← Use cases & case studies

Case study · AI cost · Customer zero · Ayurak, on its own platform

150 paid AI calls a night, down to a handful

Our own AI cost report showed the same spend every night at 04:30. Aribot traced it to one scheduled job, showed which calls were wasted, and after two small fixes the first two nightly runs made 4 and 7 model calls instead of about 150.

Download as PDF
~150 → 4–7

paid model calls in the nightly cost run, before the fix and on the first two nights after

$4.85 → $0.26

cost of that run on the last night before, and the higher of the first two nights after

2,254

earlier calls now shown as not metered, instead of being counted as free

The situation

AI spend was small enough that nobody looked closely, and regular enough that it never stopped. Every night between 04:30 and 04:32 the platform made around 150 calls to its most expensive model, costing four to five dollars a night.

The first suspect was wrong. An integrity check runs ten minutes earlier, so it got the blame. Its log shows it makes no AI calls at all.

What we built

Find the caller, not just the bill

The AI cost report breaks spend down by feature, by caller and by model. The 04:30 calls all came from the economic cost analysis that re-prices every architecture diagram each night.

See which calls were wasted

The job re-priced all 80 diagrams every night whether anything had changed or not. Six of them belonged to a company that no longer existed. For those, the job paid for an AI call first, then failed, and saved nothing. That had happened every night since February.

Fix the job, keep the result

The job now re-prices a diagram only when its components, rates or region change, or when the last analysis is a week old. It also checks that the company exists before spending anything.

Measure every provider

While tracing the spend, we found that calls to two providers were recorded with zero tokens. They are now metered, and the 2,254 earlier calls are shown as not metered instead of being treated as free.

Results

  • On the first two nights after the fix, the cost run made 4 and then 7 model calls, for $0.06 and $0.26. On the night before, it made 146 calls for $4.85.
  • No AI call is made for a diagram whose company no longer exists.
  • Every provider's calls are now metered with the tokens the provider reports.
  • These are the first two nights. We will report the monthly figure once there is a full month.

Related use cases

Customer zero: this is Ayurak running Aribot on its own platform. Figures come from our own AI usage ledger, counted on 2 October 2026. The 'after' figures are the first two nights; we will update them after a full month. Products: CXO Workbench & Economic Intelligence.