Productivity
See whether AI is helping your engineering team get more done, and whether that work holds up.
Productivity answers one question: is AI helping the team get more done, and does it hold up? Every part of the page compares work written with AI against the rest of the team's work, so you can see whether AI makes the team faster without making the work worse. It builds all of it from your source control history, so nobody has to track anything by hand.
Productivity reads from GitHub or GitLab. Connecting one takes the right permissions, so if you cannot, ask an admin.
Connect your source control
Step 1: Connect GitHub or GitLab
Open Productivity and choose GitHub or GitLab. You can also connect from Org Settings, then Integrations.
Step 2: Choose which repositories to sync
Pick the repositories Productivity should read. Nothing syncs until at least one is chosen. To change the list later, open the menu at the right end of the controls and choose Configure.
Step 3: Wait for the first sync
Proxon fetches your recent history first, then keeps reaching further back. The page fills in once the first sync finishes. An organization with many repositories can take a day or more to sync completely.
Each date range needs twice its length in history, because every figure is compared with the period before it. A range you cannot pick yet unlocks on its own as more history arrives.
Step 4: Map authors to people
From the same menu, choose Map people to match source control accounts to people in Proxon. This is what lets Productivity show figures by team and by person.
Focus on the work you care about
Every figure on the page answers to the controls at the top. Narrow the view to particular teams or repositories, and look back over the last 7, 14, 30, 60, or 90 days by day, week, or month. Productivity counts people and AI agents such as Devin by default and leaves out bots such as Dependabot, so routine dependency updates do not pad the numbers. You can add bots back in when you want the full picture.
How Proxon decides work used AI
Proxon counts a pull request as written with AI when one of its commits names an AI tool as a co-author, the way tools such as Claude Code mark the commits they help write, or when its description says it was generated by an AI tool. Every part of the page uses this same rule.
Treat the result as a floor rather than a full count. A tool that marks neither its commits nor the pull request is not seen, so the real share of AI work can be higher than the page shows.
The answer at the top
The card at the top gives the answer in one sentence, such as how much of what engineering merged AI helped write, and whether the team merges more per engineer than before. Below the sentence are the four numbers it is built from, each with a trend line and how it compares with an earlier point in the range:
- Merged with AI is the share of merged pull requests written with AI.
- Merged per engineer is how many pull requests the team merges per engineer each week.
- Cycle time is the typical time from first commit to merge, for work with AI and work without it.
- Merged without human review is the share of merged pull requests that no person other than the author reviewed, for AI-written work and the rest.
Click a number to open the pull requests behind it.
Are we getting more done?
This section covers throughput and cycle time, split by whether AI helped write the work.
Merged pull requests per day, week, or month stacks work with AI on top of work without it, and its title line says whether output is up or down.
Cycle time draws the typical time from first commit to merge for work with AI and work without it, and says which of the two got faster or slower. Click any point to open the pull requests behind it.
Where the time goes splits a pull request into its stages and shows the typical time in each, with AI and without: Ticket to first commit, Writing the code, Waiting for review, In review, and Approval to merge. Its title line names the stage where AI work saves the most time.
Cycle time starts at a pull request's first commit. When Proxon has no first commit for a pull request, it starts when the pull request was opened.
Ticket to first commit needs a work tracker such as Jira or Linear. Without one, the other stages still work from source control alone, and the row offers a link to connect a tracker.
Is it holding up?
This section runs the same quality checks on AI work and everything else, over time. Each card shows the two figures side by side, a line for each over the range, and a chip that says which way AI work is heading.
Merged without human review is the share of merged pull requests that no person other than the author reviewed, for AI-written work and the rest. A review from an AI reviewer, such as GitHub Copilot's code review, does not count as a human review. Its chip reads Rising, About the same, or Falling.
Typical pull request size is measured in lines changed (additions plus deletions). Three in four pull requests are this size or smaller. Its chip reads Growing, About the same, or Shrinking.
Click any point on either chart to open the pull requests behind it.
Where is it working?
This section shows where AI work merges faster without skipping human review, and where it does not yet. Break it down by AI tool, Repository, Kind of work, or Teams with the switch beside the heading. A sentence above the table names where AI pays off most.
Every breakdown shows the same three columns:
- Verdict sums up the row.
- AI work merges says how much faster or slower AI work goes from first commit to merge than the rest, with both times.
- Skips review shows the share of AI work merged without human review, next to the share for the rest.
Each breakdown adds its own columns. AI tool shows each tool's share of AI work, and a pull request that used two tools counts under both. Repository shows how much of the work used AI and which AI rule file the repository carries, the file that tells AI tools how to work there. Click a repository for more detail. Kind of work sorts pull requests by their titles into Upkeep, Refactoring, Bug fixes, Docs, and New features. Click a kind to open its pull requests. Teams shows how much of each team's work used AI and how many of its repositories have AI rules. It fills in once you upload an org chart.
How the verdict is decided
| Verdict | What it means |
|---|---|
| Paying off | AI work merges at least 20% faster than the rest, and does not skip human review more often. A gap of up to 3 points counts as the same. |
| Mixed | AI work is at least 10% faster, but not by enough to pay off, or it skips human review more often. The main reason shows under the verdict. |
| Not yet | AI work merges less than 10% faster than the rest, or slower. The reason shows under the verdict. |
| Too few to tell | Fewer than 20 pull requests with AI, or fewer than 20 without it to compare against. No speed or review comparison is shown. |
What to do next
Cards point out the changes most worth making, ranked by how much each would move the team. A card marked "What's working" names something going well, and one marked "Worth a look" names something to look into, such as a stage of the cycle that slowed down or how often pull requests get a review from a bot. Cards never single out a person. When there is nothing to point out, the section does not appear.
All metrics
The last section puts every measure in one table, from the whole organization down to each team and its members. Expand a row to go a level deeper, and sort by any column. The table counts pull requests merged in the selected range, and its cycle time starts at the first commit, like the rest of the page. Hover a column header to see exactly what it counts. Each row counts the people beneath it once, so a person on two teams never inflates the organization's total. The table fills in once source control authors are mapped to people.
Who can see Productivity
Every role can see Productivity by default, and an admin can turn that off for a role. Rows for individual people in All metrics, and insight cards that name a person, show only to admins and the owner by default; everyone else sees teams and the organization. Connecting source control, choosing repositories, and mapping people are limited to admins and the owner by default.