Step one
Assisted audit
4 to 6 weeks
- 30 to 90 days of gateway and provider logs
- Findings report ranked by recoverable dollars
- A fix pack for each finding, ready for your gateway
- A read-out with your platform and finance leads
A Crux Digital managed service
Much of every enterprise AI bill pays for work a smaller model, a script, or an overnight job would do as well. Fit Review reads the usage records you already keep, flags where AI was the wrong tool or a bigger model than needed, prices the fix, and measures what it saved.
Built for companies spending six or seven figures a month on AI APIs.
What the 2026 research says
From the sample report: a sales team's email tool, from the first signal to savings confirmed after the fix.
Sales' email tool sent 1.2M short rewrites a month to a premium AI model.
Polishing a short email doesn't need the most capable model. A smaller one does it for about 1/50 the price.
Recommended: one routing change the platform team can make centrally. The sales team's tool works exactly as before.
$102k a year saved, confirmed from the next two weeks of usage records.
The missing question
Gateways route traffic. FinOps tools allocate spend. Both are good at what they do, and Fit Review replaces neither. It sits on top and judges the work itself.
| Question | AI gateway | FinOps tool | Fit Review |
|---|---|---|---|
| How much did we spend, and on which model? | ● | ● | uses yours |
| Which team or key spent it? | ● | ● | uses yours |
| Should this work have used AI at all? | – | – | ● |
| Was it the right size of model, called the right way? | – | – | ● |
| What would the better approach cost? | – | – | ● |
| Did the fix actually save money? | – | – | ● |
How it works
Every day, Fit Review reads the usage records your AI systems already keep. It looks for patterns across thousands of requests rather than judging single ones, because waste shows up as a habit: the same job, tool, or team, day after day.
Millions of individual requests become a list you recognize: each app, agent, scheduled job, and team.
Each workload is checked against eight known waste patterns, with the evidence and how certain the finding is.
Your actual usage at current list prices, compared with what the better approach would cost.
Uncertain findings are held for a person to review. Nothing is shared until your team decides.
The pattern catalog
Each one is a plain business problem with a measurable fix. Underneath, each is a precise rule over your usage records.
Routine, rule-based work like tagging tickets or reformatting records is sent to an AI model thousands of times a day.
A simple script handles most of it; AI only sees the cases that need judgment.
Deterministic work on an LLM: thousands of near-identical prompts with structured input and short structured output
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.
Default to the right-sized model; keep the premium one for hard cases.
Oversized model: top-tier model on short prompts with short, simple outputs
Every question about a contract or report re-sends the entire document, and you pay for all of it every time.
Send the document once and reuse it, or send only the relevant pages.
Context stuffing: large, mostly repeated input on every call
Most automated tasks finish in a few steps, but some loop for hundreds, repeating the same work and running up the bill.
Put a step limit and a clear finish line on every agent.
Runaway agent loop: one task with 100+ calls, repeated tool calls, or rising context per turn
Large scheduled jobs use the instant, full-price service even though nobody is waiting on the results.
Run them through the overnight batch service at half the price.
Real-time bulk job: large volume sent in a burst, no user waiting on results
When a request fails, a system immediately retries it, often several times, and each failed attempt is billed.
Retry patiently and fix the underlying error.
Retry storm: same request repeated after errors or timeouts
Employees ask a general AI about internal systems and policies it was never connected to: costly, and often wrong.
Point those questions at the company's own search or system of record.
Lookup with no source: factual q&a with no retrieval or tool use
Many people and systems ask for the same summary or answer each day, and each one is generated and paid for from scratch.
Generate it once and share it.
Duplicate work: same or near-same prompt run by many people or jobs
Proof, not promises
The headline number is realized savings: what shipped fixes actually saved, measured from your logs after the ship date. We compare cost per unit of work before and after, so usage growth can't hide savings or fake them.
Privacy, trust, and tone
If people feel watched, they route around you and you lose visibility entirely. Every default is the least invasive one.
Token counts, timing, model, and call structure. Prompt text stays opt-in per deployment.
When an export carries message bodies, they're hashed on import and never stored.
Findings attach to apps and teams. Individual attribution stays with the platform team, only to route a fix.
No per-person scores. Nothing in the product ranks employees.
Findings stay with your team. Nothing reaches other teams or employees unless you choose to share it.
“This job could run as a batch,” never “you're wasting money.” Every output follows it.
Getting started
Crux Digital engineers run the first review on your own logs, with your team. After that we keep watching every week, and the fixes stay yours.
Step one
4 to 6 weeks
Step two
Ongoing, every week
Questions
Crux Digital, a Denver software and AI engineering firm. Crux engineers run your audit, check each finding before your team sees it, and help ship the fixes if you want a hand. The Fit Review platform is the tool we use to do it, and your team gets its own login.
No. It reads logs your gateway or provider already keeps, once a day. It recommends changes; your team decides whether and how to make them. Nothing about how traffic is routed changes to use it.
Not by default. Most patterns show up in token counts, timing, model, and call structure. When a log export includes message bodies, they are hashed on import (to match retries and duplicates) and dropped. Prompt classification is an opt-in for later, per deployment.
Today: LiteLLM spend logs, Portkey log exports, the OpenAI Usage API, and any gateway that can export one row per request as CSV or JSON Lines. If you don't have a gateway yet, setting one up can be part of onboarding.
It shouldn't, and we measure it. AI adoption is a guardrail metric: calls and active users should hold or rise while cost per unit of work falls. If total usage drops, the product is doing something wrong.
Nothing is acted on or shared without your team. Every finding comes with a confidence level and the evidence behind it, and low-confidence ones are flagged for a closer look. Your team decides what's accurate, what to fix, and who hears about it.