Sample report · synthetic data
AI usage fit review
Generated by the Fit Review engine from 30 days of synthetic AI usage records for a fictional enterprise with about 2,485 people and services calling AI (Sep 12, 2026 to Oct 11, 2026): 22,764,322 calls across 12 workloads. Costs are at current list prices. The company and its data are made up; the analysis is not.
- Annual AI spend
- $7.1M
- Recoverable per year
- $2.2M
- Findings
- 7 + 3 held for review
- 01
Product's AI agent keeps working long after the task is done
Most automated tasks finish in a few steps, but some loop for hundreds, repeating the same work and running up the bill.
The fix: Put a step limit and a clear finish line on every agent.
Evidence from the usage records
- 2,964 of 23,400 tasks ran long; the longest took 259 calls
- Context grows by about 601 tokens per turn in long tasks
- 34% of calls in long tasks repeat an earlier request exactly
- Typical tasks finish in 13 calls
Most tasks finish quickly, but 2,964 ran long (up to 259 calls) repeating the same tool calls. A step cap plus a stop condition ends them.
$1.2M$147ksaves $1.1M a year - 02
Legal's tool re-sends the whole document with every question
Every question about a contract or report re-sends the entire document, and you pay for all of it every time.
The fix: Send the document once and reuse it, or send only the relevant pages.
Evidence from the usage records
- Average call sends 84,977 input tokens to get 501 back
- Only 0% of input tokens are served from cache
- 100% of calls share the same system prompt
Each call resends ~84977 input tokens, most of them the same. Put the stable part first and mark it cacheable, or retrieve only the relevant sections.
$590k$177ksaves $412k a year - 03
Data's automated job pays AI to do a spreadsheet's job
Routine, rule-based work like tagging tickets or reformatting records is sent to an AI model thousands of times a day.
The fix: A simple script handles most of it; AI only sees the cases that need judgment.
Evidence from the usage records
- 3,120,000 calls a month with the same prompt shape (100% share one template)
- Inputs vary by 5% around 2,400 tokens; outputs average 18 tokens
- No tool use and no one streaming the result
The job sends ~3120000 near-identical structured requests a month and gets short structured answers. A rules or lookup step handles most rows; send only the ambiguous 10% to claude-haiku-5-5.
$280k$992saves $279k a year - 04
Growth's automated job pays rush rates for overnight work
Large scheduled jobs use the instant, full-price service even though nobody is waiting on the results.
The fix: Run them through the overnight batch service at half the price.
Evidence from the usage records
- 100% of calls arrive in bursts, peaking at 130,000 calls in one hour
- 30 burst windows in the period; nothing is streamed back to a user
- Sent through the real-time API instead of the batch API
Up to 130000 calls an hour arrive in scheduled bursts and nobody waits on the result. The batch API costs half as much and finishes within 24 hours.
$304k$152ksaves $152k a year - 05
Operations' tool answers the same question hundreds of times a day
Many people and systems ask for the same summary or answer each day, and each one is generated and paid for from scratch.
The fix: Generate it once and share it.
Evidence from the usage records
- 99% of calls repeat a request already answered the same day
- The most repeated request is run separately by 40 people or jobs
- 175,396 of these calls could be served from a shared cache
99% of calls ask a question already answered that day, by up to 40 different people or jobs. Run it once and share the result.
$96k$1,297saves $95k a year - 06
Sales' tool uses a premium model for quick edits
Quick rewrites and yes/no checks go to a premium model when a smaller one gives the same answer for a fraction of the price.
The fix: Default to the right-sized model; keep the premium one for hard cases.
Evidence from the usage records
- 100% of calls go to claude-opus-4-8, a top-tier model
- Average call is 519 tokens in and 150 tokens out
- claude-haiku-5-5 handles the same token volume for about 1/50 the cost
Requests are short and outputs are simple. claude-haiku-5-5 handles this work; keep claude-opus-4-8 as an explicit opt-in for the hard cases.
$95k$1,900saves $93k a year - 07
Finance's automated job pays repeatedly for the same failures
When a request fails, a system immediately retries it, often several times, and each failed attempt is billed.
The fix: Retry patiently and fix the underlying error.
Evidence from the usage records
- 843,440 calls repeat a request that just failed, within 120s
- 41% of calls error; most common: timeout (561,600), rate_limit_429 (281,840)
- Longest retry chain: 7 attempts with no backoff
843,440 calls retried a request that had just failed, up to 7 times in a row. Exponential backoff stops paying for repeats; the upstream timeout is the real fix.
$175k$114ksaves $61k a year
Held for review
These can't be confirmed from usage records alone, so they wait for a person to check them before anyone else sees them. A wrong finding costs more trust than a missed one saves.
- Support staff ask a chatbot for facts it doesn't have (3 people, about $97 a year if confirmed)