McKinsey State of AI 2026: 32% Skipped a Software Buy
The Short Answer
McKinsey’s State of AI 2026 global survey found that 32% of organisations decided against buying at least one software product or feature because agentic coding tools let them build it internally. In the technology sector the figure rises to roughly 41%.
In the same survey, the share of respondents reporting EBIT impact from AI stayed flat at 37% year over year.
Those two numbers together are the story: enterprises are changing how they acquire software faster than they are changing what they earn.
What the 32% Actually Measures
Read the wording carefully, because the headline is being over-read everywhere.
The survey asked whether organisations decided against buying one or more software products or features because agentic coding tools made building it viable. It did not ask:
- whether the in-house build shipped
- whether it is still running a year later
- whether total cost came in under the licence it replaced
- whether anyone is on call for it
“We decided not to buy” is a procurement event, not an engineering outcome. The gap between those two is where most of the interesting risk sits. A third of organisations have made a decision whose consequences arrive on a two-to-three-year lag.
Why This Is Happening Now Rather Than in 2024
Three things converged:
- Agentic coding crossed a threshold on small, well-specified internal tools. Not on complex systems — on the CRUD app, the internal dashboard, the integration glue, the reporting layer. That is precisely the category that mid-market SaaS has monetised for fifteen years.
- The buy path got more expensive and more annoying. Seat-based pricing, AI surcharges on existing products, and procurement/security review cycles that take longer than the build now does.
- The build path got a credible demo. An engineering lead can produce a working internal tool in an afternoon and walk it into a budget meeting. Nothing persuades a CFO like a working thing.
The vulnerable purchase has a recognisable shape: thin logic over your own data, few external integrations, low regulatory surface, and a licence cost that looks large next to a week of engineering time. Anything that description fits is now contested.
The Flat 37% Is the Warning Label
Adoption is up. Agents in production are up. Software purchases avoided: a third of respondents. And the profit attribution did not move.
Three explanations, all partly true:
1. Savings landed in the wrong line. Cancelling a $60,000 licence and absorbing the work into an engineering team that now maintains another service does not create EBIT — it moves cost from a visible line to an invisible one. Headcount did not fall. Attention did.
2. Value is accruing to the smallest units, not the enterprise. Individual developers and teams genuinely go faster. That speed does not aggregate to the P&L when the constraint on the business is demand, regulation, or organisational throughput rather than engineering capacity.
3. It is early. Deployments from 2026 will show up in 2027 and 2028 financials, if they show up at all. Flat is not the same as zero.
The uncomfortable reading is the first one. A build decision that prices agent tokens against an annual licence is comparing construction cost to total cost of ownership — and the survey’s flat profit line is what that mistake looks like in aggregate.
What Building Actually Costs After the Agent Stops
The generated code is the cheapest artefact in the lifecycle. What follows is not:
| Cost | Who pays it |
|---|---|
| Security review of code nobody wrote by hand | Your AppSec team, forever |
| Dependency and CVE patching | Your platform team |
| On-call when it breaks at 02:00 | Your engineers |
| Compliance evidence (SOC 2, GDPR records, audit trails) | Your compliance function |
| Onboarding the next maintainer | Everyone, repeatedly |
| The rewrite in year three | The team that inherits it |
| Accessibility, i18n, edge cases the vendor handled | Nobody, until a customer complains |
A vendor amortises all of that across thousands of customers. You amortise it across one. That is the entire economic argument for SaaS, and coding agents do not change it — they change the cost of the first 20% of the work, which was never where the money went.
What It Means If You Sell Software
The 32% is a real signal for vendors, but the threat is specific rather than general:
At risk now: single-purpose tools, thin analytics layers, form builders, internal-workflow apps, anything whose demo can be reproduced in an afternoon, anything priced per seat for occasional users.
Not at risk yet: systems of record, products carrying regulatory certification, anything where the data network effect is the product, anything where the vendor’s support obligation is what the buyer is really purchasing, deep vertical software with a decade of domain edge cases encoded.
The defensible move is to stop selling capability and start selling the obligations — uptime, liability, certification, integration maintenance, the roadmap. Those are the parts a coding agent cannot generate, and they are what actually survives a build-vs-buy review in 2026.
What It Means If You Buy Software
Before you convert the next licence renewal into a build ticket, price the whole thing:
- Estimate the run cost, not the build cost. Three years of maintenance, on-call, security review and one migration.
- Ask who owns it when the person who built it leaves. If the answer is nobody, you have not saved money, you have deferred a cost.
- Check the regulatory surface. If the vendor’s certification was doing compliance work for you, rebuilding it in-house transfers that obligation to you at full price.
- Build where the logic is yours. The strongest build cases are workflows that encode something specific about how your business works — precisely the things vendors serve badly.
- Buy where the obligation is theirs. Anything with an SLA, an audit report, or a liability transfer attached.
The honest summary of McKinsey’s 2026 data: a third of enterprises changed their acquisition behaviour, and none of them can yet show it in profit. Both halves of that sentence should inform your next decision.
Last verified: September 8, 2026.