Deploy free

AI-assisted PR percentage: measuring AI coding adoption, honestly

AI-Assisted PR Percentage is the share of merged pull requests carrying an AI-assist marker — a commit trailer or a PR marker — rolled up across the whole window and weighted by count, so one slow current week can't drag the number around, with a weekly sparkline underneath. It's the closest single number you'll get for measuring AI coding adoption across a team. It also comes with a real asterisk, and this page is mostly about that asterisk.

deckgauge · Board · Intelligence
AI-assisted PRs
34%

AI-assisted PR %, trending up.

What it measures, and where the data comes from

For every merged PR in the window, Deckgauge checks whether it carries a recognized AI-assist marker — most commonly a Co-Authored-By commit trailer left by a tool like Claude Code, but also PR-level markers some tools stamp instead. The headline % is the count of marked PRs over total merged PRs across the entire window, not an average of each week's percentage — that distinction matters, because averaging weekly percentages lets one quiet week (five PRs, three marked) swing the number as hard as a busy one (two hundred PRs, sixty marked). Count-weighting fixes that.

  • GitHub — merged pull requests, checked for a trailer on any of the PR's commits or a marker in the PR body.
  • GitLab — merged merge requests, same trailer/marker check.
  • Azure DevOps — completed pull requests, from the repos you've opted into code intel for, same check.
Absence isn't evidenceThis is a marker detector, not a lie detector. It only counts PRs where the tool actually stamped something machine-readable. If your team's assistant doesn't leave a trailer, or a squash-merge strips it, or someone's using a tool Deckgauge doesn't recognize yet, adoption reads as zero here no matter how much AI actually touched the code. Read a low or 0% number as "no markers found" — never as "no AI used" — and don't hand this figure to a VP as if it were exact.

How to read it

Look at the sparkline before the headline number. A single week's % is noisy on a small repo — eight PRs merging is enough to swing a percentage ten points on its own. The trend across the sparkline is the signal: is the line generally rising, flat, or falling over the last eight weeks.

The headline number itself is best read as a floor, not a ceiling, on real adoption — see the callout above. Two teams both reading "20%" don't necessarily have the same adoption; one might genuinely be at 20%, the other might be at 60% with a tool that doesn't tag its commits.

What it tells you over time

A sparkline that keeps climbing week over week, even slowly, is the strongest read this widget gives you: adoption spreading through the team, not just one early adopter having a big month. A flat line after a tooling rollout is worth a direct question in your next 1:1s — either people aren't using the new tool, or they are and it isn't leaving a marker. Those are very different problems with very different fixes.

Example situations

1. A suspiciously low reading right after a rollout

deckgauge · AI-Assisted PR %
AI-assisted PRs
3%

3% two weeks after announcing an AI-assist tool.

What you're seeing: a flat 3% line, barely moving week to week, two weeks after the team was told to start using an AI coding assistant. On paper this looks like adoption failed.

How to react: before you conclude anything about the team, go check whether the tool actually stamps a marker. Plenty of assistants only add a trailer if you use their CLI or IDE integration a specific way — a plugin misconfigured, or people pasting code manually instead of committing through the tool, will produce exactly this flat-low line with real adoption happening underneath it.

Managerial playAsk two or three engineers directly whether they're using the tool, separate from what this widget says — a five-minute Slack thread settles the "is it adoption or is it detection" question faster than staring at the sparkline. If they are using it and the marker isn't showing up, that's a tooling/config fix, not a people problem: get the trailer wired correctly before you draw any conclusion about who's on board. Don't schedule an "AI adoption" all-hands over a number that might just be a broken commit hook.

2. A sparkline climbing steadily for two months

deckgauge · AI-Assisted PR %
AI-assisted PRs
52%

Adoption spreading past the early adopters.

What you're seeing: a steady, uninterrupted climb from roughly a fifth of merged PRs to over half, across eight weeks. This isn't one contributor going all-in on a new tool — a climb this smooth and this long usually means it's spreading person by person.

How to react: this is a genuine, believable adoption trend, and it's worth confirming who's driving it before you assume it's evenly spread. Pair it with the AI Adoption table to see whether this is broad or three people carrying the whole number.

Managerial playIf it's broad, say so out loud — a real, visible adoption number is one of the few AI metrics worth actually celebrating in a team update. If it's narrow, go talk to the people driving it: what made the tool click for them, and can that get turned into a five-minute internal demo instead of a top-down mandate. Adoption that spreads because a peer showed someone a shortcut sticks a lot better than adoption that spreads because a policy said so.

Frequently asked

How do you measure AI coding adoption across a team?
AI-Assisted PR Percentage is the share of merged pull requests carrying an AI-assist marker — a commit trailer or PR marker — across the selected window, weighted by count so a single slow week cannot swing the figure, with a weekly sparkline beneath it.
What is a good AI-assisted PR percentage?
There is no benchmark tier, deliberately. The number is only meaningful against your own trend and your own rollout: what matters is whether it is rising after you bought seats, not how it compares to another company.
Why might the percentage understate real AI usage?
Because it is marker-based. If a team uses an assistant that does not write a Co-Authored-By trailer or PR marker, that work is invisible here. A low reading is never proof AI was not used — check tool configuration before drawing conclusions.

Related widgets

Last updated