McKinsey's annual survey on the state of AI, published on 25 August 2026, contains a lot of expected numbers and one surprising one. The expected numbers say adoption is deepening, large enterprises are moving faster than small ones, and financial impact is not budging. The surprising one fits on a single line: 32% of respondents report that their organisation decided against purchasing at least one software product or feature because it could now be built in-house using agentic coding tools.
Nearly one in three. And among the 6% of organisations McKinsey classes as "high performers" — those attributing at least 5% of their EBIT to their use of AI — it is nearly one in two.
Let us say up front what this measurement is worth, because everything that follows depends on it. The survey ran from 4 May to 8 June 2026 among 1,719 respondents across 97 countries, weighted by each country's contribution to global GDP; 36% work at organisations with more than $1 billion in annual revenue. It is self-reported, on a panel that is not a random sample of companies. And McKinsey sells AI transformation consulting: the firm has an obvious interest in readers concluding that they should accelerate. None of that disqualifies the numbers — they are consistent year on year, and the survey publishes the gap between what respondents predicted and what they observed, which is to its credit. It simply means reading the recommendations differently from the data.
"Not buying" is not "having built"
First caveat, and it matters: the question is about a decision, not a delivery. A respondent reporting that they walked away from a purchase is not reporting that the replacement software is running in production, that it is maintained, or that it cost less.
That is a nuance, but it is the one separating a signal from a result. What the 32% measures is the moment the trade-off flipped in someone's head: the cost of building has fallen far enough that "we'll do it ourselves" is a sentence you can once again say out loud in an investment committee. That is already considerable. It says nothing about what happens next.
And the build-or-buy trade-off was never about the cost of building alone. It was about what buying included on top of the software.
What the purchase included, and what you now own
When a company buys a piece of software from a vendor, it is not buying features. It is buying four things bundled into a subscription that never appear on the invoice:
- a roadmap — someone, elsewhere, decides what the product becomes and funds it;
- maintenance — version upgrades, security fixes, compatibility with the rest of the ecosystem;
- compliance carried by a third party — certifications, contractual commitments, legal liability when something is breached;
- someone to call — the person on the other end of the line when it breaks on a Friday evening.
None of those four lines disappears when you build. They change owner. The decision McKinsey measured is therefore less "we saved a licence" than "we internalised a lifecycle" — with everything that implies for governance: knowing who decides what changes, who still understands how it works in eighteen months, and on what basis you would one day decide to retire it.
This is not an argument against building. It is the argument that decides whether building holds. A coding agent collapses the cost of the first draft; it does nothing for the cost of ownership. That is exactly the shift we described in 40 days and 2.3 billion tokens: the work does not disappear, it moves downstream — and downstream, here, spans years.
The paradox at the top of the table
The most interesting passage in the survey is the one nobody will quote in a LinkedIn post, because it contradicts the savings promise.
McKinsey reports that about 20% of respondents say AI operating costs (tokens included) constrain their use of AI. A moderate share, broadly consistent across organisation sizes. But isolate the high performers — the ones building the most, deploying the most coding agents, walking away from the most purchases — and the number changes character: they report being constrained by the cost of their coding agents about three times as often as everyone else. And the gap is specific to them: on other kinds of AI tools they are constrained no more than average.
Even as per-token costs have declined, the number of tokens consumed and generated has increased even faster.
That sentence is from Michael Chui, a McKinsey senior fellow, in the commentary accompanying the survey; he adds that finance leaders are discovering AI is not "too cheap to meter" once you get to agentic software development and complex reasoning tasks.
In other words: succeeding with coding agents means meeting their bill. Success does not protect you from the cost constraint, it leads you to it. It is the same mechanism as Uber burning its annual AI budget in four months — an episode that reads better as a symptom of maturity than as an overrun. A company that never hits its token ceiling is rarely a frugal company; more often it is a company that has not yet put its agents to work.
The practical consequence comes from Lieven Van der Veken, a senior partner, and I keep it because it is actionable tomorrow: the organisations moving fastest treat operating costs as a design constraint, not an afterthought. Treating cost as a design constraint means deciding the model, the context you send and the scope of a task before launching it — not discovering the bill at the end of the month.
80% against 37%: the gap value falls into
The third number is the one that comes back every year, and it still has not moved. 80% of respondents say AI has improved their individual productivity, and about half say it helps them make better decisions. But only 37% attribute any positive EBIT impact to AI — give or take, the same share as last year, despite another year of deployment.
Resist the easy reading ("so AI is useless") as much as its opposite ("so the value is there, hidden"). The two numbers do not measure the same thing: the first is an individual perception, the second an accounting attribution. Twenty minutes saved per person per day is perfectly real and perfectly invisible in a P&L, as long as nothing in the organisation converts it into something — one more release, a shorter lead time, a person freed for other work.
It is the same blind spot the Institut Montaigne pointed at when it concluded that the problem is not technology but organisation. The McKinsey survey puts it its own way, through Tara Balakrishnan: the limiting factor is increasingly the organisation's ability to absorb change.
One detail deserves more attention than it gets, because the experience is not uniform by level. 47% of mid-level managers and individual contributors report having experienced at least one AI-related negative effect, against 31% of executives and senior managers. The enthusiasm measured at the top is not the one lived below — and a transformation steered on executive sentiment alone will misdiagnose itself.
The number that should make you careful with all the others
McKinsey publishes a comparison one wishes were more common: what respondents predicted last year, set against what they observe this year.
In 2025, 32% of respondents expected AI-related declines in their organisation's head count over the coming year. This year, 14% report an actual decline — less than half. Two-thirds report little or no change. And for the year ahead, the forecast climbs back to 39%.
There is a methodological lesson here that applies to every survey of this kind, including this one: a questionnaire measures expectations well and the future badly. The same people who were off by a factor of two on head count also answered the questions on productivity and financial impact. That does not make their answers useless — it is a reminder that a stated intention is not a fact, and that a feeling of productivity is not a measurement of productivity.
It is also what makes the trajectory of that 32% worth watching. If next year's survey shows those abandoned purchases turned into internal software that is live and maintained, the shift is real. If they turned into projects stopped halfway, we are back to the scenario we described in companies cutting AI projects without knowing why, with one aggravating factor: you will not only have lost a project, you will also have lost the software you did not buy.
What the 6% do differently
McKinsey describes a small group of organisations — 6% of the panel, unchanged from last year — that attribute at least 5% of their EBIT to AI. Three traits set them apart, and none of them is technological:
- they redesign workflows instead of inserting AI into them. Nearly three-quarters report fundamentally redesigning workflows (up from 55% last year), against one-quarter of everyone else;
- their senior leaders visibly commit — they are twice as likely to report it;
- they have a defined process for measuring the impact of their AI initiatives — again twice as often as the rest.
The third point is the one that interests me most, because it is the cheapest to put in place and the first to be sacrificed. Measuring impact returns nothing while everything is going well; it only becomes valuable at the moment of arbitration, which is to say too late to install retroactively. That the best-performing organisations are also the ones that measure does not establish causation — the survey cannot support that — but the correlation is clear enough to make you wonder which side you want to be on.
What it changes for whoever decides to build
If you are in the third that walked away from a purchase, you have just taken on a roadmap, a maintenance burden and a compliance obligation. Three questions follow, and none of them is technical:
- Where does the specification live? A vendor documented its product for you. In-house, if the only place the expected behaviour is written down is the code the agent produced, the knowledge is already lost. That is the subject we covered as the spec as an asset, and the decision to build makes it suddenly a lot less theoretical.
- Who is accountable, by name? The vendor contract designated an owner by default. Software built in-house without a named owner does not have a default one: it has none, and that only becomes visible on the day of the incident.
- What is the real cost, per feature? A licence compares to another licence. An in-house build compares to what it cost — in tokens, in review time, in retries. A cost attached to a month supports no decision; a cost attached to a feature and a version does.
That is precisely the work Survol does, and I am not going to pretend it is a coincidence: we build a steering tool for products developed by agents, so those three questions are our raw material. But they arise with or without us, and they arise on the day the decision is made — not eighteen months later, when nobody remembers why the module does what it does.
So the real lesson of the survey is not that AI lets you build instead of buying. It is that the decision to build has become a governance decision again, after twenty years in which it was mostly a budget decision. Which is far less comfortable — and considerably more interesting.
Sources
- Dan Tinkoff, Lieven Van der Veken and Michael Chui, with Tara Balakrishnan, "The state of AI in 2026: On the road to ROI", McKinsey & Company (QuantumBlack), 25 August 2026. Online survey run from 4 May to 8 June 2026 among 1,719 respondents in 97 countries, weighted by each country's contribution to global GDP. Every figure and quotation in this article comes from it.
