Train for the decisions the agent hands over: repairing failed payments, handling returns, releasing payments held over a limit, deciding beneficiary changes. The programme runs learning, then calibration on past exceptions, then two weeks of shadow review beside a certified analyst, then sign-off with zero red-line misses. Approval rights follow the certificate, depth by depth.
What does the payments operations team review?
A payments agent can work at depth 4, where it moves money, but only for capabilities certified on the business's own cases. Everything else reaches a person. In a typical payments team that means repair queues (a payment failed on a format or account error), returns and recalls, duplicate payment checks, held payments over the business's limit, beneficiary changes and payee name mismatches.
The head of payments operations owns the programme. A senior analyst with a current certificate runs calibration and shadowing. The financial crime team owns screening decisions; payments reviewers route screening hits, they do not clear them. That boundary belongs in the red lines from day one.
Why does payments review need its own training?
Speed and irreversibility. Many payments cannot be recalled once settled, and some reversals are withheld at the platform floor for every agent at every depth. A reviewer who approves the wrong release cannot undo it with a second click.
Payments review also turns on provenance. The ledger and the settlement rail are sealed sources; an emailed instruction is external. A reviewer must see at a glance which facts are strong enough to support a release, and treat any external claim as unable to authorise a payment on its own.
The steps
- Person (head of payments operations): lists decision types and red lines. Red lines might include: releasing a payment with an unresolved payee name mismatch; approving a beneficiary change from an external source without call-back; clearing a screening hit.
- Person (analyst): completes the learning: depth by capability, the minimum chain of five bounds, the four verdicts, provenance classes and writing a decision reason.
- Person (senior analyst): runs calibration on 60 anonymised past exceptions over three sessions, weighted towards the rare but costly ones: name mismatches, duplicates that are not duplicates, recalls after settlement.
- Agent, depth 4 for certified capabilities, held otherwise: processes live work in shadow mode. The trainee decides each held item in a shadow queue; the senior analyst decides for real.
- Agent, depth 1: logs both decisions, the bound that decided the verdict and any red-line flag.
- Person (senior analyst): debriefs daily on mismatches, reading the trainee's written reasons aloud.
- Person (head of payments operations): signs off after at least two weeks of shadow decisions at the bar.
- Person (workspace admin): grants approval rights by capability, starting with repairs and returns, adding releases over limit after a further period.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Routine payment to a verified beneficiary | Executes within limits | 4 | Analyst samples |
| Payment failed on a format error | Proposes the repair with its source | 2 | Analyst approves the repair |
| Payment over the business's limit | Prepares and holds | held | Analyst with rights releases |
| Payee name mismatch | Holds with both names shown | held | Analyst and second approver decide |
| Screening hit | Routes to financial crime | held | Financial crime decides |
| Reversal of a settled payment on the platform's own rail | Withholds | withheld | Nobody; analyst escalates through the proper channel |
What should calibration focus on?
Disagreement. Pick cases where experienced analysts have historically split: a duplicate that turns out to be two genuine invoices of the same amount; a recall request from a customer who has been defrauded; a return whose reason code does not match the story. Ask each person to decide alone and write one line of reasoning, then compare.
Make the bound visible. For every held item, ask the room which bound decided it: role scope, certificate, sources, the business's limits or the platform guardrails. An analyst who knows that a payment was held by the business's own limit, not by a source problem, will look for the right thing.
Which controls matter, and what evidence do they leave?
- Separation from screening. Evidence: every screening hit routed to financial crime, with their decision.
- Second approval on name mismatches. Evidence: two names on each decision.
- Independent shadow decisions. Evidence: trainee and senior decisions filed before comparison.
- Sign-off against a written bar. Evidence: correct decisions out of the total, red-line count, signature.
- Approval rights tied to certificates. Evidence: the admin view of who may release what, and expiry dates.
What should you measure?
Agreement with the senior analyst by decision type. Red-line misses, which must be zero. Time to decide held payments, especially near cut-off. Repairs that failed a second time. And releases approved without opening the facts panel, which the workspace record makes visible.
Where does it usually go wrong?
Calibration uses only clean cases because they are easy to anonymise. The team then meets its first real recall dispute live.
Shadowing is cut short when volumes rise. That is exactly when reviewers learn most.
The agent's certificate lapses or a run that clears no depth drops it to depth 1, and routine payments arrive as held items overnight. Reviewers who have never practised a surge freeze. Rehearse it.
The general operations version of this programme is training the operations team to review agent decisions; legal teams adapt it in training the legal operations team. The reviewer's daily routine is set out in how payments operations staff review an agent's decisions, and readiness can be recorded on the reviewer onboarding checklist for payments operations. Sequencing across the business is in the guide to training teams for AI agents.
Questions people ask
Can a reviewer approve a payment the agent refused?
No. A refusal means the action fell outside the agent's bounds or the business's limits. If the refusal looks wrong, the reviewer raises it so the limits or brief can be reviewed by the people who own them.
Should every analyst be trained on releases over limit?
Not at first. Train everyone on repairs and returns, then extend release rights to analysts who have passed sign-off on those. Fewer, better-trained approvers make a stronger control.
How does this relate to anti-money laundering training?
It sits beside it. FATF-based AML training covers suspicion and reporting; this programme covers how to decide an agent's held work. Both are needed, and neither replaces the other.