Turning our own AI usage into a report anyone can read

Challenge: we couldn't read our own AI bill

Horizon AI's team builds AI applications for several client projects at once, calling more than a dozen models day to day — GPT, Claude, Gemini, DeepSeek, and others. Before we consolidated, we faced the same problems as every enterprise adopting AI:

• Keys were scattered across each project's environment variables, and no one knew how many the company held.
• Each provider billed separately, so we only learned the month's spend at month-end.
• When a client asked what their project's AI cost that month, we had to piece the answer together from three dashboards.

Solution: one governance plane for the whole company

We migrated every model call to our own ATP Petrichor platform and applied the Organization → Workspace → Project hierarchy:

1. One Project per client project, each with its own key and its own budget.
2. Model access is controlled at the project level — the system enables only the models a client's contract allows.
3. Every request is logged: model, token count, status, and the project it belongs to, all queryable in real time.

The migration was the same process we run for client adoption: inventory, consolidate, set budgets, and put it on a dashboard — done in two weeks.

Results: from month-end surprise to a daily dashboard

• Each client project's AI cost is visible in real time, so we can calculate margin.
• During a client's security audit, the logs are the evidence — no documents to reconstruct after the fact.
• New models are available the moment they are added, and engineers no longer request their own keys. Governance is not a constraint; it is what lets everyone work with confidence.

This case study is also our product validation: every practice we recommend to clients, we run on ourselves first. The scattered keys, runaway bills, and last-minute audit paperwork your business faces — we have been through all of it, and we have solved all of it.