Yes, for every release and block they approve. An AI agent can score transactions, hold suspicious payments and draft a case summary, but releasing a held payment is the analyst's call. A fraud the agent never flagged, because its sources or rules missed it, is not the analyst's alone. The record shows where each decision was made.
Why is the release decision the heart of fraud accountability?
Fraud teams make two kinds of mistake: letting fraud through and stopping genuine customers. Both cost money and trust. An agent helps by watching every transaction, holding the risky ones and gathering context: device, location, payee history, prior contact.
Releasing money is depth 4 work. In most firms the agent can hold a payment on its own but cannot release one it held; that waits for a named analyst. The analyst's release is therefore the moment the risk is accepted, and the moment accountability attaches.
What may the agent do, and what stays with the analyst?
- The agent may: score and hold transactions, pull device and payee history, draft a summary, and send a scripted verification message through a channel the customer already uses.
- The agent refuses: releasing a payment it held, accepting a customer's word as verification, and changing payee details.
- The analyst keeps: deciding whether a customer is being coached by a scammer, judging a story told on the phone, releasing or blocking, and deciding when to involve the customer's bank or the police.
Some actions are withheld at the platform floor for every agent at every depth, such as reversing a settled payment on Pryme's own rail. Those are not the analyst's to switch on, and not the agent's to attempt.
What does the record show about who decided?
The case record lists every signal the agent used, with provenance. The bank's own transaction history is sealed. A payee's name check result from a verified scheme is signed. What the customer says on the phone about who they are paying is declared. A message forwarded by the customer from the payee is external.
When an analyst releases a held payment, the record shows what was on screen, any notes they made, and the time between opening the case and releasing it. If the agent flagged that the payee was newly created and the customer's explanation was declared only, and the analyst released anyway, the decision is theirs.
An example: the convincing customer
Take a retail bank. A customer tries to send 900 to a new payee. The agent holds the payment: a new payee, an unusual time, a recent change to the customer's phone number. The customer calls in, explains it is a deposit for a holiday rental, and sounds calm. The analyst notes the call and releases the payment.
It was an authorised push payment scam. The customer had been coached. The record shows the three signals, the declared explanation and the release. The analyst is accountable for the release. The team is accountable for whether analysts were trained to treat a calm, coached customer as a known pattern.
What controls keep it fair?
- Declared explanations cannot release a high-risk hold alone. Ask for something signed or sealed.
- Time pressure is visible. The record of time-to-release lets leads spot rushed decisions.
- Analysts practise on real cases first, through a two-week practicum for fraud analysts recorded in a practicum log template for fraud analysts.
- Skills keep pace with patterns. Agent certificates lapse after 90 days and fraud typologies move faster; see how to prepare fraud analysts for 90-day recertification.
Onboarding teams face a related question earlier in the customer's life; see KYC analysts and AI agent mistakes. The wider topic is in AI literacy and certification for people who work with agents.
Where the Academy fits
Certified Pryme Operator, the Operator-track credential from Pryme Intelligence Academy, requires a written exam and two weeks of real review decisions with no missed red line. For a fraud team, releasing a high-risk hold on a declared story alone is the sort of red line the practicum exists to catch before a reviewer works unsupervised.
Questions people ask
Is the analyst accountable for fraud the agent never held?
Not on its own. If the signals were absent from the agent's sources, the gap belongs to the agent's design and its certification on the bank's cases.
Should analysts be measured on speed?
Partly, but never alone. Speed targets without quality checks push analysts to release on declared stories.
Can the agent block a card without a person?
Often yes, within bounds, because blocking is easier to undo than releasing. The bounds and their owner should still be named in the record.