Payment investigations are one of the best places for a bank to start with agents. Tracing a missing payment, repairing one that failed, answering another bank's query and pursuing a recall are rule-bound, document-heavy and mostly made of reading and writing. The few steps that move money are easy to isolate and hold. That combination is rare, and it makes investigations a strong first job.
What does payment investigation work involve?
Every bank runs a team, often called payment investigations or payment exceptions, that handles what goes wrong after a payment is sent or received. The work arrives as cases.
- Tracing. A customer says a payment has not arrived. The team follows it through the bank's systems and, for cross-border payments, through correspondent banks, to find where it is.
- Repair. A payment failed or was held because information was missing or malformed. The team finds the missing detail and resubmits.
- Enquiries from other banks. Another bank asks for information about a payment, or queries a beneficiary name or purpose.
- Returns and recalls. A payment needs to come back, because it was sent in error or to the wrong account. A payment recall is a request to the receiving bank to return funds, which that bank may or may not be able to meet.
- Charges claims. Fees deducted along the way are queried by the customer or another bank.
Most of these cases follow a known pattern: identify the payment, gather its history, determine what happened, communicate with the right party in the right format, and close the case.
Why does this suit an early agent?
A first agent in a bank needs to deliver visible value while keeping risk low and building the controls the bank will need later. Investigations meet each test.
| Test for a first agent | How payment investigations measure up |
|---|---|
| Is the work mostly reading and drafting? | Yes: most steps are looking up records and writing messages |
| Are the rules clear? | Largely: payment schemes and message formats are well defined |
| Is there a backlog with measurable outcomes? | Usually: case volumes, ageing and resolution times are already tracked |
| Can the irreversible steps be separated? | Yes: crediting a customer or returning funds is a distinct, holdable action |
| Do customers notice the improvement? | Yes: "where is my payment?" is a common and stressful question |
| Does it build reusable capability? | Yes: provenance, holds and records built here serve every later agent |
The question of which AI agent a bank should deploy first has several reasonable answers. Investigations are often among the strongest because so little of the work is irreversible.
What can the agent do, and at what depth?
The split between agent and person follows the money.
- Depth 1: read the payment's full history across internal systems, the customer's account and any previous cases.
- Depth 2: draft replies to customers that explain what happened and what is being done; draft decisions on charges claims.
- Depth 3: send standard tracing and enquiry messages to other banks; repair and resubmit payments where the missing information comes from a strong source the bank already holds; update case status and close resolved cases.
- Stays with a person: crediting a customer's account for a lost or delayed payment, returning funds to a sender, accepting or refusing a recall request received from another bank, and any case where fraud is suspected.
Refunding charges is a borderline case. Small refunds that clearly fit policy may suit depth 3, as discussed in whether an AI agent can refund bank charges. Larger or disputed amounts stay with a person.
Where are the traps?
Investigations look safe, and mostly are, but three points need care.
Repairs that change the beneficiary. Filling in a missing reference is a repair. Changing the beneficiary account or name is not; it redirects money. Any repair that touches the destination of funds should be held, and the new details must not come from an external source.
Recalls on instant payments. On instant rails, funds reach the beneficiary within seconds and recall depends entirely on the receiving bank and the beneficiary. An agent should never tell a customer their money will be returned. It can say a request has been made and explain what happens next.
Fraud hiding in ordinary cases. A "where is my payment?" case can be the first sign of a scam: the customer paid someone who promised goods and they never came. An agent should recognise those patterns and hand the case to the fraud team rather than treating it as a tracing exercise.
What should the bank measure?
Measure what the investigations team already measures, before and after: case volumes, time to first response, time to resolution, ageing backlog and reopen rates. Add two agent-specific measures: the share of cases the agent closes without a person, and the rate at which people overturn the agent's drafts. A falling backlog with a rising overturn rate is a warning, not a success.
Where Pryme fits
In Pryme, investigation capabilities each carry their own depth, earned through Attest on cases from your own investigations queue. The minimum chain holds any step that credits, returns or redirects money for a named person, and every message and verdict is filed against the case. See AI agents in banking: what they may and may not do.
Questions people ask
Can an agent talk directly to other banks?
Through standard message formats for tracing and enquiries, yes, at depth 3. Messages that commit the bank to returning funds or accepting liability should be approved by a person.
How does an agent know where a cross-border payment is?
It follows the payment's message trail and the bank's own records, and sends tracing requests along the chain. It reports what it found and what is still unknown.
Should the agent close cases automatically?
It can close cases where the payment is confirmed delivered and the customer has been told. Cases involving money owed to or by the customer should close only after a person has resolved them.