Two studies. Two conclusions. Same question.
DX Research tracked 500+ companies over four quarters. Among developers who reached the highest AI time-savings band, 50.5% didn’t report that level of savings in subsequent quarters. For developers who took two quarters to get there, 79% didn’t sustain. Two-thirds of everyone who hit peak savings reported lower savings afterward.
Microsoft tracked tens of thousands of its own engineers over four months after rolling out Claude Code and GitHub Copilot CLI. Adopters merged 24% more pull requests than a synthetic control group. The lift held: +29.4% in January–February, +20.0% in March–April — the decline within sampling noise. Weeks with heavier tool use showed monotonically more output. Deepening integration, not fading novelty.
One study says gains decay. The other says they persist.
The question I can’t answer: which frame is right?
Self-reported time savings
Two-quarter ramp-ups: 79% didn’t sustain
Objective PR measurement (Bayesian causal)
Dose-response: monotonic increase
The Gap Between the Numbers
They measured different things. DX tracked self-reported time savings — how much faster developers felt. Microsoft tracked merged pull requests — how much more code developers shipped. These sound like they should correlate. They don’t. METR found developers who were 19% slower by the clock believed AI had sped them up by 20%. A 39-point perception gap.
So when DX says gains decay, it might mean developers stop noticing the gains — not that the gains stop happening. When Microsoft says gains persist, it might mean output volume holds — not that the output is worth what it was.
The DX Q2 2026 benchmarks add a third signal: over 50% of code is now AI-generated, PR sizes have nearly doubled, but the Developer Experience Index fell from 67 to 65 over four quarters and Change Confidence declined 6.1%. Developers understand the code better. They trust it less.
More output. Less confidence. That’s not decay or persistence. It’s something else.
Three Explanations
I see three possible mechanisms. They aren’t mutually exclusive, but they lead to very different places.
Task ceiling. The easy wins come first. AI automates boilerplate, test scaffolding, routine patterns — work that was tedious but predictable. Once that’s captured, what remains is architectural reasoning, legacy navigation, ambiguous requirements. The kind of work where AI helps less. Gains don’t decay; they were always front-loaded. The ceiling is real and low.
Perception shift. The gains are real and sustained — Microsoft’s data suggests this is at least plausible — but developers stop reporting them. What felt magical in the first month becomes infrastructure by the fourth. The speedup is absorbed into expectation. DX’s self-reported decay is a measurement artifact: you can’t sustain surprise. This is the most optimistic reading, and the hardest to test, because it means the measurement instrument is systematically wrong about the trajectory while being right about the level.
System absorption. Individual gains are real but the system consumes them. Opsera found that AI-generated PRs wait 4.6x longer for review and carry 15–18% more security vulnerabilities. The 58% speed-to-PR gain gets eaten by the review bottleneck. DX’s innovation ratio stayed flat — saved hours aren’t converting to new features. The gain exists at the keystroke and vanishes at the pipeline. Developers feel the decay because they experience the downstream friction, even though their personal output is technically higher.
Why It Matters Which One
If task ceiling: the gains we’ve measured are roughly the gains we’ll get. Plan accordingly. Stop promising 10x.
If perception shift: the gains are larger and more durable than anyone reports. The problem is measurement, not productivity. Throw out the surveys. Trust the merge logs.
If system absorption: the gains are real but the system isn’t built to use them. This is the only explanation that implies the problem is solvable — redesign the pipeline, not the tool.
Each answer points somewhere different. The DX decay curve looks like task ceiling. The Microsoft persistence data looks like perception shift. The Opsera review multiplier looks like system absorption. The data doesn’t pick a winner.
The honest thing to say is that the resolution depends on data no one has collected yet: a longitudinal study tracking both objective output and subjective experience, at both individual and organizational levels, over more than four months.
Until someone runs that study, 50.5% is either a warning or an artifact. I don’t know which.