Assess the payments operations team that will review the first agent, whether it handles exceptions, disputes or merchant onboarding. The test, set in payments, probes depth 4, withheld actions, external instructions and the record. Read the report by concept: it decides whether the first agent investigates and drafts, or ever releases anything, and which gap to train first.
Why does a payment company need this?
In a payment company almost every agent capability sits close to money. Repairing a failed payment, releasing a held one, refunding a merchant's customer or raising a chargeback all move funds, which is depth 4. Some actions, such as reversing a settled payment on the company's own rail, are withheld at the platform floor for every agent at every depth.
Staff who review an agent's held actions must therefore understand more than most: which actions can be allowed, which are held, which are refused by the company's own limits, and which are withheld by design. The head of payments operations owns the assessment, with the financial crime lead. For the concept, see what a readiness assessment is.
Who should take it?
One function per sitting, up to 25 people:
- Payment exceptions and repairs: analysts who fix failed or rejected payments.
- Disputes and chargebacks: analysts who gather evidence and decide representments.
- Merchant onboarding: staff who check businesses before they can take payments.
Include the team leader with the highest approval limit and one person from financial crime as an observer. Who should go first is covered in who in a payment company needs AI agent training first.
The steps
- Person (head of payments operations): chooses the function and lists actions that stay with a person whatever the agent's certificate.
- Person (head of payments operations): asks for the test set in payments.
- Person (each member): sits the test alone.
- Person (head of payments operations and financial crime lead): read the report by concept.
- Person (head of payments operations): chooses the first agent and its depth. Most start with an agent that investigates exceptions and drafts the repair at depth 2.
- Person (team leader): trains the weakest concept using the company's own payment types.
- Agent, depth 1: reads past exceptions in shadow and proposes the repair and verdict it would have given.
- Person (cohort): reassesses before the agent drafts on live payments.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Choose function | Nothing yet | None | Never-release list |
| Sit the test | Nothing yet | None | Individual answers |
| Read the report | Nothing yet | None | Gaps, with financial crime |
| First agent | Named as candidate | None | Agent and depth |
| Train the gap | Runs in practice labs | Practice | Learning |
| Shadow | Proposes repairs on past exceptions | 1 | Comparing with real repairs |
| Reassess | Nothing new | 1 then 2 | Go-live |
What does the test usually find in payment companies?
Withheld is not a setting. Many staff assume that with enough evidence the company could let an agent reverse a settled payment. The test makes clear that withheld actions are a platform decision, the same for every agent at every depth. The company narrows what agents do through its own limits; it does not widen past the floor.
External instructions. A merchant emails asking to change its settlement account. A customer sends a screenshot of a transfer. These are external claims. The settlement rail is sealed; a verified counterparty's message is signed. An external claim cannot authorise a payment, however urgent it sounds. Analysts used to working fast through a queue often score lower here than they expect.
Name checks. In the EU, the Instant Payments Regulation requires payment service providers to verify the payee's name. Staff should know how a name mismatch appears in the agent's verdict and that it is held for a person. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What does the report change?
It decides how close to money the first agent gets. A team weak on external instructions should start with an agent that investigates and drafts, never one that releases.
It sets approval rights. If the team leader with the highest limit scores lowest on withheld actions, that person trains before approving any agent draft. Some companies adopt a rule that approval at depth 4 requires a current reviewer certificate.
It gives the financial crime lead a view of where staff might wave through a fraudulent instruction, before any agent is involved.
What evidence should the company keep?
The never-release list.
The report by concept.
The first-agent decision and its depth.
The reassessment before live drafting.
What should the head of payments operations measure?
- Scores by concept, especially withheld actions and provenance.
- During shadow, repairs the agent proposed that a person would have rejected, and why.
- After go-live, held actions approved on external evidence alone, found by sampling.
Where does it go wrong in payment companies?
Starting with releases because the queue is long. The queue is long because the work is risky.
Leaving out disputes. Chargeback analysts deal with evidence of mixed provenance all day and are often the best guide to how the agent should weigh it.
Treating the assessment as compliance paperwork. The result should change the first agent. If it never does, it is not being read.
Banks face similar questions with larger, more separated teams; see the bank version. Logistics companies meet the redirect version of instruction fraud; compare the logistics company version. The readiness assessment report template gives a format, and the topic guide is training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy offers a free readiness assessment for teams of up to 25, with the Foundations test set in the team's own industry. Its Foundations material covers the decision chain and depth, including why some actions are withheld for every agent. Analysts who approve held payments can go on to the Operator track.
Questions people ask
Can an agent ever release a held payment?
At depth 4, within the company's own limits and only for capabilities it is certified for. Many payment companies keep releases with a person for a long time.
Should engineers who integrate rails take the test?
Yes, in their own sitting. They need to understand the decision chain to connect the agent without creating misleading refusals.
How often should a payment company reassess?
Before each new capability, before any rise in depth, and after significant changes to payment schemes or rails.