Procurement has been told it is about to be transformed by AI roughly once a quarter for several years. Some of that is noise. Some of it is now working in production, and it is worth separating the two before committing budget.
Working now: classification and extraction
Spend classification was always a tedious, error-prone, rules-based job, and it is exactly the kind of pattern-matching that current models do well. The same is true of pulling key terms out of a contract portfolio nobody has read end to end. Both produce immediate, checkable value.
Working now: drafting and summarising
First-draft RFP questions, supplier communications, summaries of long submissions — these save real hours. The output needs review, but reviewing a draft is faster than producing one.
Not yet: judgement under accountability
Awarding business, deciding what a relationship is worth, judging whether a supplier is telling you the truth about capacity — these involve accountability that cannot currently be delegated to a system that cannot explain itself. Tools that support the judgement are useful; tools that claim to make it are not ready.
The prerequisite everyone skips
Every one of these depends on data that is complete enough to be worth analysing. Organisations that cannot say what they spent, with whom, on what, will not be rescued by a model. The unglamorous data work comes first, and it is usually where the delay actually is.