Buried in the methodology notes of Linear’s How Teams Build report is a sentence that deserves to be printed above every chart: “We count pull requests opened, not merged.”
The chart above that footnote is the one doing the rounds. Teams that connected a coding agent went from 21 to 65 pull requests a week over two years, while teams without one went from 8 to 10. Tripled. Across all 47,900 workspaces in the sample, PRs opened are up 111% on a June 2024 baseline. That number is now turning up in board decks and budget reviews as proof the AI spend is paying for itself.
An opened PR is a unit of activity, not a unit of shipped software. Linear knows this and says so in the same report: an opened PR says nothing about the value of the change. The caveat is right there on the page; plenty of people quoting the chart skip the footnote.
The denominator is in someone else’s report
Linear does not publish merge rates for its cohort. That absence is the whole story. Without a denominator, “3× the PRs” is unfalsifiable as a productivity claim, which is precisely what makes it so quotable.
You have to go to LinearB for the denominator. Their 2026 benchmarks, built on 8.1 million PRs across 4,800 teams, found AI-generated pull requests merge at 32.7%, against roughly 84.5% for human-authored code. The AI PRs also wait over 16 hours for a reviewer to pick them up, compared with about 200 minutes for unassisted work. More than five times longer in the queue, and less than half as likely to land.
Run the arithmetic the chart invites
To be clear about what follows: the calculation is mine, it splices two datasets that measure different populations, and Linear makes no such claim. But it is the estimate the volume chart begs for.
Take the agent-connected team opening 65 PRs a week. Apply LinearB’s 32.7% merge rate for AI-generated code and you get about 21 merged PRs a week. That team’s 2024 baseline was 21 PRs opened; at the 84.5% human merge rate, call it 18 merged. So the honest version of “we tripled our output” may be “we merge about three more PRs a week than we did two years ago.” Roughly 20% more shipped code. Real, worth having, and an order of magnitude short of the headline.
If Linear’s agent-connected teams merge AI code at better than 32.7%, the gain is bigger. Nobody can say, because the number isn’t published. What we can say is what happens every time someone counts merged code with a control group.
Every careful count lands in the same band
Atlassian studied its own Rovo Dev agent across 3,400 repos from 2,500 customers, using propensity score matching to compare adopters against similar non-adopters, and counted merged PRs. Result: 19% more merged pull requests per month, or three to five extra PRs landing. DX, measuring across 400+ companies, found a 65% increase in AI tool usage moved median PR throughput by about 8%, with most organisations in the 5–15% range.
Atlassian’s matched repos, DX’s cross-company medians, and my envelope arithmetic on Linear’s own cohort all land between 5% and 20%. Whenever someone counts merged code with a credible comparison group, the answer sits in that band. The only way to reach 3× is to stop counting at “opened.”
And volume carries a cost even when it never merges. The two-thirds of AI PRs that don’t land still consume review attention, the scarcest resource an engineering team has. Underneath, GitClear’s analysis of 623 million code changes shows what the surviving volume is made of: copy/paste up 41%, duplicated blocks up 81%, two-week churn up 15%.
Be fair to Linear, hard on the usage
Linear’s data is honest about what it measures, and its sample skews the way you would expect: teams that chose Linear and then wired up coding agents sit at the enthusiast end of the market, so their numbers are a ceiling for everyone else. The report prints its caveats. The failure mode is entirely downstream, in the founder who screenshots the 21-to-65 chart into an investor update, and the engineering leader who lets a volume metric stand in for a delivery metric because it is the flattering one.
So before you cite it: count merged PRs, or better, deployed changes. Watch pickup time, because a review queue filling faster than it drains is a cost dressed up as output. And when any vendor shows you a productivity chart, ask the one-word question that deflated this one: opened, or merged?
A 20% gain in shipped code is a good outcome and a defensible budget line. Claiming 3× sets you up to be the person explaining, two quarters from now, why delivery didn’t triple. If next quarter’s AI spend rests on a volume chart, have us check the denominator first.
Sources