← all work.

AI audit skill / in production

AI Rating Engine

An AI freight audit that rates every shipment again against the client's carrier contracts, finds what was overbilled and delivers one workbook auditors use to file claims. Nothing is estimated and nothing is made up.

role.

built and own it

status.

in production

type.

AI audit skill

stack.

claude, python, mcp, smc3 rateware, pc*miler, xlsxwriter, sql

Mockup of an audit workbook: shipments rated against the contract, overcharges highlighted, with an AI audit note
illustrative mockup with sample data. no client data.
195numbered rules the audit follows
15checks that must pass before any rating starts
31checks a separate verifier runs on the finished workbook
100%of payment rows land in exactly one bucket and the totals are checked

the problem.

Carrier contracts are long PDFs, scanned amendments, rate cards in a dozen layouts, fuel tables and emailed quotes. Checking thousands of LTL, TL, air and courier bills against all of that by hand is slow and easy to get wrong in ways that turn into bad claims.

what it does.

You give it a client's freight payments and their contract folder. It reads every document (including scanned PDFs, files embedded inside spreadsheets and emails), matches each carrier to its contract and rates every shipment with the rate in effect on its ship date.

LTL is rated on the SMC3 CzarLite tariff the contract names, plus the contract discount and minimum. Truckload uses the contract's lane rates with PC*MILER miles, air uses the forwarder's rate grid and fuel is checked against the weekly DOE diesel price and the contract's fuel table.

Every row ends up in one bucket: claim, review, excluded, unauditable or OK, with an error type and a plain note for the auditor. A carrier only becomes claimable when most of its bills reconcile to the penny, so one wrong assumption can't flood the claims list.

why it can be trusted.

It never assumes. If a rate, date or tariff isn't in a document, it asks or marks the shipment unauditable. No averages, no nearby lanes, no guesses.

A verifier that's independent of the builder reads the delivered workbook back and works out every rated row again from the source files. On one recent run that meant 6,427 payments worth $18.4M, with every check passing. What it learns about a client is written to that client's rule file, so the next run starts smarter.

how it works.

  1. read every contract, rate card and fuel table
  2. run the pre flight checks and ask what's missing
  3. rate each shipment on its ship date
  4. sort every row into one bucket with a note
  5. verify the workbook independently, then deliver