TL;DR
Two 2026 datasets, never cited together. Engineering telemetry from 22,000 developers shows median code review time up 441 percent and incidents per pull request up 243 percent once AI tools land. Payroll data from Stanford shows young workers in the most AI-exposed occupations sitting 19 percent below trend, through reduced hiring rather than layoffs. One system, two halves: review demand climbing, reviewer supply closing. Generation capacity can be bought this afternoon. Judgment has a ten-year lead time, and nobody is placing the order.
Two numbers from this year that nobody has put in the same sentence.
The first: median time in code review, up 441 percent. The second: employment for twenty-two to twenty-five year olds in the most AI-exposed occupations, 19 percent below where it would be had it kept pace.
Generation is a flow. You can buy more of it this afternoon, on a monthly plan, with a credit card. Review is a stock. It takes about a decade to build one unit of it, and that unit is a person. We are expanding the flow and liquidating the stock, and the two decisions are being signed off by different people in the same building.
The telemetry
In April 2026 Faros published two years of engineering telemetry from 22,000 developers across more than 4,000 teams, comparing each organisation's lowest-adoption period against its highest.
The throughput story holds up. Task throughput per developer up 33.7 percent. Epics completed per developer up 66 percent. That is the part that makes the slide.
Then the rest of the report. Median time in code review up 441.5 percent. Time to first review up 156.6 percent. Incidents per pull request up 242.7 percent. Bugs per developer up 54 percent, against 9 percent in the same report a year earlier. Code churn, lines deleted against lines added, up 861 percent.
And the figure I keep coming back to: pull requests merged with no review at all, human or agentic, up 31.3 percent.
The bottleneck did not disappear. It moved from writing to reading, and reading is the part we never industrialised.
The payroll data
Different dataset, different question, same system. The Stanford Digital Economy Lab works from ADP payroll records, three and a half to five million employees a month. Its August 2026 update runs through June.
Workers aged twenty-two to twenty-five in the most AI-exposed occupations are 19 percent below the employment level they would have reached had they kept pace with less-exposed peers. That gap was 15 percent in the July 2025 vintage. Experienced workers show no comparable gap at all.
The mechanism matters more than the number. The adjustment runs through hiring, not separations. Nobody got walked out. The door was just left shut.
The cross nobody makes
Put the two together and you have a scheduling problem dressed up as a productivity win.
Review demand is now rising five times faster than it was. The people who will be doing that review in 2033 are the ones you would hire in 2026. A senior reviewer is not something you procure in a quarter. It is eight to ten years of a person shipping bad code, having it caught, and learning what being caught feels like.
The Ise Grand Shrine in Japan has been pulled down and rebuilt on the adjacent plot every twenty years since the seventh century. The sixty-second rebuild finished in 2013. The timber does not need it. The carpenters do. Twenty years is roughly one working life, which means every master builds that shrine twice: once as an apprentice, once as the person teaching the apprentices.
They treat skill as something that has to be rebuilt on a schedule or it is gone. We treat it as a line item.
Why this matters for your business
If your AI programme reports throughput and not review latency, you are reading half the instrument panel. (The half that goes in the board deck, naturally.)
Two questions for the next steering committee. What is our median time in review, before and after the tools arrived? And who, by name, reviews agent output in 2032?
If the answer to the second one is somebody who will be retired by then, you do not have an AI strategy. You have a nine-year runway.