Building software got cheap. Knowing what to build didn't — the 2026 data
The cost of building software collapsed — 41% of all code was AI-generated in 2025 — while the cost of knowing what to build didn't move. The data behind the shift.
The cost of building software collapsed. The cost of knowing what to build didn't move. Here's the data behind the shift — and why almost nothing in the tooling stack was rebuilt for the expensive half.
Building got cheap — fast
- 41% of all code written in 2025 was AI-generated.
- 84–93% of developers use or plan to use AI coding tools, up from 76% a year earlier (DORA 2025, JetBrains AI Pulse).
- In a controlled study of 4,800 developers, GitHub Copilot users finished coding tasks 55% faster (1h11m vs 2h41m).
Cursor, Claude Code, Bolt, Lovable, v0 — the time to turn an idea into working software dropped by an order of magnitude in about two years. If your bottleneck was ever "can we build it," it mostly isn't anymore.
But most of what gets built is never used
- 80% of software features are rarely or never used. Just 12% of features drive 80% of usage (Pendo, aggregated across finance, HR, education, logistics, healthcare, e-commerce).
- Public cloud-software companies invested up to $29.5 billion building features that are rarely or never touched (Pendo).
- ~43% of startups fail from poor product-market fit / no market need — the #1 or #2 root cause across a decade of post-mortems (CB Insights).
Cheaper building doesn't fix this. It amplifies it. When shipping is the constraint, a wrong bet costs weeks. When shipping is free, you can build the wrong thing faster than ever.
And the people who should catch it can't
- Product managers spend only about 27% of their time on strategy — ~73% goes to tactics and execution (Pragmatic Institute).
- Roughly 52% of a PM's time goes to unplanned "firefighting."
- 74% of product professionals spend fewer than 5 hours a month with customers or evaluators.
Keep reading
Spec Hallucination: The 15-Minute Claim Trace for AI-Written Specs
AI writes specs that sound sourced but often aren't. The Claim Trace is a 15-minute audit that labels every claim Sourced, Inferred, Assumed or Invented before engineering builds from it.
What Counts as Evidence? The Risk Receipt for Value, Usability, Feasibility, and Viability
AI made feasibility cheap to prove. Value, usability, and viability still need their own evidence — and a great prototype demo counts for less than it feels like. Here's the Risk Receipt framework for checking what you've actually retired.
What Product Managers Should Never Delegate to AI
AI can summarize evidence and propose options, but some product decisions should remain human-owned. Here are the judgments PMs should never delegate without review.