‘AI for procurement’ covers a lot of ground. Some of it is real; some is a rebrand of features that existed before the term became fashionable. Worth separating the two.
Where AI is already useful
- Spend classification. Mapping a messy general-ledger transaction list to a clean category taxonomy is exactly the kind of fuzzy, pattern-matching task ML handles well. The category accuracy on first pass is now typically above 90%, which beats the manual baseline and beats rules-based systems that have to be rewritten every time the ERP changes.
- Anomaly detection in supplier behavior. Bid patterns, lead-time slippage, quality incidents — surface them before they show up as problems.
- Negotiation strategy suggestions. Pattern-recognition across historical events: ‘in the last three auctions for this category, the lowest bidder was set within the first 12 minutes. Consider shortening the event window.’
Where it’s still mostly marketing
- ‘AI-negotiated’ contracts where the model writes back to the supplier. These exist in narrow, tactical categories. They do not work for anything strategic.
- Predictive should-cost. Genuine should-cost modeling is engineering work — material specs, process flow, labor rates. A model trained on past prices is just predicting past prices.
The honest framing
The valuable AI in procurement looks like a co-pilot, not an autopilot. It surfaces patterns, drafts options, flags anomalies — and leaves the negotiating to the negotiator. Anyone selling the autopilot version either has a very narrow use case in mind, or isn’t being precise.