Banks are not short of agent pilots. What they are short of is agents that survive the second meeting with risk and compliance. The question in that meeting is never whether the agent is clever. It is what the agent is allowed to do with a customer’s money, and who answers for it when it does.

Where agents earn their place

The work that pays back first is high in volume, bounded and well documented:

• Servicing. Balances, statements, card status, transfer status and how-to questions, answered from sources the bank has approved.
• Dispute and complaint intake. Gathering the facts, the transaction trail and the customer’s account history into one file a person can decide on.
• KYC refresh preparation. Spotting expiring documents, requesting updates and checking what came back against the file.
• Reconciliation and operations. Matching entries, flagging breaks and drafting the journal that clears them.
• Compliance preparation. Summarising alerts and screening hits so the analyst starts from a finding rather than a list.

The four lines a bank agent does not cross

The first two are enforced by the chain at the moment of action; the last two are scored in Attest before the agent acts at all. None is a sentence in a prompt.

• It does not move money below depth 4. Refunds, payments and transfers are depth 4. A servicing agent’s role stops at 3, so it drafts the movement and holds it for a named person.
• It does not promise what the rail cannot deliver. On an instant rail a settled transfer is final. The agent says so, raises the return request and flags operations; it never implies a reversal it cannot make.
• It does not make a credit decision. It can assemble the file, check it for gaps and summarise it. The decision belongs to a person with the authority to make it.
• It does not answer from outside its sources. If no approved source covers the question, the agent returns a miss and says why. In a bank, a confident answer from nowhere is worse than no answer.

A worked case: the wrong account

A customer writes at 9pm: they sent $3,000 to the wrong account an hour ago and want it back. The servicing agent reads the transfer: settled, on an instant payment rail where settlement is final. Reversal is not a capability any agent holds, at any depth, because on this rail it does not exist.

What the agent can do is depth 2 and 3 work. It drafts the return request to the receiving institution, opens an operations case with the transaction reference and the customer’s statement, and tells the customer exactly what happens next and when. The ledger records the refusal, the reason and the case it opened. When the auditor asks how the bank handles misdirected transfers, the answer is a query, not a reconstruction.

Data, regulation and accountability

Customer data is personal data, governed wherever the bank operates: the GDPR in the EU, the UK GDPR, state privacy laws in the US, Nigeria’s Data Protection Act 2023 and their equivalents elsewhere. The agent reads only what its role permits, and every decision it takes, reads included, is recorded. Under the EU AI Act, systems that assess the creditworthiness of individuals are classed as high-risk; after the 2026 Digital Omnibus those obligations apply from 2 December 2027. Our regulation guide maps each obligation to the control that meets it.

Accountability stays where it was. The agent prepares, a named person decides what is above the agent’s depth, and the record shows which one did what.

Choosing the first use case

The first agent a bank deploys sets the tone for every one after it. Pick work that scores well on four tests:

• Volume. Enough cases a day that the time saved is visible within a month.
• Reversibility. Mistakes can be caught and corrected before they reach a customer or a ledger.
• Written policy. The rules already exist in a document the agent can read and cite. If they live only in people’s heads, write them down first; the agent will only be as consistent as the policy.
• A measurable outcome. Handling time, backlog, error rate or customer wait, measured before the agent starts.

Servicing questions and dispute intake usually pass all four. Payments and credit fail the second test, which is why they come later and stay held.

What your supervisor will ask

Banking supervisors differ by market, but their questions about AI in operations are converging. Expect these, and have the answer in the record rather than in a slide:

• Who approved the agent’s level of autonomy, on what evidence, and when is it reviewed?
• How do you know it is still performing as approved, today?
• Where are its refusals and holds recorded, and who reviews them?
• How is the vendor managed as a third party, and where is customer data processed?
• What happens if you need to switch it off, and how quickly can a person take the work back?

The first 90 days

In a standard deployment the agent spends up to four weeks in training against the bank’s own policies and past cases, then two weeks shadowing live work where every output is a draft. Certification follows, depth by depth. By day 90 a servicing agent can be answering routine questions at its certified depth, drafting the rest, and producing a record the bank’s risk team has already reviewed once. Our deployment guide walks through each step.

Questions people ask