulate it. The objections you get back within a day are the real risk register.

5. Bring the page to every steering meeting and change only the dates. If a date moves twice, escalate the owner, not the decision.

The kickoff where I ask who has to say yes is still quiet. That has not changed. What changed is that I now write the silence down, put a date on it, and hand it back to the room before anyone has spent a dollar building anything.

I spend most of my week inside delivery organisations where the build is fine and the decisions are late. If you are about to sign a scope document that names forty deliverables and no owners, book a call first.

That is the promise inside almost every buyout model signed this year. Bain’s 2026 Global Private Equity Report, published on 23 February, put a number on the new regime: a deal now needs 10 to 12 percent EBITDA growth a year to earn the benchmark 2.5 times return over five years, against about 5 percent through the 2010s. Cheap debt and multiple expansion stopped doing the work, so the operating plan has to. Seventy six percent more earnings in five years, from the same company, run by mostly the same people.

Now halve it.

The thesis

A value creation plan is a bet on how much change a company can absorb, and diligence never measures that. It measures the market, the margins, the contracts and the code. It does not measure the rate at which this organisation has historically turned a decision into a finished change.

So sponsors underwrite a plan that needs three years of change delivered in eighteen months, launch all of it at once to beat the clock, and replace the chief executive when it does not land. The industry calls that an execution problem. I call it an underwriting error. The capacity was never there, and nobody checked.

Halving it

Alvarez & Marsal published its European private equity value creation report on 20 May 2026. Two hundred fund investors and C-level executives across ten countries, interviewed by Statista Q in February. Two findings sit side by side in the release, and I do not think they were meant to be read together.

First, 58 percent of sponsors now deploy resources within the first 100 days, double the 29 percent recorded a year earlier. Second, 65 percent say they have achieved less than half of the value targeted in plans developed over the past two years.

The response to plans that under deliver is to start them sooner. Starting sooner means more initiatives running at the same time. More initiatives running at the same time is the one thing an organisation with finite change capacity cannot carry.

Now run the arithmetic on “less than half”. Take a plan written to the new regime: 12 percent a year, 76 percent over five years. Capture half the value and you land 38 percent. Compounded, that is 6.7 percent a year.

A half captured plan in the 12 percent world delivers something much closer to the 5 percent world. The sponsor paid a 2026 entry price for 2015 performance, and the model will not show it until year three.

Where the capacity actually sits

Change capacity is not headcount. It is the small number of people who run the business today and also own a piece of the change. In most mid market companies I have worked inside, that is the chief executive, the finance lead, one or two operations heads and whoever runs technology. Five or six people.

Every initiative in the plan lands on them. A pricing reset lands on finance and sales. A procurement programme lands on operations and finance. A systems consolidation lands on technology and finance. An AI programme lands on everyone, and on technology twice. Notice which name keeps appearing.

The people underneath them are already full. Gartner’s HR research found the average employee experienced 10 planned enterprise changes in 2022, up from 2 in 2016, while willingness to support change fell from 74 percent to 43 percent. Its April 2025 survey of more than 2,850 employees found 79 percent have low trust in change, and only 32 percent of business leaders report healthy change adoption. A new owner does not arrive at a rested organisation. It arrives at one that has already been changed ten times this year and has stopped listening.

Then the owner column fails, and the sponsor does the one thing it can do quickly. AlixPartners’ 11th annual private equity leadership survey, published 25 March 2026, covered 427 executives. 65 percent of PE firms report chief executive turnover during the hold. 83 percent say unplanned turnover lengthens the hold, and nearly half say it reduces returns. Bain has buyout holding periods at around seven years at exit, against five to six from 2010 to 2021.

Replacing the chief executive is a capacity fix applied after the fact, at the most expensive moment available, and it resets the owner column to zero.

How it compounds

The failure runs in a straight line. The plan is written to the 12 percent requirement. Diligence never asks whether the company can carry it, so every initiative is launched inside the first 100 days. The same five or six people end up owning all of it. Less than half the value lands. The chief executive goes, the hold stretches, and the new leadership inherits a half finished portfolio of change with a shorter clock.

The alternative path is not slower. It is ordered. Test the capacity before close, sequence the initiatives to fit it, and the value lands inside the hold instead of in the next owner’s model.

What I actually look at

On 1 September I argued here that technical diligence audits the codebase and ignores the delivery system. This is the same leak, one level up. The delivery system is where a company’s change capacity becomes visible, because almost every initiative in a modern value creation plan eventually becomes a change to a system that someone has to build, configure or migrate.

So when I read a value creation plan now, I skip the initiative column and read the owner column. I count distinct names. Then I ask the company one question: in the last 24 months, how many cross functional changes did you start, and how many did you finish? The ratio between those two numbers is the most honest estimate of change capacity I know. I have not seen it in a standard diligence pack.

After 30 years of shipping software, the pattern I trust least is a plan where the finance lead’s name appears against more initiatives than there are quarters in the first year.

Where this breaks

The Alvarez & Marsal respondents blame the world first, not their plans. 62 percent name geopolitical volatility as the top obstacle to value creation and 58 percent name tariffs and inflation. Some of that “less than half” is macro, not capacity, and a self-reported survey of 200 European respondents cannot separate the two. I am reading a pair of numbers in a way the authors did not.

The Gartner figures measure attitude, not throughput. Low trust in change does not prove an organisation cannot deliver it. Some deliver anyway.

The clock is real. With seven year holds and a 12 percent requirement, a sponsor who sequences may lose more to time than it saves in capture. My answer is that sequencing is ordering, not waiting, but I concede the trade is tighter than this post makes it sound.

Some sponsors do test for this. Large operating partner benches and management assessments exist for a reason. The limit is that a management assessment grades the individual, not the throughput of the system around them. A strong chief executive with a full owner column is still full.

And the strongest objection: sometimes the chief executive change is the right capacity fix and the plan was fine. Nearly half of AlixPartners’ respondents say unplanned turnover reduces returns. The other half do not.

What I would do on Monday

1. Add one line to diligence: cross functional changes started versus finished over 24 months. Ask for the list, not the number.

2. Read the owner column. If one name owns more than three initiatives, the plan is a wish list.

3. Launch no more initiatives in the first 100 days than the company finished in its best six months. Put the rest in a dated queue.

4. Underwrite at the capture rate, not the target. If the evidence says half, model half, and let the price move.

5. Hire the second line before close. The finance and technology owners you will otherwise hire in year two, after the chief executive has gone.

1.12 to the fifth is 1.76. Half of it is the 2010s.

The gap between those two numbers is not strategy. It is how many things one finance lead can finish in a year. Nobody asked.

I build software for private equity, which means I spend a lot of time inside the systems a value creation plan has to run through. If you are underwriting one this quarter and nobody has counted the owner column, book a call.

Heavy AI users are producing about 25 percent more code than their own past selves. Duplicated code is up 81 percent, from 40.3 duplicated lines per million changed lines to 73.0. GitClear pulled both out of 623 million code changes recorded between 2023 and 2026.

Volume rose a quarter. Copying rose four fifths. Copying is growing 3.2 times faster than the output it sits inside, which means the extra duplication is not a by-product of writing more. It is a change in what is being written.

The thesis

AI did not automate engineering. It automated production, left verification exactly where it was, and the cost of verification scales with how much code you already own. So the price of people who can produce code is falling towards the price of a subscription, and the price of people who can look at plausible code and say what is wrong with it is going up. Thirty years at a keyboard just became an appreciating asset.

That is contestable, and most of the market disagrees with me. The prevailing 2026 argument is that models now write senior-level code, so you need fewer seniors and more cheap operators driving agents.

The grid nobody reads as a grid

Stanford’s software engineering productivity programme, run by Yegor Denisov-Blanch, has measured nearly 100,000 engineers across more than 600 companies. Split the work four ways and the gains separate hard.

Greenfield, low complexity: 30 to 40 percent. Greenfield, high complexity: 10 to 15 percent. Brownfield, low complexity: 15 to 20 percent. Brownfield, high complexity: zero to 10 percent, with a material share of teams going net negative. Net of rework, across everything, the average lands around 15 to 20 percent.

Read that as a table and it is a productivity study. Read it as a grid and it is a depreciation schedule on your own estate. Every line of code your company already owns lowers the return on the next line a model writes for you. Not because the model is worse at old code. Because someone has to decide whether the new thing is consistent with the old thing, and that decision costs roughly the same whether the change took six hours to write or six seconds.

That is the whole move. Production time collapsed by an order of magnitude. Verification time did not move at all. When one side of a two-sided process gets 20 times cheaper, the other side becomes the entire cost, and whoever can do the other side becomes the entire constraint.

It gets worse with scale. The same programme reports models losing roughly half their performance at around 32,000 tokens of context. Your enterprise codebase is not 32,000 tokens. The context that decides whether a change is correct lives in people, in decisions made four years ago that were never written down, in the reason a function looks stupid.

What the numbers do next

GitClear’s Maintainability Gap research, January 2026. Refactoring fell to 3.8 percent of changed lines in 2026, from 21 percent in 2022. Copy and paste rose to 15.7 percent in the first half of 2026. Function reuse is down 35 percent since 2023. Long-term maintenance activity is down 74 percent since 2023. The supply of context maintenance collapsed at exactly the moment demand for context rose. That is the shape of a system where the cheap half got cheaper and the expensive half got skipped.

DORA published its ROI of AI-assisted Software Development report in May 2026, led by Nathen Harvey at Google Cloud, building on research covering close to 5,000 technology professionals. Their model for a 500-person engineering organisation at 176,000 dollars fully loaded: 8.4 million dollars of investment, 11.6 million dollars of first-year value, a 39 percent first-year return, payback around eight months. In the same model, moving change failure rate from 5 percent to 6 percent costs 344,000 dollars.

Run those two numbers against each other. 344,000 dollars is 3 percent of the entire first-year value, per single point of change failure. Ten points, which is not a wild outcome for a team merging on plausibility, costs 3.44 million on a straight line. That takes 11.6 million of value down to 8.16 million against 8.4 million of spend. The programme goes negative. The whole 39 percent return sits inside ten points of a metric most engineering organisations have never baselined.

Stack Overflow’s developer survey, fielded 29 May to 23 June 2025, 49,009 respondents. 84 percent use or plan to use AI tools. Trust in accuracy fell to 33 percent and only 3 percent trust the output highly. The single biggest frustration, at 66 percent, is “almost right, but not quite”. 45 percent say debugging AI code takes longer than expected. Near universal adoption, collapsing trust, and the top complaint is a category of error that only a person holding context can catch.

I have been writing software for 30 years. For most of that time the scarce thing in any delivery organisation was someone who could make the machine do the thing. That stopped being scarce, and I noticed it in my own week before I saw it in anyone’s dataset. Almost none of my hours now go on producing something that works. They go on deciding whether something that already works should exist in this system.

So I changed the first question I ask. Walking into a delivery organisation I no longer ask for velocity. I ask what share of merged changes last quarter had a reviewer who had touched that file before. Most places cannot answer it. The ones that can are the ones where AI is actually paying.

Where this breaks

METR ran a randomised controlled trial on experienced open-source developers working in repositories they knew well, published 10 July 2025. They were 19 percent slower with AI and believed they were 20 percent faster. That is the best argument against me on the table: the veteran is the person the tool slows down. METR’s follow-up, reported 24 February 2026 across 57 developers, 143 repositories and more than 800 tasks, measured minus 18 percent for returning participants and minus 4 percent for newly recruited ones. The slowdown concentrated in the people holding the most repository context. My reading is that they were slow because they were doing the verification, and elapsed task time does not price a defect that never shipped. That is a reading. It is not what the study measured.

Second, METR withdrew that design. They now say the data gives only very weak evidence on productivity, partly because 30 to 50 percent of developers declined to submit tasks they did not want to attempt without AI, and recruitment kept failing because people would not give up their tools at any offered price. I am leaning on a study whose own authors have told you to lean on it less.

Third, GitClear counts code, not outcomes. Duplication is a maintainability proxy. A product rewritten inside two years does not care about maintainability, and if your estate has a short half-life none of this applies to you.

Fourth, the ten-point extrapolation on DORA’s 344,000 dollars is mine, not theirs. They priced one point. Failure cost is almost certainly non-linear and I do not know in which direction.

Fifth and strongest: greenfield is real. A company whose entire codebase is eighteen months old sits in the 30 to 40 percent corner of that grid, and for them the right hire genuinely is a fast operator driving agents. My argument is about accumulated context. If you have not accumulated any, ignore it.

And one against my own vanity. Thirty years is not judgement. Plenty of long-tenured engineers hold deep context about exactly one system and nothing transferable to any other. Tenure is a proxy and a weak one. What is repricing is the ability to reject a plausible thing, and some people have that at twenty six.

What I would do on Monday

1. Place yourself on the grid before you sign the seat contract. Split last quarter’s merged changes into greenfield and brownfield, low and high complexity, then weight the Stanford ranges against your actual mix. Most enterprises will find the majority of their work in the bottom right corner, where measured gain is zero to 10 percent. Budget against that, not against the vendor deck.

2. Baseline change failure rate this week. It is the one number in DORA’s model that can take a 39 percent return negative, and you cannot reconstruct a clean pre-AI figure once the seats are live.

3. Track duplication growth as a ratio to output growth, monthly. If duplication is outrunning volume the way it did 3.2 to one in GitClear’s data, you are buying throughput and selling coherence, and that invoice arrives in eighteen months.

4. Fund review capacity as a role, not as an unpaid tax on the same three people. If verification is the bottleneck, staffing it as an afterthought is the same error as one tester per hundred engineers.

5. Change what you test for in interviews. Stop asking candidates to produce a correct function. Hand them a plausible 200-line diff against a system you briefed them on twenty minutes earlier and ask what is wrong with it. That separates people faster than any exercise I have used.

Close

Back to those two numbers. 25 percent more output. 81 percent more duplication.

One of them is what your AI budget was bought to produce. The other is what it gets judged on in three years. The gap between them is not a tooling problem, and no seat licence closes it.

It gets closed by someone who reads the diff and says no.

I spend most of my week inside delivery organisations working out which changes should have been stopped. If you are about to buy another thousand seats and nobody has baselined your change failure rate, book a call first.

Writing a real one takes most of a week. You read everything the company has published. You sit with the people who do the actual work. You find the thing nobody has said out loud yet, and you write the plan that follows from it. That plan is the work. It is the most valuable thing I produce in any given month. And the convention in this industry is that I hand it to someone who has not yet decided to buy anything, for free, and then ask to be paid for the part that comes after.

I stopped doing that. Not out of principle. Out of arithmetic.

The thesis

A proposal was never a document. It was a costly signal. Thirty three hours of senior attention that could not be faked was the proof that you had understood the problem, and the buyer read the effort, not the pages. Generative AI made that effort cheap to counterfeit. A signal that is cheap to counterfeit carries no information. So the proposal still exists, it still gets written, and it no longer tells the buyer anything.

The arithmetic that made me stop

Loopio published its sixth annual RFP Response Trends and Benchmarks Report on 11 March 2026, from 1,533 response management professionals across more than seventeen industries. Read the numbers next to each other rather than one at a time.

The average organisation submitted 166 RFPs last year, up from 153. The average response took 33 hours, down from 35. Nearly 80% of teams used generative AI during 2025, ten percentage points more than the year before, and 62% used it to write specific answers.

So: 153 responses at 35 hours is 5,355 hours. 166 responses at 33 hours is 5,478 hours. The technology cut the unit cost of a proposal by 5.7% and the volume went up 8.5%. The total time these teams spent writing proposals did not fall. It rose by 123 hours.

That is Jevons paradox running inside a sales organisation. Make a thing cheaper to produce and you produce more of it, and the saving never reaches the P&L.

Now put the win rate on top. Loopio’s benchmark average across 2019 to 2026 is 45%. The current figure is 39%. Adoption of the tool that was supposed to help crossed 80% in the same window that the hit rate fell six points below its own seven year average.

Take the 61% that lose. That is 3,341 hours a year, in one company, producing documents nobody bought. At a 1,950 hour working year it is 1.7 full time people whose entire job is losing. Half of these teams already name bandwidth as their number one problem, which is the correct diagnosis of the wrong system.

What replaces it

The rule is one line. If the problem is already agreed in writing, I will quote a fixed scope for the first block of work. If it is not agreed, the diagnosis is the engagement, it is priced, it runs two weeks, and it ends with a document the buyer owns whether or not they ever hire me again. If they will not pay for the diagnosis, I decline the delivery, because a buyer who will not fund two weeks of thinking will not fund six months of change.

This is not a pricing trick. It changes who is holding the risk. A proposal asks a room to commit to a large decision on the strength of a document written by the person being paid. A paid diagnostic asks one person to commit to a small, bounded, reversible decision, and it produces the evidence the room needs to make the big one.

Why the room is the real constraint

Matt Dixon and Ted McKenna analysed 2.5 million recorded sales conversations for The JOLT Effect, using machine learning to isolate roughly 8,300 factors behind wins and losses. Between 40% and 60% of lost deals end in no decision. The driver is not preference for the status quo. It is fear of failure: the buyer wants the thing and cannot carry the personal risk of being the one who chose wrong.

A proposal does nothing about that. It is an argument for a decision, delivered to someone who is afraid of the decision. More pages make it worse.

Gartner surveyed 632 B2B buyers in August and September 2024 and found 74% of buying teams showing unhealthy conflict during the decision. Delainey Kirkwood put the group at five to sixteen people across as many as four functions. Groups that reach consensus are 2.5 times more likely to report a high quality deal. Content aimed at one individual reduced consensus by 59%.

That last figure is the one that ended proposals for me. The document is written to a champion. The decision is made by a committee that is already fighting. A paid diagnostic is the only format I have found where the committee is in the room while the conclusion is being formed, so the output arrives with their fingerprints on it instead of mine.

Where this breaks

Government and regulated procurement runs on unpaid written responses and no amount of principle changes that. If your revenue depends on tenders, you write tenders, and the only available move is to qualify harder before you start.

Loopio’s sample is people whose job title is response management, so it over represents companies where the RFP is the sales motion. The 33 hours may be elapsed time rather than person hours, and with nine contributors on an average response the real number could be far worse than the one I used. I have taken the conservative reading.

The win rate fall from 45% to 39% is not proof that AI caused it. More responses to lower quality opportunities would produce the same line.

And the honest limit: charging for the diagnosis requires enough demand to say no. I could not have done this at thirty. If you need the next engagement to make payroll, you will write the proposal, and you should. The method is downstream of the pipeline, not a substitute for it.

What I would do on Monday

1. Add up last year’s losing proposals. Count the hours, name the people, and put the number in front of whoever signs their invoices.

2. Write the two week diagnostic as a real product. Fixed fee, fixed dates, named output, and a sentence saying the client owns it outright.

3. Change the first question on the call from what do you need to who has to agree, and what are they arguing about. Then price the argument.

4. Stop sending anything longer than two pages to a buyer who has not paid you. If they need thirty pages to decide, the problem was never described properly.

5. Keep one exception list. Tenders, framework agreements, and the two or three relationships worth the free work. Everything else follows the rule.

The close

I still do that week of work. I still read everything, sit with the people who do the job, and find the thing nobody has said out loud. That part was always the value and it still is.

The only thing that changed is that it now has a price, a start date, and an invoice.

I spend most of my week being paid to find out what is actually wrong before anyone commits to fixing it. If you are about to ask three firms for a free plan, book a call instead.

In the first quarter of 2026, Gartner surveyed 12,004 employees and managers across 40 countries. Of the ones whose employer had already given them enterprise AI access, 88% were using their own personal AI tools for work anyway.

That is a verdict. The people the centre exists to serve have already priced its output and routed around it.

The argument

A centre of excellence can only supply the things it controls. It controls tools, standards and approvals. It does not control workflows. Workflow redesign is the one variable that separates the companies getting AI value from the companies getting AI invoices. So the centre is structurally incapable of producing the outcome it was funded to produce. It is not a capability function. It is a permission desk with a budget.

What the numbers actually say

McKinsey ran its global AI survey between 4 May and 8 June 2026: 1,719 respondents across 97 nations, 36% of them at companies above one billion dollars in revenue. Thirty-seven percent attribute at least some EBIT impact to AI, which is about the same share as the year before. Six percent are high performers who attribute 5% or more of EBIT to AI use. The thing that separates them is not their model choice. Nearly three quarters of the high performers report fundamentally redesigning workflows, against 25% of everyone else.

Now put the spend next to it. BCG’s AI Radar 2026, published January 2026, surveyed 2,360 executives including 640 CEOs across 15 markets. AI investment is set to roughly double this year, from about 0.8% of annual revenue to about 1.7%. On a one billion dollar business that is a move from roughly 8 million dollars to roughly 17 million dollars in twelve months. In the same survey, 94% of CEOs said they would keep investing even if 2026 initiatives underperformed.

The spend doubles. The share of companies reporting any EBIT impact does not move. The marginal nine million dollars bought no measured change in the odds of impact.

Here is why. That money is being spent by a function whose authority stops at the tool. A centre of excellence can buy a model, set a standard, run an evaluation, approve a vendor and publish a prompt library. It cannot tell the claims team to stop doing the second review. It cannot collapse three handoffs into one. It cannot change who signs. Those are the acts that move a line in the P&L, and they belong to whoever owns the line.

MIT’s NANDA study on the GenAI divide ran from January to June 2025 across 52 structured interviews, 153 survey responses and more than 300 publicly disclosed initiatives. Its finding on organisational design is one sentence long and almost nobody quotes it: the organisations that succeeded sourced their AI initiatives from frontline managers, not central labs. The same study found pilots built through external partnerships reached full deployment 66% of the time, against 33% for pilots built internally. Task-specific internal tools reached production 5% of the time. General purpose tools bought off the shelf reached production 40% of the time.

A centre of excellence is an internal build shop staffed by people whose standing depends on internal builds.

I stopped opening engagements with the AI strategy deck about two years ago. I ask for two artefacts instead: the list of people who can change a business process without seeking approval, and the central team’s intake queue. The list usually fits on one hand. The queue usually runs to dozens. Everything that goes wrong over the following year is already legible in that ratio, and thirty years of writing software has not shown me a tool that repairs it.

The mechanism

A use case is found by the person doing the work. Everything after that is decided by one question: who is allowed to change the process. If the central team holds the tools and nothing else, the idea goes onto an intake form, waits for a monthly rubric, and comes back as an approved tool sitting on top of an untouched workflow. That is a pilot with no line in the P&L. If the unit that owns the process holds the decision, the workflow is redesigned first and tooled second, and the impact lands in that unit’s own numbers. Same idea, same model, same vendor. Different owner, different outcome.

Three things that have already been measured

The shadow economy was counted twice, a year apart, and it grew. In mid 2025 MIT found that only 40% of companies had bought an official LLM subscription while workers at more than 90% of surveyed companies were using personal AI tools for work. The obvious reading was a procurement gap: buy the licences and the shadow disappears. Gartner’s first quarter 2026 data says it did not. Among employees who had enterprise access, 88% still used personal tools, and those hybrid users were 1.7 times more likely to report significant time savings than the compliant ones. The shadow is not a licensing failure. It is people avoiding the latency of the approved path, and they are right to, because it makes them faster.

Uniform governance is the only kind a central body can produce. On 26 May 2026 Gartner predicted that 40% of enterprises will demote or decommission autonomous AI agents by 2027 after governance gaps surface in production. Shiva Varma, the analyst behind it, put the cause plainly: enterprises treat agent governance as binary, either locked down or fully trusted, and that is the root of the failure. Over-restriction of simple agents, in his words, slows delivery and drives shadow development. A central body cannot hold the context of forty different workflows, so it writes one rule and applies it to all of them. That is not a staffing problem. It is what centralisation is.

The talent leaves. On 13 May 2026 Gartner predicted that half of enterprises without a people-centric AI strategy will lose their top AI talent to competitors by 2027. In the same research, 27% of executives reported having a comprehensive AI strategy and 20% of leaders believed their workforce was AI-ready. A centre of excellence takes the scarcest engineers in the company, concentrates them in one team, and hands them a queue of intake forms to triage. There is no faster way to lose them.

Where this breaks

Four places, and I will not pretend they are small.

Comparing BCG’s spending figure with McKinsey’s impact figure is not a panel study. Different samples, different questions, different months. The doubling and the flat line are directional, not a measurement, and anyone who wants to reject that inference is entitled to.

The MIT sample is small, partly recruited at conferences, and the fieldwork is now more than a year old. Its 95% headline has been over-read by almost everyone who has quoted it, including people arguing my side. The frontline-managers finding is qualitative and it is one sentence.

Regulated industries are a real exception. Model risk management in a bank, clinical safety in a hospital, anything touching defence. There a central body is a legal requirement and removing it is not an option. In those settings the right question is how fast the central body can say yes, not whether it should exist.

And the strongest objection, which I hold myself: some things must be central. Identity, the evaluation harness, data contracts, model procurement, the audit trail. That is a platform team, and platform teams belong in the middle. The failure mode is not centralisation. It is merging the platform team with the permission desk, so that the group who builds the substrate is also the group who decides who is allowed to use it. Below roughly two hundred people none of this applies, because there is no centre to dismantle.

What I would do on Monday

1. Write down every person who can change a business process without asking anyone. If that list does not contain a person who touches the workflow you want AI inside, nothing you buy this year will show up in the P&L.

2. Split the central team on paper into two functions. Platform keeps identity, evaluation, data contracts and model procurement. Permission is deleted, and the decision moves to whoever carries the P&L line.

3. Set a dollar threshold under which no approval is required, only a note afterwards. Pick the number that makes your CFO slightly uncomfortable and publish it where everyone can see it.

4. Measure approval latency. Count the days from a named person having an idea to that same person being allowed to try it. If the answer is longer than five days, the 88% is your number too.

5. Default to buying. Sixty-six percent against thirty-three percent. Put the ratio into the intake criteria and make anyone proposing an internal build argue against it in writing.

The intake form is not the problem. The form is the confession. It says the company has decided that the risk of one person trying the wrong thing is larger than the risk of nobody trying anything, and most of the money burned in the last three years sits inside that sentence.

Keep the platform. Keep the evaluation harness. Burn the form.

I spend most of my week inside delivery organisations working out why an approved tool never reached the P&L. If that is the gap you are looking at, book a call.

That page was the best converting page on most of those sites for about fifteen years. It was the middle of the funnel, written down.

I have not watched one of those pages earn a click in months. It still ranks. It is still being read. It is being read by a model, and the model is not sending anyone.

The thesis

Zero-click is not the death of search. It is the death of one stage of the funnel, and it is the middle one. Buyers still start with a question and they still end with a purchase. What they no longer do is visit you while they are deciding between you and three other people.

The decline is not spread evenly, and that is the whole story

Start with the number everyone quotes. SparkToro, using Similarweb clickstream data for January to April 2026, found 68.01% of United States Google searches ended without a click. That was 60.45% in 2024 and 49% in 2019. Rand Fishkin also put AI Overviews in more than 20% of searches, cutting clickthrough by roughly 60% where they appear. Ahrefs, tracking Google referral share across more than 75,000 domains, measured an eight percentage point fall between June 2025 and May 2026, about a 22% drop.

Most people read those numbers as a flat tax. Everything goes down a fifth, so write more content and hold on.

Then look at where the answer box actually went. Semrush tracked more than 600,000 keywords across ten industries on the United States desktop database from November 2025 to April 2026. AI Overview presence on commercial-intent queries grew 71% over those six months. On transactional-intent queries it fell 5%. In finance, commercial-intent presence grew 231.25%.

Commercial intent is not a technical term you can wave past. It is the query class where somebody types “best”, “vs”, “alternatives”, “compared”, “review”. Transactional intent is “buy”, “pricing”, “demo”, “login”. One of those is your middle funnel. The other is your bottom.

So the answer box is not taxing your site evenly. It is expanding fastest at exactly the point where a buyer is choosing between vendors, and retreating at the point where they have already chosen and want to transact. If you average the decline across your whole site you will conclude you have an SEO problem. You have a stage problem, and the stage that took nearly all of the damage is the one that used to produce your pipeline.

The shortlist is now built before you know the account exists

6sense surveyed more than 4,000 B2B buyers across North America, EMEA and APAC for its 2025 Buyer Experience Report. The split between independent research and seller engagement moved from 70/30 to 60/40. Then the finding that should stop you: 94% of buying groups had already ranked a preferred vendor before they made contact with any vendor at all. 77% of purchases went to that preliminary favourite. The pre-engagement favourite wins 80% of deals. And 94% of buyers said they used a large language model to synthesise and organise their research.

Put the two datasets next to each other. The stage where the shortlist gets built is the stage where the click disappeared, and the tool doing the building is a model that never visits.

And the model is not reading the page that ranks

This is the part most marketing teams have not absorbed. Ahrefs ran 15,000 long-tail queries through Brand Radar and checked where the cited URLs sat in Google. Across ChatGPT, Gemini and Copilot, an average of 12% of AI-cited URLs ranked in Google’s top 10 for the same query. ChatGPT in-text citations: 8.0%. Gemini: 8.6%. Copilot: 8.2%. Perplexity was the outlier at 28.6%. Roughly 80% of citations did not rank anywhere for the query they answered.

Your comparison page won the middle of the funnel by ranking. Ranking is now a poor predictor of being quoted. Those are two different retrieval systems with two different tastes, and you have been optimising for one of them for fifteen years.

I stopped asking for the rankings report about a year ago. When I sit down with a marketing team now, the first artefact I ask for is the list of domains cited underneath the answers their buyers get, and almost nobody has one. I have asked for it often enough to know the reaction: a pause, then someone offers me the Search Console export instead. That export measures a channel that has stopped being the one that decides.

What replaces the middle of the funnel

Not more content. Three changes, and they are uncomfortable in a way that content calendars are not.

Comparison content has to stop being self-serving. A model asked to compare four vendors will not lift a table where one vendor wins every row, because that table is not useful for the job it was given. It lifts the source that concedes something. The page that says “if you need X, buy the other one” is the page that gets quoted, and being quoted is now the thing you wanted from ranking.

Your presence has to exist off your own domain. Given that 80% of citations rank nowhere, the sources being pulled are review sites, forums, documentation, analyst write-ups and third-party comparisons. You cannot publish your way to a citation on a domain you control. You can be the vendor those sources have something specific to say about.

The metric has to change from sessions to mentions. If your dashboard counts visits, the middle of your funnel now looks like it has stopped existing. It has not. It has become invisible to the instrument you are holding.

Where this breaks

Four honest limits, and I would not publish this without them.

The 68% figure comes from a clickstream panel, United States only, four months. Panels over-represent some users and under-represent others. The direction has held across seven years of this study, so I trust the trend. I would not build a board slide on the second decimal place.

The Ahrefs overlap result is long-tail queries specifically. Head terms almost certainly overlap more, because head terms are where the well-known sources sit. If your category is three words wide, your mileage differs.

The 6sense numbers are buyers self-reporting after the fact, and people reconstruct their own decisions to look more orderly than they were. Nobody remembers being influenced. Everybody remembers deciding.

And the honest one: Semrush also reports that a visitor arriving from a language model is worth 4.4 times an organic search visitor, which is the statistic every AI search vendor has been quoting all year. It has no published sample size or method, and it has an obvious selection effect. Someone who arrives from a model arrives late, already informed, already close to buying. Of course they convert better. That does not mean the model created the demand, and treating it as proof of channel value is the same error as claiming credit for branded search.

If you sell into formal procurement, or you sell a commodity, or your revenue is renewals, your middle funnel was never on the open internet and none of this touches you.

What I would do on Monday

1. Tag every landing page as informational, commercial or transactional, then pull twelve months of traffic by tag. If the loss is concentrated in commercial, you have this problem. If it is flat across all three, you have a different one and this post is not it.

2. Write down the five questions a buyer asks in the four weeks before they shortlist. Ask them in ChatGPT, Gemini, Copilot and Perplexity. Record whether you were named and what was said about you. That is under an hour and it is a baseline.

3. Read the citations under those answers, not your rankings. Most will not be you and most will not rank. That list is your actual distribution channel now.

4. Take one comparison page and concede a real case where a competitor is the better choice. Name it. Then watch whether it starts getting quoted.

5. Put named-in-answer rate on the dashboard next to sessions. A number nobody owns is a number nobody moves.

The page is still there

That “X vs Y” page is still sitting on the site, still ranking, still being read more than it ever was. It is doing the job it was built for. It is comparing two products for a buyer who is deciding.

The buyer just never arrives to say thank you.

I spend most of my week inside companies working out why the pipeline thinned before the traffic did. If that is the month you are having, book a call.

The money arrived in the technology seat. Almost nobody announced it.

The thesis

The CTO job split in two this year. One half became capital allocation: deciding what gets funded, at what unit cost, and what gets stopped. The other half, the half whose authority was denominated in headcount, is being deleted by the same technology that delivered the money. Most people holding the title are still doing the second job and calling it the first.

What actually moved

Three things happened in about fourteen months, and they are usually reported separately.

The amount got large. Gartner put worldwide AI spending at $2.52 trillion for 2026 in its 15 January release, up 44% on $1.757 trillion in 2025. That is not a line item. In most large enterprises it is now the biggest discretionary number the technology function touches, and it is growing faster than the function that governs it.

The composition is hostile to a builder. Inside that $2.52 trillion, AI infrastructure is $1.366 trillion. AI services are $588.6 billion. AI software is $452.5 billion. AI models are $26.4 billion. AI data is $3.1 billion.

Divide the first by the last. For every dollar the world spends on AI data this year, it spends $441 on AI infrastructure.

I have not seen a capital allocation ratio like that survive contact with a board. Infrastructure is bought at a price three vendors set, and it reprices every time they do. Data is the only line in that list that compounds inside your company and cannot be bought by your competitor on the same terms. A capital allocator looks at 441 to 1 and asks whether that is the return maximising split. A delivery manager looks at 441 to 1 and approves the invoice, because approving invoices was never the part of the job anyone was measured on.

The org chart made a new seat. IBM’s Institute for Business Value surveyed 2,000 CEOs across 33 countries and 21 industries with Oxford Economics between February and April 2026. 76% of organisations now have a chief AI officer. Twelve months earlier it was 26%. In the same study, 77% said talent and technology leadership roles are converging, and 85% said every functional leader has to become a technology expert inside their own domain.

Read those two findings together. The mandate is leaving the technology seat at the same time the money is arriving in it. Whoever can price a decision keeps both. Whoever cannot keeps neither.

What an allocator has to say out loud

A delivery manager is measured on shipping. An allocator has to state three things before spending anything, and I have watched a lot of capable technology leaders fail to state any of the three for AI.

The unit. One sentence: what is one unit of output from this, and what does it cost today. Not productivity. A countable thing with a dollar figure attached to it.

The hurdle. What return clears. The FinOps survey found a growing number of organisations now requiring AI investment to be self funded out of efficiency gains. That is a hurdle rate. It was set by finance, not by technology. If you did not write it, you are not the allocator. You are the applicant.

The kill criterion. The date and the number that ends it. Deloitte surveyed 3,235 business and technology leaders across 24 countries for its 2026 State of AI in the Enterprise report. Only 25% had moved 40% or more of their AI pilots into production. Only 21% reported mature governance for agentic AI. A portfolio where three quarters of the positions never reach production and four fifths carry no governance is not underperforming. It is unmanaged.

The half being deleted

Gergely Orosz published a piece on 18 August 2026 on engineering leaders taking career breaks. He interviewed roughly twenty of them. Six of ten CTO level contacts told him they were on the way out. The reasons he recorded are worth reading in order: unrealistic AI expectations, cost cuts of 20 to 50%, equity sitting behind a preference stack, and smaller teams reducing the need for the leadership layer at all. Several said they would rather work fractionally than take another full time seat.

I have a direct interest in that last sentence. I do fractional CTO work through PADISO and enterprise AI delivery through Brightlume, so the pattern Orosz describes is also my order book. I am not neutral about it. I also think he is describing a structural change rather than a mood. If your authority came from managing 60 engineers and the organisation now needs 25, the authority left with the 35. Seniority does not replace it. The people leaving are the ones who worked that out first.

Where this breaks

Four honest limits.

Gartner’s category boundaries are doing a lot of work in my 441 to 1 figure. AI data at $3.1 billion almost certainly excludes the data engineering labour, the warehouse spend and the integration work already booked under other lines. The real ratio is better than 441 to 1. The direction holds. The magnitude is arguable, and I would not put that number in a board pack without the footnote attached.

The chief AI officer figure is self reported titles collected from CEOs. A title can be created with no budget behind it. Twenty six to seventy six in twelve months is also too fast to be a stable structure, and some of those seats will not exist in 2028. What it proves is that boards decided the technology seat as it stood was not sufficient. It does not prove they built the right replacement.

Orosz’s sample is about twenty people and he says so himself. It is a signal from a well connected observer, not a measurement.

And the split does not apply below a certain size. Under roughly fifty engineers the allocator and the builder are the same person on the same day, and pretending otherwise produces a CTO who writes memos while the product stalls. This is a problem of scale. It arrives at the point where the AI bill starts to rival the engineering payroll.

What I would do on Monday

1. Write the unit for every AI initiative you fund. One sentence each: the countable output, and its cost today. If you cannot write it in a sentence, you are not funding a position, you are funding an activity.

2. Find out who set the hurdle. If AI at your company has to pay for itself out of efficiency gains, name the efficiency line and the quarter it lands in. If finance wrote that and you did not, fix the order.

3. Write the kill criterion before the next invoice. A date and a number. Put it in the same document as the business case, not in a review deck six months later.

4. Pull last quarter’s AI spend apart into infrastructure, licences, models and data, and compute your own ratio. Compare it to 441 to 1. If yours is worse, you have an allocation problem, not a technology problem.

5. If a chief AI officer was appointed at your company in the last twelve months, find out this week whether they hold the pen or the press release. The answer tells you which half of the job you still have.

I built D30 to pull three statement fact files out of ASX annual reports instead of hiring analysts to read them, because one of those options has a unit cost that falls every year and the other has one that rises. That was a capital decision taken from a technology seat. It is the only kind of decision that seat gets judged on from here.

The FinOps Foundation changed one word. Cloud became Technology. It reads like housekeeping. It is a charter, and charters get written after the power has already moved.

Brightlume does this work with enterprise teams. If the gap between the AI pilot and the P&L is the problem you have, talk to me.