Key figures
- 4.15 Merged pull requests a week per engineer, up from about 2.5 (Superhuman on the AWS blog, published Nov. 10, 2026)
- 60% Increase, by Superhuman's count (Superhuman on the AWS blog, published Nov. 10, 2026)
- 20 hrs Median time for a change to get through (Superhuman on the AWS blog, published Nov. 10, 2026)
This is a before-and-after comparison, the post describes no control group, and it describes other changes to the engineering system, so the 60% is Superhuman's attribution, not a proven effect of AI.
It won't surprise me to find that many readers don't know what a "pull request" is. So let me offer a primer. A pull request, or PR, is a programmer's request to add a finished change to the code a team shares. The programmer writes the change on a separate copy, then asks for it to be pulled in. Other engineers read it, automated tests run against it, and it may be sent back for fixes. Once it is approved, it is merged into the main code, the version the product is built from. A merged pull request is one change that made it all the way through.
A piece of code gets written. Then it waits.
That is roughly how it goes at Superhuman, the company that used to be called Grammarly. The 410 engineers who were there in both the first half of 2025 and the first half of 2026 went from about 2.5 merged pull requests a week each to 4.15. Superhuman calls it a 60% increase. It describes the gain as "more output from the same people." The count comes from the company's own version-control systems, and the account appears in an AWS blog post dated Nov. 10, 2026, written by three people from Superhuman and two from AWS.
Superhuman attributes the result to rolling AI out as a platform, with support channels, documentation and tooling, rather than as a license purchase. That is attribution, not proof. The post describes no comparison group. The post also describes other changes: a fast lane that auto-approves low-risk changes, build work on one shared repository that returned more than 20 hours of CI runtime a day, and a move of services from Amazon ECS onto Amazon EKS.
The writing got faster. The rest of the walk did not. Each change still takes about as long to get through, a median of about 20 hours. "Generation accelerated, but review, CI, and merge stayed human-paced," the post says. The first reviewer does arrive sooner, after a median of 36 minutes instead of 54, in part because an AI reviewer now takes the first pass. Still, the front of the line gets farther from the back, and everyone reaches the end at about the same hour.
Then there is the heavy thing at the bottom of the pack. According to the post, 28% of pull requests require rework, and reworked changes take 5.7 times longer. Nobody remembers packing it.
So the money is going into checking too. Superhuman says its current platform investment goes into AI code review with multiple reviewers, where the strictest verdict wins, auto-approval only for explicitly low-risk classes of change, and quality gates in every pipeline. Engineers still hold the merge button.
What the Index says: Evidence grade 2. The figures are measured from Superhuman's own systems on a defined group of engineers, but they appear on a vendor-hosted AWS blog with no independent source. This is a before-and-after comparison, the post describes no control group, and it describes other changes to the engineering system, so the 60% is Superhuman's attribution, not a proven effect of AI. Merged pull requests count shipped work, not its value. The post says early adoption brought "a real dip in quality" and recommends tracking revert and change-failure rates, but it publishes neither. Engineers also report saving about 6.5 hours a week; that figure is self-reported.
Sources:
AWS Migration & Modernization Blog, "Superhuman's AI advantage: measuring the business value of AI adoption," by Artur Chiriac, Chandini Bansal, Giles Douglas, Jen Dakin and Khrystyna Shlyakhtovska (Nov. 10, 2026)
Technology Case Study Lead story
The Line Stretches
Superhuman says the 410 engineers active in both periods went from about 2.5 to 4.15 merged pull requests a week, which it calls a 60% increase. Each change still takes about 20 hours to get through. Before-and-after on a vendor-hosted AWS post.
Read the story