From an AI invoice to a feature's ROI, in four joins.
TraceROI does one thing carefully: it connects the systems that already hold the answer (your AI vendors, your code host, your tracker, your billing) and shows its working at every step.
Cost → person
Anthropic, OpenAI, Cursor and Copilot report usage by user, API key or seat. We map every identity (work email, GitHub login, API key) to one person and team.
Anthropic · OpenAI · Cursor · CopilotPerson → PR & ticket
We read PR metadata (author, branch, files, timestamps, reviews) and the ticket key in branch names or titles, e.g. feature/PAY-123-upi.
GitHub · Jira · LinearTicket → sprint, epic, feature, module
Tickets roll up to sprints and epics. Epics map to features. Changed file paths map to modules (via CODEOWNERS or your own path rules).
your structure, not oursFeature → revenue (optional)
Connect Stripe, or upload a simple CSV of feature, period and revenue. Revenue per AI rupee is shown as a directional signal, never claimed as cause and effect.
Stripe · CSVHow a rupee finds its ticket
For each person, for each day, for each tool, we take that day's AI cost and place it using the most precise evidence available. Every allocation carries a confidence score, and every dashboard shows the mix.
| Method | Evidence | Confidence |
|---|---|---|
| Exact session | ledger-hook recorded the session, its branch/ticket and its cost | 0.95 |
| Session time split | Sessions tagged with tickets but without per-session cost: the day's cost is split by session duration | 0.85 |
| Active-PR split | No hook: the day's cost is split across the PRs that person had open that day | 0.60 |
| Unattributed | No linked work. Shown openly as exploration / other, never hidden | 0 |
Splits are exact to the micro-rupee: nothing is rounded away, so the traced total always equals the invoice.
ledger-hook, the optional precision layer
A tiny open-source CLI each developer can install in two minutes. It registers as a Claude Code hook and a git hook, and sends one small event per AI session: who, which tool, start and end time, repo, branch, ticket key and cost when known. No prompts, no code, no file contents.
It also covers what vendor APIs can't. For example, Anthropic's Claude Code analytics covers API-billed usage but not Pro/Max/Team subscription seats.
How we calculate ROI (and what we refuse to claim)
1. Time saved
For each AI-assisted ticket we compare its cycle time against a baseline: the same team's non-AI tickets of the same type and size over the previous 90 days. The difference, multiplied by the loaded hourly cost you set per role, is time saved in rupees.
2. Efficiency ROI
Time saved (₹) ÷ AI cost (₹), per ticket, feature, team or tool.
3. Revenue per AI rupee
Feature revenue ÷ AI cost attributed to that feature. Useful for comparing where AI money goes. It is labelled directional, because revenue has many causes.
4. Bottlenecks
Time from PR open to first review, AI vs human PRs; PRs waiting more than 48h; rework rate (follow-up fix PRs touching the same files within 14 days) per module.
What we won't do
- Rank or score individual engineers. Views are team-level by default.
- Use "lines of AI code accepted" as a productivity measure.
- Hide unattributed spend to make the numbers look better.