Agentic procurement pilots meet the three-way match

Purchase-to-pay is the natural first home for agentic automation. The process is high volume, rule-bound and well instrumented. Several vendors are now demonstrating agents that raise a purchase order, chase the goods receipt, resolve the match exception and release payment with no human in the loop below a threshold.

The demonstrations are impressive. The control narratives accompanying them are not yet. Segregation of duties assumes that the person raising an order and the person approving payment are different. An agent performing both is a single actor, however many internal steps it takes, and describing it otherwise in a control matrix is an argument you will lose with an auditor.

None of this makes the technology unusable. It makes the design decisions load-bearing. The threshold below which no human approves is the single most important number in the deployment, and it should be set by the finance function before the pilot begins rather than negotiated afterwards when the efficiency numbers are already in a slide. Anything above it needs a named human, and the log needs to show them.

The other requirement is determinism in the record. It is not sufficient for the system to produce a narrative explanation of why it paid an invoice. The audit trail must be a factual record of inputs, rules applied and outputs, reproducible on demand. Several products currently offer the narrative and call it an audit trail. They are different things and the distinction will matter the first time a payment goes wrong.

Our view: run the pilot, set the threshold low, and raise it deliberately as the exception rate justifies. The teams getting this right treat the threshold as a dial they own, not a setting the vendor recommends.