← All work Next caseA staffing plan that stopped planning endless lines →
An AI bill cut by $4,050 a month
$48.5Ka year, saved
- Scale
- Every guest comment, sorted into themes
- Methods
- Spend traced step by step, retry limits, incremental processing
The problem
The output looked right. The bill did not. A pipeline read every guest comment and sorted it into themes. Nothing errored.
What I did
I traced the spend step by step and found three bugs, none in the model: retries that never stopped, old records that looked new so they ran again, and no "only do the new stuff" logic, so every run redid everything.
FIG.02 ยท a month of waste; flip the fixestry it · sample data
What changed
- All three bugs fixed.
- Same output as before.
- Spend is tracked at every step now, so it can't creep back.
What this looks like for you
If you run AI at volume, the model is rarely the expensive part. The plumbing is.
Let's talk data →