AI Gave the World More Code. Now the World Needs More Reviewers.
High-AI teams merge 98% more pull requests, yet review time is up 91% and 31% of an engineer's day is invisible work. Here is where delivery stalls.
Key Takeaways
Teams with high AI adoption merged 98% more pull requests, but PR review time rose 91%, per Faros AI telemetry from more than 10,000 developers across 1,255 teams.
Harness's 2026 survey of 700 engineering practitioners found 31% of a developer's day now goes to AI-related validation work. Only 38% of organizations track the time spent reviewing AI-generated code.
Plandek's benchmarks of 2,000+ teams show top-quartile teams ship in under 22.5 days and bottom-quartile teams take more than 62. AI narrowed that gap without closing it.
Teams that directly attack their planning, review, and quality constraints deliver more than 2x the output per engineer of teams that don't.
The job moved from writing code to judging it. Most companies still hire, manage, and measure as if it hadn't.
What does everyone believe about AI and engineering output?
The conventional view says AI coding tools make engineers faster, so more tools mean more shipped product. Buy the licenses, watch pull requests climb, and delivery follows. At the individual level that holds up: Faros AI found developers on high-adoption teams complete 21% more tasks and merge 98% more pull requests.
The engineers themselves agree. Info-Tech's June 2026 study of 578 software leaders found 94% of developers reporting meaningful productivity gains, and 83% reporting meaningful defect reduction in shipped code. If your team's pull request count jumped after you rolled out an AI coding tool, you've seen the same thing.
So why is your roadmap still slipping? A merged pull request isn't a shipped feature, and the belief treats the two as one event.
Why is faster code producing slower shipping?
Because writing code was never the constraint. It sits upstream of review, testing, and release, and those stages didn't speed up. Faros found PR review time rose 91% and average PR size grew 154% on high-adoption teams. Engineers now hand reviewers far more code, and the reviewers are the same people working the same hours.
Why doesn't individual speed add up to company speed?
The gains leak out at every handoff. Faros reported that the correlation between AI adoption and company-level performance metrics disappears once you look past the individual developer. Review queues, brittle test suites, and slow release pipelines absorb the extra volume before it reaches a customer.
What is the invisible work taking a third of the engineering day?
It's validation: reviewing AI code for accuracy, fixing subtle bugs the AI introduced, and explaining generated code to teammates. Harness's 2026 survey of 700 practitioners puts it at 31% of a developer's day, and only 38% of organizations track any of it.
Sapio Research ran the survey for Harness in April 2026. Of the engineering leaders who responded, 81% said code review time has risen sharply since they deployed AI. Among developers, 53% named accuracy review as their top friction, 52% named fixing subtle bugs, and 48% named explaining AI-generated code to teammates.
That's nearly a third of every engineering salary on your payroll going to work that appears on no dashboard. A function that used to take an hour to write now appears in seconds, and the hour goes to reading it, testing it, and working out why it does something slightly different from what you asked. Your engineers now edit, audit, and judge code they didn't write, and that work is harder to see than typing ever was.
What does the data actually show about which teams win?
The teams that fixed their process, not the teams with the most licenses. Plandek's 2026 benchmarks of more than 2,000 teams show bottom-quartile teams using AI cut lead time to value from 62 days to 33, while top-quartile teams moved from 22.5 days to 20. The gap narrowed but didn't close.
Metric | Low performers | Top performers |
|---|---|---|
Average delivery time | More than 62 days | Under 22.5 days |
Time to merge a pull request | 35+ hours | Under 21 hours |
Planned sprint work completed | Less than half | Over two-thirds |
Lead time improvement from AI | Nearly 50% | 10 to 15% |
AI helps struggling teams roughly four times more than it helps strong ones. Even so, a low-performing team lands at 33 days, about ten days behind where top-quartile teams started. Everyone has the same tools, so the difference sits in the planning, review, and quality practices around them.
Plandek found that teams that directly address planning, review, and quality constraints achieve more than 2x the output per engineer compared to teams that don't. Info-Tech's data shows how far most companies are from that. Only 37.4% of organizations using AI at the build stage describe their AI maturity as formal or better, and 67% agree that AI-generated code needs more testing than human-written code. Adoption is moving much faster than process.
Measurement makes all of this worse. Harness found 54% of developers fear AI productivity data will be used against them in performance reviews. If your engineers believe honest numbers get punished, you'll receive flattering ones, and you can't manage a bottleneck you're never shown.
When is the conventional view right?
It's right about raw output, and it's right when your pipeline had slack to absorb the extra volume. AI does raise task throughput, and weaker teams gain the most from it. The evidence against it is also thinner than the headlines suggest.
The best-known counterexample is the METR trial, which found experienced open-source developers took 19% longer on tasks with AI while believing they'd been 20% faster. That study covered 16 developers in early 2025, and its confidence interval ran from 2% to 39%. In February 2026 METR said the result likely understated AI's benefits, though it also called its newer data an unreliable signal. What it firmly shows is that people's sense of their own speed can be badly wrong. It doesn't show that AI slows engineers down.
Two more caveats apply. Info-Tech's 94% is self-reported, and self-reports are precisely what METR caught misfiring. And the vendors behind three of these studies (Harness, Plandek, and Faros AI) sell tooling that measures engineering teams, so they have a reason to say measurement is broken. The numbers still point the same way across all four sources.
What should you do differently?
Measure the validation work, then staff and plan around it. Track review time on AI-generated code, make ship rate your headline metric instead of pull request count, and fix the slowest stage in your pipeline before you buy another license. Harness's own five recommendations point the same direction.
Instrument review. Only 38% of organizations do. Start with hours from pull request opened to merged, and the share of review comments that land on AI-written code.
Report ship rate, not volume. Features delivered, share of planned sprint work completed (top teams finish more than two-thirds), and lead time. Pull request counts come off the executive dashboard.
Find your constraint. Plandek's data points at planning, review, and quality. Pick whichever is slowest in your pipeline and put your best people on it.
Hire for review judgment. Give candidates a flawed AI-generated diff in the interview and ask what they'd send back. An engineer who catches the subtle bug is worth more now than one who types fast.
Make the data safe to share. Say in writing that review-time metrics won't feed individual performance reviews, since 54% of developers fear exactly that.
International Digital Partners places senior nearshore engineers in DevOps, software engineering, and cloud who work inside U.S. business hours. That matters more once review is your bottleneck, because a pull request that waits overnight for a reviewer costs a full day, and overlap hours remove that wait.
Sources
The Invisible Burden: How AI is Redefining Developer Productivity in 2026 (SD Times, on the Harness State of Engineering Excellence 2026 report)
AI Helps Low-Performing Engineering Teams 4x More Than High-Performing Ones, New Benchmarks Show (Plandek 2026 Engineering Productivity Benchmarks)
94% of Developers Report AI Productivity Gains, but Governance Maturity Lags Behind Adoption (Info-Tech Research Group)
The AI Productivity Paradox Research Report (Faros AI)
We are Changing our Developer Productivity Experiment Design (METR)
More Articles
AI Gave the World More Code. Now the World Needs More Reviewers.
High-AI teams merge 98% more pull requests, yet review time is up 91% and 31% of an engineer's day is invisible work. Here is where delivery stalls.
Nearshore Teams Are Outpacing US Shops on AI Coding Tools
Nearshore teams already work async and document everything. That habit, not geography, is why they adopt AI coding tools faster than most U.S. shops.
Nearshore vs. Offshore Cost Comparison: The Real Numbers
Offshore developer rates undercut nearshore by 40 to 70 percent on paper. Attrition, lost overlap, and rework close most of that gap in practice.