How to train the procurement team to review AI agent decisions

Procurement reviewers need to learn three decisions above all: when to approve a purchase order the agent raised, when to accept a supplier the agent has checked, and how to settle a match exception. Teach how the agent decides, calibrate on past cases, shadow two weeks of live work beside a certified buyer, then sign off with zero red-line misses.

What does the procurement team review?

A procurement agent can raise purchase orders from approved requisitions, check new suppliers against the business's onboarding requirements, compare invoices against orders and goods receipts, flag contracts due for renewal and draft supplier correspondence. Much of this runs at depth 2 or 3 within the business's limits.

What reaches the team: orders over a limit or outside the approved route, suppliers with missing or self-declared documents, three-way match exceptions, renewals with changed terms, and anything the agent refused. The head of procurement owns the programme. A senior buyer with a current certificate runs calibration and shadowing. Accounts payable owns payment; procurement reviewers approve commitments, not payments.

Why does procurement review need training of its own?

Procurement decisions commit money before anyone pays it, and they are hard to unwind once a supplier has delivered. They also sit next to conflicts of interest. A reviewer who approves a single-source order for a supplier they know must be able to show why, from the record, not from memory.

The other reason is document quality. Supplier evidence varies from signed certificates to scanned letters of uncertain origin. The agent stamps each fact: sealed if it came from the business's own systems, signed if a named party answers for it, declared if the supplier simply states it, external if it arrived from outside. Reviewers must learn which class is enough for which decision.

The steps

1. Person (head of procurement): lists decision types and red lines. Red lines might include: approving a supplier whose insurance or tax evidence is only declared; approving an order that splits a purchase to stay under a limit; approving any order where the reviewer has an undeclared interest.
2. Person (buyer): completes the learning on depth, the minimum chain of five bounds, verdicts and provenance.
3. Person (senior buyer): runs calibration on 50 anonymised past cases: split orders, renewals with quiet price rises, suppliers with expired certificates, match exceptions with partial deliveries.
4. Agent, depth 2 and 3: processes live requisitions and invoices in shadow mode. The trainee decides held items in a shadow queue; the senior buyer decides for real.
5. Agent, depth 1: logs both decisions and the deciding bound.
6. Person (senior buyer): debriefs mismatches twice a week, focusing on the reasons written.
7. Person (head of procurement): signs off after two weeks at the bar.
8. Person (workspace admin): grants approval rights by capability and value band.

What should calibration cover?

Split orders and quiet renewals, because they are where experienced buyers disagree. Show two orders from the same requester to the same supplier on consecutive days, each just under the limit. Ask whether to approve. Some will; the agent may have held the second because a business limit counts orders over a period. Discussing why turns a rule into judgement.

Then show a renewal where the agent's comparison shows a price rise buried in a revised schedule. Ask each buyer what they would do and what they would write as the reason. The best reasons name the clause, the change and the action, such as "Declined renewal at revised schedule rates; returned to supplier for price justification."

Which controls matter, and what evidence do they leave?

• Value bands tied to certificates. Evidence: the admin record of who may approve which band, with expiry dates.
• Conflict declarations at approval. Evidence: the declaration against each approved order.
• Provenance thresholds for onboarding. Evidence: the class of each supplier document.
• Independent shadow decisions. Evidence: trainee and senior decisions filed separately.
• Append-only record. Evidence: every verdict, so an audit can reconstruct any award.

What should you measure?

Agreement with the senior buyer by decision type and value band. Red-line misses, which must be zero. Orders approved and later cancelled. Suppliers onboarded and later found to have missing evidence. Time held items wait, which matters to requesters. And reasons quality, scored on a sample.

Where does it usually go wrong?

Procurement trains the buyers and forgets the requesters. Requesters then push buyers to approve held orders. Brief them on what a hold means.

Calibration avoids conflict-of-interest cases because they feel personal. They are the ones most worth discussing.

Approval bands are copied from the old manual process without asking whether the agent's evidence changes them. Revisit them once the team is certified.

Fraud and lending teams make similar approve-or-hold decisions under pressure; compare training the fraud team and training the lending team. The officer's routine is in how procurement officers review an agent's decisions, and readiness can be logged on the reviewer onboarding checklist for the procurement team. The guide to training teams for AI agents sets the overall order.