Measure adoption from shipped work
Editor telemetry records interactions with an assistant. Commit and pull-request attribution records evidence associated with code entering the delivery system; the two measure different forms of adoption.
Adoption measures include AI-assisted commit and pull-request rates, detected tool distribution, and trends over time. Model information is available only when the attribution source supplies it.
- AI commit rate: detected AI-assisted non-merge commits divided by non-merge commits.
- AI PR rate: merged pull requests with at least one AI signal divided by merged pull requests.
- Tool breakdown: which supported coding assistants contributed to detected work.
- Confidence and method: the evidence used to classify a commit or pull request.
Compare adoption with delivery outcomes
More AI activity does not establish better delivery. Compare detected AI-assisted and unclassified pull requests within similar repositories, work types, and time windows.
Use the comparison to investigate whether AI-assisted work is moving faster, arriving in larger pull requests, or bypassing review more often. The result is a starting point for a team conversation, not an individual performance score.
- PR cycle time for detected AI-assisted versus unclassified pull requests.
- Median PR size based on additions and deletions.
- No-review rate for both groups.
- Changes in adoption and delivery across selected time windows.
Local attribution and coverage
Tempo CLI runs as local git hooks and can match committed files with local AI session data. It produces structured attribution metadata before an optional sync to Tempo cloud.
Tempo also detects explicit commit trailers and pull-request patterns. Attribution records retain a method and confidence; repositories with Entire Checkpoints supply a separate source of session context.
A missing signal does not prove that a commit was written without AI. Detection coverage depends on the evidence available.
Questions to investigate
Use adoption and delivery data to investigate changes in review, batch size, and throughput.
- Which tools are actually present in shipped work?
- Is AI-assisted work waiting longer for review?
- Are AI-assisted pull requests larger or less frequently reviewed?
- Which repositories have adopted AI without degrading flow?
- Where should enablement, guardrails, or review capacity change?
Measure AI coding across four layers
A useful AI measurement program separates adoption from impact. Adoption describes where assistants are present; flow, quality, and cost describe what changed after that work entered the delivery system.
Read the layers together. A higher AI commit rate with shorter coding time can still create a review bottleneck if pull requests become larger or arrive faster than reviewers can absorb them.
| Layer | Signals | Decision it supports |
|---|---|---|
| Adoption | AI commit rate, AI PR rate, active tools | Where enablement or license coverage is changing |
| Flow | Cycle time, pickup time, review time, throughput | Whether AI-assisted work moves through delivery faster |
| Quality constraints | PR size, no-review rate, rework and failure signals | Where guardrails or review capacity need attention |
| Cost and coverage | Tool mix, attribution coverage, spend when available | Whether the evidence is complete enough to support an investment decision |
Review cadence
Start with a baseline before a rollout or policy change. Review adoption weekly while behavior is changing, then compare delivery outcomes over a longer window that contains enough completed pull requests to reduce noise.
Use repository or team comparisons to generate hypotheses, not verdicts. Investigate the pull requests behind a change, agree on one intervention, and watch both the intended metric and its balancing signals.
- Baseline adoption, cycle time, PR size, and review behavior before the change.
- Segment by repository or work type before comparing outcomes.
- Choose one intervention, such as smaller batches or explicit review ownership.
- Revisit the trend and document what the team learned.
Frequently asked questions
What is AI coding analytics?
AI coding analytics connects evidence of assistant use to engineering artifacts such as commits and pull requests, then compares that work with delivery outcomes including cycle time, review load, size, and deployment flow.
Does more AI-generated code mean higher productivity?
No. Adoption is an input signal. Productivity requires context from delivery speed, quality constraints, developer experience, cost, and the outcomes the team is trying to create.
Can Tempo identify work from multiple AI coding tools?
Tempo records supported evidence from Claude Code, GitHub Copilot, Cursor, Codex, and Aider. Coverage depends on the local session formats and explicit metadata available for each tool.
What data does AI attribution send to Tempo?
Tempo CLI sends structured attribution metadata, without source contents, diffs, prompts, or conversation transcripts. The separate Entire Checkpoints integration reads prompts, context, and compact transcripts from the repository; those texts may contain code supplied in the conversation.
Sources and further reading
- DORA: Impact of generative AI in software developmentResearch on how AI adoption interacts with software delivery performance and organizational systems.
- SPACE framework for developer productivityThe original multidimensional framework for reasoning about developer productivity.
- Tempo AI coding detection methodologyThe evidence, confidence, privacy, and aggregation rules behind Tempo attribution.