Thirty-two percent of organizations have decided against buying at least one software product or feature because they could build it internally with agentic coding tools (McKinsey, 2026). That is not a prediction about procurement. It already happened, inside the last budget cycle.
The build-vs-buy line moved, and it moved quietly — no vendor announced it, no one ran a business case past the board. Someone opened a renewal email, looked at the price, looked at a coding agent, and declined.
Here is the part that should slow a Head of Operations down before Q4 renewals close. In the same survey, the organizations furthest into that substitution — McKinsey's AI high performers, of whom nearly half have skipped a purchase this way against 31 percent of everyone else — report being constrained by cost in their use of software coding agents about three times as often as other respondents (McKinsey, 2026). The cohort with the most experience of building instead of buying is the cohort that found the ceiling first.
That ordering matters more than the headline number.
What the Survey Actually Says
McKinsey's tenth annual State of AI was published on 25 August 2026, fielded 4 May to 8 June, with 1,719 respondents across 97 nations, weighted by each nation's contribution to global GDP. Thirty-six percent of respondents work at organizations above $1 billion in annual revenue.
One caveat to install now, because it determines whether any of this applies to you: McKinsey cuts the data by revenue, not by headcount. "Smaller organizations" in this survey means under $1 billion in revenue, which includes companies well outside the 50–500 FTE band. Read the directional findings; do not read your own org chart into them.
With that said, three findings interlock.
The substitution is real and concentrated. Thirty-two percent overall, most common in technology and healthcare, then professional services and energy and materials. Nearly half among high performers.
The ceiling is real and lands on the same tools. About 20 percent of all respondents say AI-related operating costs, including token costs, constrained their AI use. For coding agents specifically, high performers hit that constraint roughly three times as often as others — and they are not disproportionately cost-constrained on other tool types. The friction is specific to the thing they substituted into.
The capability is not evenly distributed. Organizations above $1 billion in revenue went from 27 percent to 40 percent scaling AI agents in at least one function year over year. Smaller organizations stayed essentially flat at 22 percent. About two in ten respondents overall report scaling software coding agents, rising to 31 percent at larger enterprises.
So the firms best positioned to replace software with agents are also the ones already paying for the privilege. And the firms most tempted by the arithmetic — the ones whose SaaS bill hurts most — are the least likely to have scaled an agent at all.
Why the Swap Looks Cheap in September
The pull is not imaginary, and it is not vendor hype. Software prices are rising faster than almost anything else on your P&L.
Vertice's index put SaaS inflation at 12.1 percent in April 2026, 14.2 percent in May and 16.4 percent in June — a new record, surpassing the previous 14.7 percent peak of November 2025, and the fastest two-month acceleration they have recorded (Vertice, 2026). Vertice puts that at nearly five times the general rate of inflation. The same index also flags shrinkflation running alongside it: list prices up while feature access at the same tier quietly narrows.
Sit in a renewal meeting with that number and a working coding agent, and the trade feels obvious. A 16 percent increase on a tool three people use, against a build you could ship in a fortnight.
The problem is that the comparison being made is not the comparison being bought.
What Actually Transfers in a Build-Instead-of-Buy Swap
A SaaS contract is a fixed, supported, externally maintained liability with a known renewal date. What replaces it is three separate things, only one of which appears in the business case.
A variable token bill. The licence was predictable; inference is metered and scales with use, retries, and whatever an agent decides to read. This is precisely the line McKinsey's high performers report hitting.
A permanent internal maintenance owner. Not the person who built it. The person who holds it in 2028, when the API it calls changes and the one engineer who understood it has left.
The maintenance characteristics of agent-written code, which are measurably worse than the code it replaces.
The Maintenance Owner Nobody Names
GitClear's Maintainability Gap study analysed 623 million code changes from 2023 to 2026 and found the quality signals moving the wrong way at once: code block duplication up 81 percent, within-commit copy/paste up 41 percent, error-masking constructs up 47 percent, two-week churn up 15 percent — while refactoring line moves fell 70 percent, cross-file function calls (the signal for reuse) fell 35 percent, and long-term legacy maintenance fell 74 percent against 2022 levels (GitClear, 2026).
Their phrase for what this produces is the useful one: perpetual V1 components. Code that ships and is never consolidated, because nobody refactors what an agent can regenerate.
Google's DORA programme reaches the same place from a different direction. Across roughly 5,000 respondents, AI adoption among developers rose from 76 percent to 90 percent, and AI-assisted delivery throughput recovered to neutral or better — but delivery instability persisted as a cost of adoption, and DORA's framing of AI is as an amplifier of whatever organizational system it lands in (DORA, 2025). The prior year's edition had measured a 7.2 percent decrease in delivery stability for every 25 percent increase in AI adoption (TechTarget, 2025).
Amplifier is the word to hold onto. A 300-FTE firm with one overloaded internal developer and no deployment pipeline does not get a software vendor's discipline by generating a replacement for the vendor's product. It gets its own discipline, faster.
The Cohort Furthest In Found the Ceiling First
Most technology arguments are about whether the sceptics will be proved right later. This one has already run the experiment, and the result is in the same dataset as the enthusiasm.
The organizations that substituted most aggressively are the ones now reporting cost constraints on coding agents three times as often as everyone else (McKinsey, 2026). Not cost constraints in general — they report those at roughly normal rates. Specifically on the tool class they swapped into.
That is what a ceiling looks like when it is discovered from the inside. You replace a fixed line with a variable one, usage grows because the thing is useful, and the variable line stops being smaller than what you cancelled.
For a mid-market operator the inference is not "don't build." It is that the cost curve you are being shown at the moment of decision is the first month of it.
The Honest Counter
Three limits, stated plainly.
Self-reported survey data is not an audit. McKinsey's figures are what respondents say about their own organizations. Nobody reconciled the claimed in-house build against a general ledger, and "decided against buying" includes decisions that were never going to close anyway.
Build genuinely wins in a definable set of cases. Where the workflow is specific to your firm, where no vendor fits without heavy configuration, where the thing is small and stable and reads from systems you already own — building has always been defensible, and agents lower the threshold where it becomes so. The 32 percent is not uniformly a mistake.
One year of quality data is not a verdict on a decade. GitClear measures what AI-assisted authorship looks like now, while tooling and practice are both immature. Those signals could improve. They have not yet.
What none of that changes: the swap converts a vendor's balance-sheet problem into yours. That is a transfer, not a saving, unless you price the transfer.
Which Renewals Make the Trade Defensible
Four moves before Q4 closes. None requires a new vendor, and none requires abandoning the build case.
- Price the token line at twelve months of grown usage, not at pilot volume. If the replacement is only cheaper at today's call volume, you have not found a saving — you have found a lag. Model it at three times current usage and see whether the decision survives.
- Name the maintenance owner before you cancel the contract. A name, a percentage of their time, and a successor. If no one will put a number on it, the build is being funded from unallocated headcount, which is to say from someone's evenings.
- Sort renewals by who absorbs the breakage. Anything touching payroll, compliance, customer-facing billing or regulated data keeps a vendor with an SLA and an indemnity. Internal reporting, glue between systems you already own, and single-team workflow tools are where the build case is honest.
- Treat a 16 percent renewal quote as a negotiation signal, not a verdict. SaaS inflation running at nearly five times general inflation means the quoted increase is market-driven and frequently negotiable. Benchmarking a renewal is cheaper than owning a codebase, and you can do it this month.
A licence you cancel is a cost you stop paying. A system you build is a cost you start owning — and the organizations furthest down that road are telling you, in the same survey that made the swap look attractive, exactly where it stops being cheap.
Before the next renewal goes out, pick one tool on the chopping block and write down who maintains its replacement in 2028. If the line stays blank, renew it.
