32% skipped a software buy
to build with AI

The next software your company buys may be one it builds instead. McKinsey’s State of AI in 2026 finds nearly a third of organizations (32%) have decided against buying at least one software product or feature because agentic coding tools could build it in-house. This is not an efficiency story; it is a structural change in procurement decisions. Our judgment: build-with-AI inverts the unit economics — a predictable license is replaced by floating token operating cost plus model dependency — so procurement needs a new evaluation grid rather than treating “not buying” as a cost saving. Source: McKinsey State of AI 2026.

Where the 32% comes from

The survey was fielded May 4–June 8, 2026, with 1,719 responses across 97 countries, weighted by share of global GDP. The core finding: 32% of respondents report their organization decided against buying at least one software product or feature because it could be built internally with agentic coding tools. This is most common in technology and healthcare, followed by professional services and energy and materials — not the industries with the least budget, but the ones whose workloads look most like APIs.

The distribution matters more. Among AI high performers (the ~6% who attribute at least 5% of EBIT to AI), nearly half have skipped a software purchase; only 31% of others have. High performers scale software coding agents at twice the rate and other agentic AI at 2.7x. The ones who can build are already building — and they are the cohort most able to show AI value.

The unit economics invert: a license becomes an operating line

The classic buy-vs-build trade was “buy = fast but expensive, build = slow but cheaper.” Build-with-AI changes both dimensions: speed approaches buying (agentic coding compresses delivery to weeks), but the cost structure becomes operating — token spend floats with usage, model dependency demands switching rights, and maintenance lands on your own teams. McKinsey also reports about 20% of organizations say AI operating costs already constrain their AI use, and the share attributing EBIT impact has been flat at 37% for two years. A third-party read captures it precisely: “build got easy, benefiting did not.”

For finance, this is a ledger migration: the saved license fee does not disappear — it becomes model API bills, orchestration infrastructure, and maintenance code nobody wants to own. Substitution without an operating model is cost-shifting, not cost-saving.

Why this is a “third option,” not the new default

32% is structural, but it does not retire buying: compliance certifications, regulatory boundaries, and core-system integration still favor purchase; core differentiation, high-iteration, and API-shaped workloads favor building. The real change is that every decision now carries a third option that most procurement processes have no evaluation standard for.

Our judgment: give build-with-AI its own evaluation grid instead of forcing it into the old “build vs. buy” template. Three tests: ① Is this core differentiation or commodity? Build only what is core; building commodity is a liability. ② Is the workload API-shaped — clear boundaries, structured inputs and outputs, describable as a spec? If you cannot write the spec, an agent cannot own it either. ③ Can you price the maintenance burden, including model churn and switching? If you cannot price it, it does not belong in a budget.

Our judgment: substitution without an operating model is cost-shifting

High performers build not because they code better, but because they have the operating foundation: in the same survey, 73% of high performers rebuilt workflows around AI versus about 25% of others. Whether a build succeeds is decided by who operates, meters, and iterates it — not by whether it can be written. Treating “skipped software purchase” as this year’s saving usually gives the difference back in token bills and model-switch costs.

For buyers, the correct reading is: 32% is the procurement fact, the flat 37% EBIT is the warning — the substitution action creates nothing by itself; the operating model after it creates value. This is our own practice: anything we build internally must carry equally strict cost lines and evaluation gates, or we renew the license.

A checklist for CIOs

Actions and the question for buyers

Actions: before your next SaaS renewal negotiation, send the three-test grid to procurement and engineering so “build or renew” becomes a standard agenda item; for existing build projects, verify each has its own operating budget line. Question for buyers: under what evaluation framework was your last “we can build it with a coding agent, so we won’t buy it” decision made? If the answer is “there was no framework,” you are running naked in the 32% wave.

References: McKinsey, The state of AI in 2026: On the road to ROI (published 2026-08-25; 1,719 responses, 97 countries, fielded 2026-05-04 to 06-08) https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai ｜ Digital Applied, A Third of Companies Skipped Buying Software and Built It (secondary analysis, 2026-09-01) https://www.digitalapplied.com/blog/a-third-of-companies-skipped-buying-software-and-built-it