Example report · example, made-up company
AI cost audit example report
This is a complete AI cost audit for a made-up company, built from made-up usage data with the same method a real audit uses. It is not a customer. Provider and model names are replaced with letters.
Last updated:
1. Summary
Example Co. spends $6,400 a month on AI APIs across 2 providers and 4 features. I found 5 changes. Together they are worth an estimated $1,395 to $2,885 a month.
Start with number 1. It is the largest and the easiest to test. Numbers 3, 4 and 5 are small code changes with low risk.
2. Where the money goes
| Provider | Per month | Share |
|---|---|---|
| Provider A | $4,800 | 75% |
| Provider B | $1,600 | 25% |
| Feature (from request tags) | Per month | Share |
|---|---|---|
| Support chat assistant | $2,600 | 41% |
| Ticket tagging | $1,500 | 23% |
| Document summaries | $1,300 | 20% |
| Internal search | $600 | 9% |
| Untagged | $400 | 6% |
6% of spend had no feature tag. Tagging it would make the next review sharper.
3. Ranked changes
| # | Change | Estimated per month | Effort |
|---|---|---|---|
| 1 | Move ticket tagging to a smaller model Ticket tagging, Provider A | $640 to $1,275 | Low |
| 2 | Trim boilerplate from summary inputs Document summaries, Provider A | $260 to $520 | Medium |
| 3 | Re-embed only changed documents Internal search, Provider B | $200 to $400 | Low |
| 4 | Cache the repeated system prompt Support chat assistant, Provider A | $195 to $390 | Low |
| 5 | Cap retries on timeouts All features, Provider A | $100 to $300 | Low |
Total: $1,395 to $2,885 a month (estimated).
Want this for your own AI spend?
Request my audit4. Method notes
Move ticket tagging to a smaller model
Tagging costs $1,500 a month on a frontier model. A smaller model costs about 15% as much for short classification tasks, so a full switch would cut about $1,275. The low end assumes only half the volume can move. How to check: Run 200 labeled tickets through both models and compare tags.
Trim boilerplate from summary inputs
In a sample of 50 calls, 20% to 40% of each input was headers, signatures and repeated legal text. Input is most of this feature's cost, so the range follows that share of $1,300. How to check: Compare 30 summaries before and after trimming.
Re-embed only changed documents
The index re-embeds every document each night. About two thirds of documents did not change between runs. The low end allows for documents that change often. How to check: Confirm search results match on 20 common queries.
Cache the repeated system prompt
The same long system prompt goes out with every chat turn and makes up about 30% of this feature's cost. Cached input is billed at about half price on this provider. How to check: None expected. Output does not change.
Cap retries on timeouts
From March 14 to 16, timeouts triggered up to 6 retries per call and added about $300. Capping retries at 2 stops a repeat. The low end assumes spikes like this happen every few months. How to check: Watch error rates for a week after the change.
5. What this audit did not check
Two months of data only, so seasonal swings may be missed. Quality was not tested; each change lists the test to run first. Prices are the providers' list prices on the date of the report.
6. Next
A 5 to 8 minute video walks through this report. Send questions in writing within 14 days and I answer them in one written round.