AI Services · Measure
Know what your AI costs, what it returns and where the bill can drop.
Independent numbers from a firm with no dashboard to sell, starting with a 1–2 week teardown that ends in a costed savings plan.
The tools are in. The numbers aren't.
You paid for coding assistants, model APIs and agents. Now the board asks what they returned, finance asks why the token bill keeps climbing, and the honest answer is that nobody knows.
Only 31% of engineering leaders measure AI productivity gains, and 68% call the lack of clear metrics their biggest challenge (LeadDev, 2026; LeadDev). Gut feel isn’t enough: in a controlled study, experienced developers felt 20% faster with AI and measured 19% slower (METR, 2025).
What you get
AI Cost Teardown
1–2 weeks · fixed fee
Read-only scripts run inside your account against your model-provider admin data or Amazon Bedrock costs. We then design the changes that cut the spend: prompt caching, cheaper models where they do the job, batch processing, agent loop limits and budget caps. You get a savings plan with every change costed.
AI Engineering Baseline
3–4 weeks · fixed fee · re-run quarterly under managed measurement
A before-and-after measure of what AI is changing in your engineering delivery, comparable every quarter. Collectors you run read your git, pull request and CI history. They also read AI-tool admin data (GitHub Copilot, Claude Code, Cursor) and token spend, plus a short developer survey. The report tracks a fixed set of metrics, including lead time, change-failure rate, review latency and model cost per merged pull request.
Observability install
6–8 weeks · fixed fee
We turn “is the AI any good?” into a nightly score. Your experts define a good answer; we build a test set of approved answers and an AI judge checked against human scores. Around it sits one governed gateway with a key and budget per team. It adds cost attribution by team and use case, tracing, dashboards and alerts, all in your account.
Managed measurement
capped monthly retainer
A quarterly re-baseline, rubric upkeep as models and processes change, and regular cost reviews, so the numbers stay honest after we leave.
Numbers with nothing to sell
Neutral by design
We sell no dashboard and no seat license, so nothing in our numbers flatters a product. Productivity dashboards tend to grade their own tool; we measure across all of them.
Quality and cost, not seat counts
Usage dashboards tell you who logged in. We measure what changed in delivery, what each result cost and whether the output was any good.
A judge you can trust
Every install checks the AI judge against your experts’ scores before it grades your output. Every score carries the rubric and judge version that produced it.
Our meters caught what our cost reporting missed
44%
under-count in AI cost reporting, caught because two independent meters check each other
1 commit
per-session cost tracking traced a sudden cost jump to a single code change
$0.09–0.14
cost per fix our AI fix agent settled at, tracked session by session
Part of the Governed AI Platform
Observability and governance share one gateway. If you’re adopting AI at scale, the Governed AI Platform installs both together, with a single discovery for both.
Questions finance and engineering leaders ask
Do you need access to our code or data?
Not for the Teardown or Baseline: the scripts run inside your environment under your credentials, and you send us only the aggregate results. For an install, you approve in writing exactly what our engineers can access, and your data is not used to train AI models.
Which tools can you measure?
GitHub history, GitHub Copilot, Claude Code, Cursor, and Anthropic, OpenAI and Amazon Bedrock spend, plus any model traffic that runs through the gateway. If your stack differs, we build the connector as part of the job.
How fast will we see savings?
The Teardown ends with a costed savings plan in 1–2 weeks. Many changes, such as caching and model routing, can be applied the same month.
Will budgets cut off our engineers mid-task?
For tools that route through the gateway, you choose what happens at each limit: an alert or a stop. We recommend an alert first for interactive coding tools and a hard stop for automated agents.
Do you score individual developers?
Only if you want us to. Free assessments report at team level. On paid work, individual scores can target coaching and training at the areas each engineer is working on, and you decide who sees them before we start. The developer survey is always anonymous.
What does it cost to re-run the Baseline?
Quarterly re-baselines are part of the managed measurement retainer, so every quarter is measured with the same instrument.
Free ways to start
Run our AI Spend Snapshot against your Anthropic or OpenAI admin data or your Bedrock costs. It shows spend by key and model, your cache-hit ratio and which calls could run on a cheaper model. Or take the AI Engineering Scorecard: 15 questions and a gap report.
Find out what your AI is really costing you.
Ask for the free AI Spend Snapshot, or name your models and tools and we’ll scope a Cost Teardown and the savings levers that apply.