AI contact center agents for banks

The useful split is not by difficulty of language. It is by consequence of being wrong.

Suited to automation: balance and transaction enquiries, card activation, temporary card block, statement requests, branch and rate information, payment status, appointment booking, and the whole class of “where is my money” questions that dominate call volume and have a definite answer.

Not suited, today: anything where the customer is in distress, disputes requiring judgement about liability, hardship and collections conversations, suspected fraud where the caller may be under coercion, and any case where the right answer depends on discretion the bank has deliberately reserved for a person.

The boundary worth designing to is this: an AI agent should handle the calls where being wrong means a customer repeats themselves, and should hand over the calls where being wrong means a customer is harmed.

Contact-centre verification has historically leaned on shared secrets — date of birth, the last few transactions, a memorable word. Those are exactly the details that appear in a data breach or that a caller can be socially engineered into revealing to someone else. Automating the same questions makes them faster to attack, not safer.

The stronger pattern moves verification off the call. The agent sends a push to the customer’s enrolled device; the customer approves with a passkey; the bank receives a cryptographic assertion that the enrolled device and its owner approved this interaction. No secret is spoken aloud, so nothing said on the call is worth intercepting, and the agent knows who it is talking to before any account data is read out.

It also handles the case that defeats knowledge-based verification entirely: a caller who knows all the right answers because someone else told them to say them.

Voice biometrics fits alongside this as a passive signal — useful for flagging a mismatch mid-call, not sufficient as the sole gate before moving money, particularly given how far speech synthesis has moved.

Supervisors across these markets are converging on a simple expectation: a customer speaking to an automated system should be told, in terms they will actually register, and should be able to reach a person.

Disclosure works when it is early, plain and unprompted. A short statement at the opening of the call — that this is an automated assistant and a person is available on request — meets the expectation. What does not meet it is a disclosure buried in terms, delivered only if asked, or worded to imply a human is present.

A related decision worth making deliberately: do not give the agent a human name and a fabricated persona. It creates precisely the impression disclosure exists to prevent, and it makes the eventual handoff feel like a deception rather than a transfer.

Complaint handling is usually the most prescriptive part of a supervisor’s consumer-protection rules: complaints must be recognised, logged, acknowledged within a defined period, and reported.

That places a specific requirement on the agent. It must recognise a complaint even when the customer never uses the word — “this is the third time I’ve called about this” is a complaint — and it must log it into the same complaints system a human agent would use, starting the same clock. An AI agent that resolves the immediate question but never registers the complaint creates a compliance gap that is invisible until it is audited.

The safe default is to treat ambiguity as a complaint and log it. Over-logging is a reporting nuisance; under-logging is a breach.

Handoff quality is what customers actually judge the system on. Three rules cover most of it.

• Always available, never earned. A request for a person should be honoured immediately, at any point, without a retention attempt. Making customers argue for it is the single fastest way to turn a working system into a complaint generator.
• Carry the context across. The person who picks up should see the transcript, the verified identity and what has already been tried. Making the customer repeat everything erases the benefit of the automated portion entirely.
• Hand off on distress, not just on request. Frustration, confusion and vulnerability signals should trigger a transfer before the customer has to ask.

Design for what happens when no person is available, too. An honest “no one is free right now, here is a callback at a time you choose” is far better than a queue with no estimate.

An AI agent does not change what must be recorded, but it does add material: alongside the audio, there is now a transcript, a record of the actions the agent took, and the reasoning trail behind them.

Treat all of it as call records. It should live under the same retention period, the same access controls and the same residency requirement as the audio. The action log in particular is what lets you answer, months later, why the agent did what it did on a specific call — which is the question an investigation actually asks.

Where a supervisor requires customer data to remain in country, that requirement covers the transcript and the model processing, not only the recording. This is the common reason a contact-centre AI deployment ends up on-premise or in an in-region cloud.

Containment — the share of calls handled without a human — is the metric most often quoted and the most misleading on its own. A system can contain a call by exhausting the customer.

• Resolution, not containment: did the customer’s problem end?
• Repeat contact rate within a few days, which exposes false resolutions.
• Handoff rate and handoff reason, which tells you where the boundary actually sits.
• Complaint capture rate, audited against sampled transcripts.
• Verification success rate, which usually surfaces enrolment gaps rather than agent problems.

Score every conversation rather than a sample. The cost of reviewing them all is now low, and full coverage is what turns quality assurance from an audit exercise into something that catches problems in the week they appear.

FortAgent takes inbound and outbound calls with the caller verified through FortAuth before any account action, clear AI disclosure at the opening, complaint logging into your existing system, and a human handoff that carries the transcript and verified identity across. Every conversation is transcribed, scored and written to an audit trail, and it deploys on-premise or in-region where residency requires it. Governance expectations by market are on the compliance pages.