Contract & terms extraction
Contracts are where the money terms live: extract rate cards, payment terms, and escalators into a reference sheet — then let invoices be checked against promises.
Invoices tell you what a vendor charged. Contracts tell you what they were allowed to charge. At most companies these live in different worlds — invoices in the AP system, contracts in a legal folder nobody opens after signature — which is why rate creep, missed discounts, and expired-but-still-billing agreements quietly leak money. Document intelligence closes the loop: extract the commercial terms once, structurally, and every invoice becomes checkable against its promise.
What to pull (the commercial skeleton, not the whole contract)
- Rate cards — Alder's carrier contracts: rate per mile by lane, fuel-surcharge formula and index, accessorial fees (detention, liftgate, redelivery). This is the reference data the rate audit runs on.
- Payment terms & discounts — net days, early-payment discounts (a 2/10 net 30 unclaimed is a ~37% annualized return, forfeited), volume rebate thresholds and their measurement windows.
- Escalators & expirations — annual increase clauses (capped at what index?), renewal dates, notice windows. An expired contract with continued billing is both a compliance gap and a negotiation opportunity — extraction makes them visible on a calendar instead of discovered in year three.
- Approval-relevant clauses — liability caps, termination rights, exclusivity. Not finance's daily business, but flagging them into the summary keeps legal's attention where it belongs.
The extraction pattern is Module 2's same contract — named fields, page/section citation required for every extracted term ('clause 4.2, page 7' — so a human can verify in seconds), UNREADABLE/AMBIGUOUS escapes, no interpretation. Contract language is where models are most tempted to summarize helpfully; the citation requirement is what keeps extraction honest, because an invented term can't cite a real clause. A human verifies every extracted term against its citation once, at extraction time — then the reference sheet is trusted data with an audit trail, reusable thousands of times.
Priya's rate audit (Module 1's map) was sampling 5% of carrier invoices against contracts, manually, quarterly. With rate cards extracted into a reference sheet, the check runs on every invoice at match time: billed rate vs. contracted rate for that lane, fuel surcharge vs. the formula at that week's index, accessorials vs. the agreed schedule. The audit becomes a continuous detective control — and in freight, where misbillings run 1-3% of spend, it's typically the single fastest payback in the whole finance-AI portfolio.