32% Cancelled a Software Purchase to Build It With Agents. 6% Are Actually Seeing Returns

Three figures from McKinsey's State of AI 2026 set against each other — 32% of organisations having decided against buying a software product or feature in order to build it with coding agents, with technology at 41% and healthcare at 39%, beside the 37% attributing any EBIT impact to AI and the 6% qualifying as high performers, both unchanged year on year, above the note that 80% report personal productivity gains that are not aggregating — The buying behaviour changed sharply. The measured organisational return did not move at all. Both findings come from the same survey of the same 1,719 people.

If you sell software, the most important number published this week is 32%.

That is the share of organisations in McKinsey's State of AI 2026 — a survey of 1,719 professionals and business leaders worldwide — who report deciding against buying one or more software products or features, because they could build it internally with agentic coding tools.

Not "considered building." Decided against buying.

1. Where it is worst

The headline number hides a spread that matters more than the average, because it tells you whether this is your problem or somebody else's.

Share who skipped a purchase to build with coding agents, by sector

McKinsey State of AI 2026

Technology — 41%. The worst exposure by a clear margin. If you sell developer tools or infrastructure to technology companies, four in ten of your prospects have already done this to somebody.
Healthcare payers and providers — 39%. Higher than most people would guess for a regulated, risk-averse sector.
Professional services and energy and materials — 38%.
Financial institutions — 36%.
Media and telecom — 34%.
Pharmaceuticals and medical products — 33%.

Notice the floor. The lowest sector in that list is 33%. There is no segment where this is not happening, which means "our buyers are too regulated to build it themselves" is no longer a position you can hold without evidence.

2. The number that sits awkwardly beside it

Here is the part worth reading twice, and it comes from the same report, the same respondents, the same fieldwork.

37% attribute at least some EBIT impact to AI. That figure is flat against 2025.

6% qualify as high performers — attributing at least 5% of organisational EBIT to AI with significant impact. That is also unchanged year on year.

The buying decision moved sharply. The measured organisational return did not move at all.

So a third of organisations are cancelling purchases on the strength of a capability that, by their own accounting, has not yet changed their earnings. The confidence is running well ahead of the demonstrated result.

Why this is not simply 'they are wrong'

Returns lag decisions, and they should. Nobody expects a build started this quarter to show up in EBIT this quarter, and a flat aggregate can hide a rising distribution. But two years of flat is not a lag, it is a pattern — and the honest description is that the build decision is being made on belief about where the capability is going rather than on where it has demonstrably arrived.

3. The productivity paradox, in one survey

The mechanism sits in two more figures from the same report.

80% of respondents using AI say it improved their individual productivity. And yet organisational EBIT impact is flat.

That is the classic shape of a productivity paradox: real gains at the level of the person, which fail to aggregate at the level of the firm. Individual time saved gets absorbed — into more meetings, more review, more rework, more coordination — rather than showing up as output or margin.

It also explains how both halves of this report can be true at once. The engineer who used an agent to prototype a replacement for a $40k/year tool genuinely did that faster than they could have last year. Whether the company ends up cheaper, once you count the maintenance, the on-call, and the eighteen months of edge cases, is a different question that the survey period has not yet answered.

4. What the reliability data says about the bet

I wrote last week about Microsoft's Thinkingbox benchmark, which runs 507 stateful business workflows twenty times each and grades what actually changed in the database. The strongest model tested managed 65.36% pass@1 and 25.25% pass^20 — right first time on two tasks in three, right every time on one in four.

Put that next to the 32%.

The workflows these teams are replacing — claims processing, billing rules, procurement approvals, account provisioning — are precisely the multi-step, policy-conditioned, state-heavy tasks the benchmark measured. And that is where measured agent reliability is weakest.

This is the same calibration gap I keep running into, now on the buy side. In the Temporal survey, 85.5% of engineers trusted agent output while half hit problems daily. Here, a third of organisations are cancelling purchases on that same trust. The gap has stopped being an engineering curiosity and started showing up in someone's revenue.

5. What actually survives, and what does not

None of this means the 32% are making a mistake. For a genuinely simple, well-bounded internal tool, building it with agents is now often the right call — and the honest response to that is to stop selling those products, not to argue with the buyer.

What survives is what an internal rebuild cannot cheaply reproduce.

A feature they can specify

Does not survive

If a competent engineer can write down what your product does in a paragraph and an agent can approximate it in a weekend, that was never a moat. It was a head start, and the head start just got shorter.

What you know that they do not

Survives

Accumulated edge cases, the data you hold, the integrations you have already certified, the compliance surface you have already passed. None of it is in the spec, all of it is in the price, and an agent cannot generate it because it is not a coding problem.

Who carries the pager

Survives

An internal build transfers the maintenance, the on-call and the liability onto a team that did not previously own them. That transfer is the actual cost, it is almost never in the business case, and it is the strongest honest argument you have.

Whether the workflow is stateful

Decides it

A read-only dashboard is a good build candidate. A multi-step process that writes to systems of record, under policy, is where the benchmark says agents are least reliable — and where your product has already absorbed years of failure modes.

6. What to change on Monday

The rebuild cost, honestly

Write

Not a scare document. A real accounting of what the buyer takes on: maintenance, on-call, compliance evidence, the integrations, and the edge cases you have already hit. Buyers respect an honest total cost of ownership and discount a defensive one.

"Have you already built something?"

Ask

Put it in discovery. A third of your pipeline has done this somewhere, and the ones who tried and abandoned it are your best references — they have priced the thing your deck is arguing about.

By sector exposure

Segment

Technology at 41% and pharma at 33% are different markets now. If your ICP is technology companies, this is a first-order threat to your pipeline. If it is not, you have some time — but the floor is 33%, so not much.

The middle card is the cheapest and nobody does it. "Did you consider building this internally, and how far did you get?" turns an objection you handle at the end of a cycle into information you have at the start of one.

7. What I would not conclude

Three cautions, because this is a survey and surveys invite over-reading.

It is self-reported intent about past decisions, not audited procurement data. "We decided against buying" covers everything from a cancelled renewal to a feature that never made a shortlist — and the second is far less costly to you than the first.

Flat EBIT is not proof the builds fail. It is proof they have not yet shown up. The interesting figure will be the 2027 report: if the 32% rises again while the 6% stays put, that is a pattern worth naming. One year is not.

McKinsey sells AI transformation consulting. The same discount I applied to Temporal's survey applies here — the raw percentages are far more trustworthy than any framing laid over them, and "on the road to ROI" is a conclusion with a commercial interest attached.

The honest summary

Nearly a third of organisations have already decided against a software purchase because agentic coding tools made building look viable. In technology it is 41%, and the lowest sector on the list is 33%. That is a demand-side shift, it is happening now, and it is measurable in somebody's pipeline this quarter.

The same survey found the share of organisations seeing any EBIT impact from AI did not move all year, and the share seeing serious impact stayed at 6%. Meanwhile the best-measured agent reliability on exactly the stateful workflows being replaced is 25% across repeated runs.

The buyers are not irrational. They are early, and they are betting on a trajectory rather than a result. Your job is not to tell them the trajectory is wrong. It is to be honest about what a rebuild actually costs them, and to be selling something a weekend of agent output cannot approximate.

If you cannot name what that is, the 32% is not the threat. It is the diagnosis.

Sources: McKinsey, "The State of AI in 2026" · The Register, "McKinsey says enterprise AI is finally 'on the road to ROI'" · Sector breakdown as reported by India Gazette · arXiv 2608.19741, Thinkingbox · The 1,719 sample size, the sector percentages and the EBIT figures are as published by McKinsey and the reporting above; the argument and the cautions in section 7 are mine.

Microsoft Ran the Same Agent Task Twenty Times. The Score Fell From 65% to 25%August 27, 2026

Trust Is Running Ahead of the Evidence, and a New Survey Has the ReceiptsAugust 26, 2026

The Vibe Coding Trap: Why Thousands of AI Startups Are Paying $50K–$500K to Rebuild Their MVPsApril 2026

IdeaToMVP Academy

Want to build with AI — not just read about it?

4-week live cohort for founders. Learn to ship AI agents, scope MVPs, and automate your business — taught by the same team that writes these guides.

Explore the Academy →