Everyone Feels Faster With AI. Almost Nobody Can Prove It.

Developers are convinced AI is making them faster. Their employers mostly can't back that up with a number. That gap, between how productive AI feels and what anyone can actually measure, is the headline of GitKraken's 2026 State of AI in Engineering report, a survey of 554 developers and engineering leaders.

The perception side is overwhelming. 84% of developers say AI made them more productive, and 43% say much more. Fewer than 5% feel slower. If you've watched an agent knock out a migration in an afternoon, that tracks.

The measurement side is where it falls apart. Only 20% of organizations measure productivity in any specific way. 39% have no way to gauge AI's impact at all, and another 33% lean entirely on developers self-reporting that it helped. Put those together and roughly 72% of organizations are running on belief instead of evidence, spending on tooling and restructuring workflows around a productivity gain they haven't quantified.

Why "it feels faster" isn't enough

Perceived productivity is a real signal, but a noisy one. It doesn't distinguish faster typing from faster shipping, and it says nothing about the parts of the job AI can quietly inflate: review load, rework, defects that surface later, the time spent verifying an agent's output. A team can feel 30% faster and be net slower once you count the cleanup. Without measurement, nobody knows which it is, and "we feel faster" isn't a number you can defend when the tooling bill arrives.

The adoption curve isn't waiting for the proof

Meanwhile, delegation to agents is accelerating. In September 2025, 7.6% of developers said handing tasks to an agent was their primary way of working. By June 2026 that was 28%, nearly a 4x jump in nine months, and two in three developers now run agents at least sometimes. Usage splits by tool: Codex and Cursor users behave like power users, with roughly 45% running parallel agents and about half keeping agents going all workday, while Copilot and ChatGPT users stay more assistive.

So the workflow is changing fast, confidence is high, and the evidence base is thin. That's the risk the report names: decisions are being made at a scale the measurement hasn't caught up to.

Closing the gap

You don't need a perfect metric to beat "we feel faster." Start with outcomes you already track and can attribute: cycle time from first commit to merge, change-failure rate, review turnaround, rework rate. Baseline them before a team adopts agents, then watch the direction. Pair that with a light qualitative read of where AI genuinely helps versus where it adds review burden. Most engineering orgs have already bet on AI productivity and are only now realizing they never set up the scoreboard; the teams that will actually know are the ones measuring before the next tool lands, not after. (If you're deciding which agent to run, the same measure-first discipline applies to model choice.)