How to teach accounts payable clerks to bank corrections that improve an AI agent

Teach clerks that fixing a draft is half the job; the other half is saying why it was wrong. A banked correction records the cause (wrong source, missing supplier fact, misapplied rule), and the right answer with its source. The agent owner turns repeated causes into changes to sources, house rules or the brief, then re-certifies. No model is retrained.

What does banking a correction mean in accounts payable?

An accounts payable agent matches invoices to purchase orders and receipts, codes them, posts those within tolerance and drafts supplier queries. Clerks review what it holds. Every day they fix things: a wrong cost centre, a supplier matched to the wrong entity, a tolerance exception that should have been a query. If they just fix the item and move on, the agent makes the same mistake tomorrow, and the day after.

Banking turns a fix into an instruction for the agent's owner. It is a short, structured record attached to the corrected item. Pryme never updates model weights for a customer; improvement comes from changing what the agent reads and the bounds it works within. Our guide to AI agent skills by role places this skill at the centre of a reviewer's job, and how accounts payable clerks review an AI agent's decisions covers the review itself.

The AP team lead owns the practice. The agent owner owns what happens to banked corrections.

What are the steps?

1. Agent, depth 2 or 3: drafts or holds the item. A coded invoice, a supplier query, a tolerance exception.
2. Clerk: corrects the item. Changes the cost centre, the supplier, the amount or the wording.
3. Clerk: banks the correction. Picks one cause from a short list, adds a one-line reason, and attaches the source that shows the right answer: the purchase order, the contract, the supplier master record.
4. Agent, depth 1: groups banked corrections by cause and supplier each week.
5. Agent owner: reviews the groups. Decides the fix: a source correction, a new house rule, a brief change, or nothing because it was a one-off.
6. Builder: makes the change and adds the corrected cases to the agent's certification set, so every future run is tested against them.
7. Agent owner: re-certifies where the change touches depth 3 posting, then tells clerks what changed.

Which causes should AP clerks choose from?

Keep the list short so clerks use it. Six causes cover most AP corrections:

• Wrong source. The agent read the right place but the data there was wrong, such as an out-of-date supplier record. Fix: correct the source.
• Missing fact. The agent did not have something a clerk knew, such as a supplier's new legal entity name. Fix: add the fact to a source the agent reads.
• Rule misapplied. The agent had the facts but coded or matched wrongly, such as putting software subscriptions to hardware. Fix: brief or house rule.
• Should have held. The agent posted something that should have been an exception. Fix: tighten a bound.
• Held unnecessarily. The agent held something routine. Fix: review the tolerance with the process owner.
• Wording. The supplier query was unclear. Fix: template.

The clerk does not decide the fix. They decide the cause and show the evidence. That keeps the clerk's job quick and the owner's decision informed.

Two AP-specific points matter. First, invoices arrive by email and are external; a correction that says "the invoice shows a new bank account" is not a fix to bank, it is a fraud flag to raise. Second, coding corrections often cluster by supplier. A dozen corrections for one supplier usually means one bad master record.

Which controls keep banked corrections honest, and what evidence do they leave?

Cause and source required. A correction without a cause is a fix, not a banked correction. Evidence: corrections with cause and source attached.

Owner decision recorded. Each group gets a decision: changed, declined, or watching. Evidence: the owner's decision log.

Corrected cases join the test set. Evidence: the case references in the certification set.

Re-certification after changes to depth 3 behaviour. Evidence: certificate dates following each change.

Fraud flags kept separate. Evidence: payee-related corrections routed to the fraud or finance route, not the improvement route.

What should you measure?

• Corrections per hundred invoices, by cause, over time. It should fall where fixes landed.
• Repeat corrections after a fix, which mean the fix did not work.
• Corrections banked without a cause. Aim to keep this small.
• Time from banked group to owner decision.

Where does it usually go wrong?

Clerks stop banking under pressure. At month end they fix and move on. Make banking one click with a default reason list, and look at the gap afterwards.

The owner never closes the loop. Clerks bank for weeks and nothing changes. Tell them what changed every time, even small fixes.

Corrections are used to judge clerks. If banked corrections feed performance reviews, clerks bank fewer. Measure the agent with them, not the people.

Every correction becomes a house rule. The agent drowns in exceptions. Group first, then decide. Compare the approach in how to teach underwriters to bank corrections that improve an AI agent and how to teach paralegals to bank corrections that improve an AI agent, where fewer, weightier corrections are the norm.

Where the Academy fits

Teaching the agent is one of the eight modules in Pryme Intelligence Academy's Operator track, alongside human review and keeping knowledge current. Its practicum is two weeks of real review decisions, so clerks practise banking on live work. Market-specific training, such as AI agent training for accounts payable clerks in logistics in France, builds on the same habit.