The escalation clause was in the contract. Nobody was claiming it.

A Tier-1 component maker supplying three vehicle manufacturers, with escalation clauses in every long-term agreement tied to published raw material indices. The clauses were negotiated carefully. They were then largely unused, because claiming under them required assembling part-level consumption, index movement and contractual eligibility inside a window that closed before the commercial team could get to it.

Nobody in the business had ever computed the gap between what was claimed and what was claimable. When we did, in discovery, the figure was large enough that the finance director asked us to check it twice.

The definition

Escalation value claimed against escalation value entitled, over trailing four quarters, computed from cost sheets, bills of material, published index movements and the specific eligibility terms of each agreement. Excluded: parts under negotiation, tooling recoveries, and anything where the agreement was silent.

The exclusions took a week and two meetings, mostly about whether volume-linked eligibility thresholds should be assessed per part or per programme. The answer was per programme, and it changed the number by more than any other decision in the model.

The build

Document extraction across agreements and cost sheets, resolved to the part and programme level, joined to index feeds, producing a claim pack that a commercial manager reviews and submits. Nothing is submitted automatically. Everything is assembled automatically.

Claim preparation went from about nine working days to one. The share of entitlement actually claimed rose from thirty-one percent to seventy-eight over three quarters, and the remaining gap is now visible and argued about rather than invisible and absorbed.

What did not work

Our first extraction of the older agreements, several of which existed only as scans of signed amendments, scored well below its gate and did not ship. Those agreements were re-keyed by hand by the client's own team over two weeks. The system reads what it can read reliably and says so when it cannot, which is less impressive in a demo and considerably more useful in a claim.