Everyone's Skipping the Software Purchase. Few Have Budgeted for What Comes Next.

McKinsey's 2026 State of AI survey — 1,719 respondents across 97 countries, published August 25 — found that 32% of organizations turned down a software purchase this year because an AI coding agent could build the same thing in-house instead. Among the "high performers" already seeing real returns from AI, that number is closer to half. This isn't a hypothetical trend piece — it's a real, current shift in how mid-sized businesses are deciding whether to buy a tool or just have an agent build one.

The appeal is obvious. If you've ever paid a recurring SaaS fee for a tool your team uses at a fraction of its capacity, "just have the agent build a simpler version" sounds like common sense. Today's coding agents genuinely can turn a plain-English description into a working internal dashboard, a lead-routing script, or a reporting view in an afternoon — work that used to mean a vendor demo, a contract negotiation, and a multi-week onboarding.

But the same survey has a warning buried a few pages later: "building has become the easy half." Only 37% of organizations can point to any actual profit impact from their AI use — flat compared with a year ago — even though 80% report individual productivity gains. And the businesses leaning hardest on coding agents hit cost and maintenance limits three times more often than everyone else. The part that used to be the vendor's job — patching, security review, migrating the tool when your other systems change — doesn't disappear when you skip the purchase. It just moves onto your own plate.

For a business without a dedicated engineering team, that's the real risk. Someone in marketing can describe a dashboard to an agent and have a working prototype by end of day — genuinely useful, genuinely fast. The trouble starts three months later, when nobody remembers who's supposed to notice it broke, or that it still has read access to customer data nobody's reviewed since launch. A tool with no clear owner doesn't fail loudly. It just quietly drifts out of sync until it breaks during your busiest week.

McKinsey's own read on who's actually winning backs this up: the organizations getting real returns redesign the workflow first, then build the tool for it — they don't bolt an agent-built feature onto a process that hasn't changed. Before you let an agent build your next internal tool, decide who owns it after it ships, the same way you'd decide who manages a vendor contract. That accountability question is exactly why marketing and IT can't run as separate, occasionally-coordinating teams anymore — someone has to actually own what gets built, not just enjoy the demo.