Build · pm · 8 min read

A Third of Companies Just Declined a Software Purchase Because They Could Build It

McKinsey's 2026 survey: 32% of organisations skipped a software buy because coding agents could do it. What that means if you're selling software.

Published September 8, 2026 ·

strategy build-vs-buy saas coding-agents product-management

Almost everything on this site is written from the builder’s side of the table: which tool to use, what it costs, what it will do to you at 2am. This one is from the other side. McKinsey published its 2026 State of AI survey on August 25, and it contains a number that should change how you think about your own product roadmap — whichever side of the table you are on.

32% of organisations report deciding against buying at least one software product or feature because they could build it internally with agentic coding tools.

Not “are considering.” Decided. Already happened.

What the number actually says

The survey drew 1,719 responses across 97 countries between May 4 and June 8, weighted by each country’s share of global GDP. That is a serious sample, and the finding is stated plainly in the published report.

Read the unit carefully before you panic or celebrate. It is one product or feature, at least once. A company that built one reporting dashboard rather than paying for a seat upgrade counts identically to a company that replaced its CRM. McKinsey gives no prior-year figure, so there is no trend line — this is a snapshot, not a slope.

Two more numbers give it shape:

• Among AI high performers — the 6% of respondents who attribute at least 5% of EBIT to AI — the share is nearly half, against 31% of everyone else.
• It is most common in technology and healthcare, then professional services and energy.

So the companies furthest along are the ones building instead of buying, and the ones closest to you in the tech sector are doing it most.

The number that should be quoted next to it

Here is the part that did not travel as far. The share of organisations attributing any EBIT impact to AI is 37% — which McKinsey describes as essentially unchanged from a year ago. High performers stayed flat at about 6%. Meanwhile enterprise-wide AI scaling rose from 38% to 44%, large-enterprise agent deployment jumped from 27% to 40%, and 80% report individual productivity gains.

Everything measuring activity went up. The one thing measuring return stood still.

McKinsey’s own line: organisations’ conviction in AI is growing faster than the financial returns they can attribute to it. Deciding not to buy something is an act of capability. Whether it pays depends entirely on what happens in year two, and a survey taken in May cannot see year two.

If you sell software, this is a demand-side story

Your buyer’s build option got cheaper. That is the whole thing, and it is not evenly distributed.

The features most at risk are the ones where your product forced the customer’s process into your shape: the custom report they always complain about, the workflow step that never quite matched how they work, the integration they asked for twice and you deprioritised. Those were never things you were uniquely good at. They were things that were previously too expensive for the customer to build. That is no longer true.

The features least at risk are the ones where scale is the product: payments, deliverability, identity, compliance, uptime, and the thousand edge cases the customer will not hit until year three. Nobody is rebuilding Stripe with a coding agent because a weekend of vibe coding does not produce PCI compliance.

If you are a PM, the practical exercise is short. List your top ten feature requests from the last two quarters. Mark each one shape or scale. The shape column is the part of your roadmap that a customer with a coding agent can now do without you — and pricing a tier around that column is a strategy with a visible end date.

If you buy software, use the high performers’ rule, not the headline

The tempting read is “we can build it, so cancel the renewal.” McKinsey’s own data argues against that being the first move. What separates the 6% who make money from AI is not tooling. The report says they fundamentally redesign workflows rather than inserting AI into existing ones, pursue growth rather than only efficiency, put budget behind it, and measure it.

Declining the vendor quote is the last step in that sequence. Taken first, it is a more expensive way to change nothing — you now own a feature that does exactly what the vendor’s did, plus the maintenance.

A workable decision rule:

Build where the workflow is genuinely yours and the purchased version would distort it. Internal dashboards, quoting logic, ops tooling that mirrors your actual steps.

Buy where the function is generic and the vendor’s scale is the value. Payments, email, auth, accounting, anything with a compliance surface.

Build the feature, keep the platform. This is the survey’s most common case, given the “product or feature” wording. Build the missing report against the vendor’s API. Do not cancel the vendor.

Build, then roll out slowly. If you are replacing something people use daily, run the new thing read-only beside the old one until it earns trust. The build is the easy half; adoption is the project.

The cost nobody prices in advance

About one in five respondents say AI operating costs — explicitly including token costs — already constrain their use of AI. And high performers report cost constraints on coding agents roughly three times as often as everyone else, precisely because they use them most.

That is the tell. The organisations building the most are already the ones feeling the run cost.

A purchased product has a price you know in advance and that mostly holds. A built one has a token bill that scales with usage and a maintenance burden you absorbed silently at the moment you declined the quote. None of that argues against building. It argues for pricing the run cost of a built feature the way the vendor would have — before the decision, not after — and naming a human who owns it in year two.

If you have watched the pricing whiplash we have tracked all year across Copilot, Cursor and the June billing reset, you already know how quickly a token bill can stop being a rounding error.

The honest summary

A third of companies can now build what they used to buy. The same survey says the number making money from AI has not moved in a year. Both are true, and holding them together is the whole skill.

Building got easy. Deciding what is worth building did not.