How are PRs considered qualified in the AI Impact report when using traces?
Last updated: October 27, 2026
The AI Impact report measures AI code from Agent Traces: Span matches each pull request's lines to the author's recorded AI sessions (Claude Code, Cursor, Codex, Copilot, and similar). A PR is "qualified" when Span has enough trace coverage across the PR's life to measure AI and human lines reliably.
Why it matters
Qualified PRs are not "PRs that used AI." Qualified PRs tell you how much of your merged work Span could measure. AI usage itself is the AI Code Ratio. If traces were missing, started after the PR was opened, or had gaps while the PR was in flight, including the PR would mean guessing. Those PRs stay in the merged-PR total and are left out of the ratio.
Column | What it answers | How it is calculated |
|---|---|---|
AI Code Ratio | How much of the measurable diff was AI-authored? | AI-attributed lines ÷ eligible lines, only on qualified PRs |
Qualified PRs | How much of the merged work could Span measure? | Qualified merged PRs ÷ all merged PRs in the range |
A low qualified-PR percentage does not mean people are not using AI. It usually means many PRs in the range were opened before traces were collected, or traces were incomplete for those PRs. Read AI Code Ratio for adoption and Qualified PRs for measurement coverage.
How it works
Which PRs are in the denominator. Every PR in Qualified PRs (the "of N") is a merged PR in your date range whose author is a development contributor, and whose repository, branches, and author are not ignored in Settings → Metrics.
Dates. The date range buckets PRs by merge date. Qualification looks at when the PR was opened, because that is when traces need to start.
A merged PR is qualified when at least one of these is true:
Strong trace match. At least 20% of the PR's eligible lines match the author's traces. For a PR opened before trace collection started for your organization or for the author (an author's collection counts as starting 3 days after their first trace), the bar is 60%, because earlier work may not be recorded.
Continuous traces, even with little or no AI matched. The author's latest trace is at or after the merge, and there is no run of two or more weekdays without traces (out-of-office days excluded) from 3 days before the PR opened through 2 days after it merged. A 0% AI PR can qualify this way, which keeps human-written work in the ratio.
Fully agent-authored. Every non-merge commit carries an agent author or agent co-author, or the PR author is an agent account with no person behind it. Span counts these PRs as 100% AI unless the PR's linked sessions show the same tools, in which case Span uses the line match.
You do not need to have used AI for a PR to qualify. You do need traces that cover the PR.
Which lines count. Attribution works on any code file, in any language. Span leaves out files that match your ignore rules, files that are not code (Markdown, plain text, CSV, JSONL, and files with no extension such as Makefile or Dockerfile), and files where the PR only deletes lines.
Why a PR is not qualified. Each non-qualified PR shows one of these reasons:
Reason in the report | What it means |
|---|---|
Author not traced at the time | The PR was opened before trace collection started for your organization or for this author, and less than 60% of its lines matched the author's traces. Merge date does not matter. |
Gaps in author's trace data | The author had no traces, the PR merged after the author's last trace, or there was a gap in traces while the PR was in flight. Span cannot confirm it has the traces needed to attribute the diff. |
No eligible code to analyze | Every changed line was excluded before attribution: by your ignore settings, as a non-code change, or as a file removal. |
Background agent, not traced | The PR was produced partly with a background agent (for example, some commits carry an agent co-author or the description says it was generated by an agent, but not every commit is the agent's). Fully agent-authored PRs can still qualify. |
PR too large to analyze | The diff exceeds the size limit even after ignore rules, so attribution was not run. |
Session too large to analyze | The only traces overlapping the PR were too large to load, so they were skipped. This is not a clean "no AI" result. Span re-checks on later runs in case a smaller overlapping trace appears. |
Too much edit history to analyze | The matching traces contain so much AI-edit history (a file rewritten many times) that attribution was skipped. |
Inconsistent diff data | The diff data is missing or inconsistent (more AI lines than changed lines), so a trustworthy ratio cannot be computed. |
Awaiting processing | Attribution has not finished for this PR yet. |
How to inspect the PRs
The page title is AI Impact. Find it under AI Effectiveness → AI insights → Comparison.
Open the page and select the AI code ratio card.
To the right of the chart, click the underlined Qualified: X out of Y total PRs.
In the drawer, find Non-qualified PRs with a count per reason, then click the View PRs link to list them.
To see one person or team, filter the report first, then click the same Qualified link. The list follows your filters.
Elsewhere in the report:
The breakdown table's Qualified PRs column shows the share and count, for example "40% of 50". These values are not links.
When a team or person has no qualified PRs, their AI Code Ratio shows "-".
Non-qualified PRs form the Unqualified bucket. They stay out of the comparison groups unless you turn on Show unknown PRs.
On a single PR, the PR drawer shows AI percentage as N/A and AI bucket as Unqualified, with the reason in a tooltip.
The date picker cannot start before your organization's trace-collection start. PRs merged in the selected range can still be unqualified if they were opened before that start.
Things to know
Qualification rises over time. Right after rollout, many merged PRs were opened before traces started. The rate climbs as collection runs.
Partial rollouts don't drag the ratio down. Authors with no traces drop out of the cohort instead of adding 0% PRs.
Agent markers matter. If developers leave "Generated with Claude Code" in a PR description but make some commits by hand, that PR is disqualified as "Background agent, not traced".
Fully agent-authored PRs can raise the ratio. A PR where every commit carries an agent marker counts as 100% AI. If many such PRs come from an agent Span does not record, the ratio rises and cost per AI line falls.
Throughput uses developer-weeks. For the throughput comparison, a developer-week qualifies when at least 30% of that person's merged PRs that week qualify.
Worked example: trace collection for your organization starts June 1. A PR opened May 20 and merged June 10 is "Author not traced at the time" unless at least 60% of its lines match the author's traces. A PR opened June 15 with 400 eligible lines, of which 60 match the author's sessions (15%), is below the 20% bar. It still qualifies at 15% AI if the author's traces ran without a two-weekday gap from June 12 through two days after merge. If the author's laptop was off for a week in the middle, it is "Gaps in author's trace data".
Related articles
How AI Code Ratio is calculated
Understanding Agent Traces in Span (auspex)
Need help?
Contact your Span account team or reach out to support@span.app.