Your gateway knows what AI cost. GitHub and Jira know what came out the other end. OptScale AI joins the two automatically, so the number you report is derived from your systems of record — not self-reported by the teams being measured.
Derived, not declared
Sound Familiar?
Every organization funding AI at scale hits the same wall at the same moment — the first serious question from finance
You can produce the invoice down to the token. You cannot produce a single unit of output to divide it by
Impact arrives as survey answers and time-saved estimates from the same teams whose budget depends on the answer
Seat counts come up for renewal and nobody can say which teams turned the licence into shipped work
The one figure you did produce collapses the first time somebody asks what it was compared against
The impact join
A proxy sees tokens leaving. It cannot see whether the pull request merged, whether it was reverted, or how many review rounds it took. OptScale AI reads both sides and puts them on the same row.
Session identifiers and commit trailers link the request to the commit deterministically. These are the only joins that feed headline numbers by default.
Timing and authorship proximity, where no trailer exists. Held separately, excluded from headline figures, and labeled wherever they are shown.
Shipped metric packs
Pre-built definitions, dashboards and scheduled reports — or define your own in the wizard in under ten minutes. No query writing, no engineering ticket.
Did the work hold up?
✓Cost per merged PR and accepted line
✓PR open-to-merge duration and review roundsModel output arbitrage
✓Revert rate and post-merge fix churn
✓Defect escape rate by cohort
Where is the budget going?
✓Spend by team, model, provider and intent
✓Cost per feature and per epic
✓Anomalous spend as a separate waste line
✓Savings net of provider prompt caching
Who is actually using it?
✓ Active AI users, measured from gateway auth
✓ AI-assisted share of merged PRs
✓ Adoption trend and model mix by team
✓ Cost per delivery unit over time
A delivery unit is an internal normalization used to compare a team against its own trend over time. It is never presented as comparable across companies or anchored to an industry percentile.
How the join is built
Keep the gateway you already run. Measurement first; enforcement whenever you want it.
1
Route through the OptScale gateway — or install the open-source telemetry plugin into the gateway you already run.
2
GitHub and Jira, read-only. We watch merges, reviews, reverts and ticket transitions.
3
Session and commit trailers give high-confidence joins. Heuristic matches stay separate and labeled.
4
Persona dashboards, scheduled reports to email and Slack, read-only MCP access for your own tools.
Built to survive review
Most AI ROI figures collapse under one follow-up question. These constraints are enforced in the product, not promised in a deck
Merges, reverts and review rounds are read from GitHub and Jira — not self-reported through an SDK call your team has to remember to make. An outcome you declare is an outcome you can flatter
AI-assisted work is compared against concurrent non-AI work and a pre-adoption baseline. The interface will not render a causal claim, and the caveat travels with the number into alerts and exports
Every linked figure carries its join confidence. Heuristic matches are excluded from headline numbers by default and labeled when included
Traffic outside the gateway is reported as a coverage figure, not quietly excluded. Every metric states whether it is measurable out of the box or requires an integration
Every metric ships with a Markdown document: formula, data lineage, inclusion rules, sample size and known limitations — circulate it and sign it off
Editing a metric produces a new version and a new series. Nothing silently rewrites a number you already reported
No rip and replace
The measurement layer does not require you to switch proxies. Drop the telemetry plugin into whatever you run today — any OpenAI-compatible gateway — keep your routing exactly as it is, and add enforcement later if you want it
Code, contracts and PII pasted into consumer LLM accounts, with no record it happened.
Teams on their own provider keys — invisible to both invoice review and DLP.
Recursion loops that surface on next month’s bill, after the budget is gone.
“Who sent what, to which model?” — a question you will eventually get in writing.
Once the numbers are trusted, the same gateway can route traffic to cheaper models, compress the context you never needed to send, screen every prompt, and hold every agent inside its cost and recursion limits. Measurement is what tells you a cheaper route did not quietly cost you quality.
Individual analytics
Per-person views exist because engineering leaders ask for them — and they ship switched off. Nothing on this page depends on turning them on
✓Off by default — org-level admin action required to enable
✓Small-group suppression — no cohort under five people is rendered
✓Audit-logged — every enablement and every view
✓Self-view transparency — people can see their own profile
Engineering-measurement tools die in procurement when works councils, unions or employee-representation bodies read them as monitoring. Shipping the safeguards on by default — and leading with them — is how the capability survives a European enterprise review.
Team and organization analysis never depends on individual views being enabled. Nothing in the metric packs above requires them.
Explore the Platform
Cost per merged PR.
Cost per resolved ticket
If it can be joined from your gateway and your systems of record, it can be a live metric before the end of the call
US Office:
1250 Borregas Avenue Sunnyvale, CA 94089, US
UK Office:
124 City Road, London, EC1V 2NX, United Kingdom
