a manufacturing company with thousands of skus and a global supplier network. the erp does what erps do — fields, tables, exports, the occasional integration. the buyers do what buyers do — open spreadsheets, prioritize by feel, escalate the ones that scare them. on any given morning, the volume of decisions waiting for a human is larger than the human can hold. buy wrong, the line stops.
competitors can buy the same model. they cannot buy the twelve months your team spent shaping it.
the obvious move is to point a model at the erp and let it "use ai for procurement." it does not work. the model arrives at the same data that exhausts the buyer — fragmented across tables, partially defined, missing the operational context the team carries in its head. it produces fluent answers that are wrong in ways nobody can audit.
what worked was the inverse. before the model, an ai-ready decision layer above the erp. not a replacement — a layer. the erp stays where it lives. the buyer stays where they sit. between them, a structured surface that turns fragmented context into prioritized action, with the rules of the business written down as rules instead of remembered as habits.
- 01 structure every source — erp, supplier files, demand signals, historical exceptions — pulled into a single canonical shape. not "a data warehouse." a representation built so an ai can reason on it without translating first.
- 02 codify the deterministic logic the buyers already used — risk thresholds, lead-time tolerances, supplier blacklists — written down as code instead of carried as judgment. what was tribal became testable.
- 03 prioritize every morning, the system orders the day's decisions by what actually matters. on a representative monday: 53 alerts in the critical bucket out of thousands of skus. the buyer opens 53, not 4,000.
- 04 assist only at the end, the model. grounded in the canonical layer, the codified rules, the historical exceptions. when the buyer wants context, it explains. when they want a recommendation, it gives one — with the data and the rule behind it cited inline.
any competitor can license the same erp. any competitor can subscribe to the same model. what they cannot buy is the twelve months it took to make this company's decisions legible to a machine. the layer captured the exceptions the team didn't realize it was making — every override, every refusal, every escalation became part of how the model thinks about this business.
this story is procurement. the same shape — structure, codify, prioritize, then assist — is what we keep building at ai compass across customer operations, hr, and commercial. the vertical changes. the move doesn't.
this is what a compound bet looks like in the trenches. it does not move a number in q1. it moves three numbers in year two — margin, service levels, response time — and by year three it is the reason customers stay. priced like a pilot, it looks expensive. priced like infrastructure, it looks obvious. the framing decides the funding.
the model is everyone's. the layer above it is yours. compound bets are financed in years, not quarters.