AI Adoption Metrics for Engineering Leaders: What to Track and What to Ignore
2026-11-14 · 9 min read · GitHub & Metrics
Every engineering leader is being asked some version of the same question: is AI making us faster, and how much are we using it?
The market has answered with a flood of numbers. Acceptance rates. Lines of AI-generated code. Seats activated. "Time saved". Most of them are easy to produce and hard to interpret, and a few are actively misleading.
This guide sorts them into three groups: the metrics worth tracking, the ones to use carefully, and the ones to ignore. Then it covers how to report AI adoption upward without overstating it.
The Principle: Measure the Work, Not the Tool
A tool's own dashboard tells you how the tool was used. It can't tell you whether the work that shipped changed. The useful AI metrics come from the work your team already produces — commits, pull requests, reviews, incidents — so they're comparable across tools and over time, and they connect to outcomes you already care about.
Keep that test in mind for every number below: does this come from the work, or from the tool's activity log?
Track These
1. Share of work with an AI trace
The share of your team's commits and PRs that a coding agent signed — through a Co-Authored-By trailer, a PR footer, or a bot author. (If that's new, start with what trailers are and how agents use them.)
Why it's worth tracking:
- It comes from the work, across every tool, in one number.
- It's comparable week to week.
- It reflects agent use — the fastest-growing kind and the one that changes how work flows through review.
How to read it: as a floor. Autocomplete and pasted chat answers leave no trace, so real AI use is higher, by a different amount per person.
2. Consistency: weeks with any AI-assisted work
Out of the weeks someone (or the team) did any work, how many included some AI-assisted work? This separates a habit from a one-off. An engineer who used an agent for one big migration and never again looks heavy on share and light on consistency.
Blending the two gives a fair single score. One formula that works well is 70% share plus 30% consistency over six weeks — covered in detail in how to measure AI adoption.
3. The trend, not the level
A team at 15% and rising is in a different position from a team at 35% and flat. Always show the weekly trend next to the current number. Most of the insight is in the direction.
4. Delivery metrics, side by side
AI adoption on its own says nothing about outcomes. Put it next to the metrics you already trust:
- PR cycle time — is work moving faster, or piling up in review?
- DORA metrics — deployment frequency, lead time, change failure rate, time to restore.
The interesting stories are in the combinations: adoption up and cycle time down; adoption up and change failure rate creeping up with it; adoption up and nothing else moving.
5. Review health
More agent-written code means more weight on reviewers. Track a couple of simple review signals as adoption grows:
- PR size — are agent PRs much larger than human ones?
- Time to first review and review depth — are big agent PRs getting a quick "LGTM"?
See code review in the AI era for what good looks like.
Use With Care
Licence and seat counts
Useful for budgeting. Useless as an adoption measure. Forty seats can mean forty heavy users or six. Track "seats" only next to a work-based number, so the gap is visible.
Vendor acceptance rates
"Developers accepted 30% of suggestions." This measures autocomplete behaviour inside one tool. It's fine for comparing settings within that tool. It doesn't tell you how much of the shipped work was AI-assisted, and it can't be compared across tools.
Developer surveys
Surveys are the right way to learn how people feel — whether the tools help, where they get in the way, what training would help. They're a poor way to measure usage or speed. Several studies have found developers' sense of how much faster AI makes them doesn't always match measured results (we discuss this in vibe coding and your team). Run them, read them, but don't put the self-reported speedup in a board deck as fact.
Ignore These
Lines of AI-generated code
The most quoted metric and the least useful. Lines reward verbosity, penalise good refactoring, and are dominated by whoever generates the most boilerplate. A trailer applies to a whole commit, so even "lines in AI commits" mostly measures who makes large commits.
"Hours saved"
Usually an estimate multiplied by an assumption multiplied by a headcount. If a number can't be checked against anything that happened, it isn't a metric.
Per-person AI rankings
Ranking engineers by AI usage turns a team insight into a target. People add trailers to climb the list, or strip them to avoid scrutiny, and the metric stops working. A per-person score is useful inside a 1:1 as a conversation starter (see how to talk about AI usage in 1:1s). It doesn't belong on a leaderboard or in a performance review.
A Simple Dashboard
If you only keep five numbers, keep these:
| Metric | Source | Cadence |
|---|---|---|
| Share of commits and PRs with an AI trace | Git history, all repos | Weekly trend |
| Team adoption score (share + consistency) | Git history, 6-week window | Monthly |
| PR cycle time | Pull requests | Weekly |
| Change failure rate | Deploys and incidents | Monthly |
| Median PR size | Pull requests | Monthly |
Five numbers, one source of truth, all from the work.
Reporting Upward
Leadership usually wants three things: the trend, an honest caveat, and a "so what". Three sentences cover it:
- Trend: "Agent-assisted work is 31% of our commits and PRs, up from 19% last quarter."
- Caveat: "This counts work an agent signed. Autocomplete isn't counted, so real usage is higher."
- So what: "Cycle time improved and change failure rate held, so we're getting faster without trading off quality. Next quarter we're spreading the two workflows that work best."
Resist the urge to convert this into "hours saved" or "engineers' worth of capacity". Leaders remember the number you give them, and an inflated one will come back.
How Emtricks Helps
Emtricks gives you the first three metrics out of the box: the share of AI-assisted work across every repo your team touches, a fair six-week adoption score for the team and each engineer, and the weekly trend — next to PR cycle time and delivery metrics in Pulse. The Team Report includes the AI adoption section with its caveat written in, ready to send upward.
Key Takeaways
- Measure the work, not the tool: commits and PRs an agent signed, across all repos.
- Track share, consistency and trend — and always next to delivery and review metrics.
- Treat licence counts, acceptance rates and surveys as context, not adoption measures.
- Ignore lines of AI code, "hours saved" and per-person rankings.
- Report upward in three sentences: trend, caveat, so what.
Track AI adoption next to cycle time and delivery — from your Git history, with nothing to install. Try Emtricks free for 30 days.