NewNow answering on US lines: healthcare, home services and insurance agents.See agents
VoiFlow
All resources

Industries 6 min read

AI voice agents for banks and lenders: service calls within the rules

Banks want speed and control at once. See how a voice agent verifies callers, blocks a lost card, records every call and hands vulnerable customers to people.

Solutions team
VoiFlow

A bank adviser in a dark suit and slim headset sits at a tidy desk beside a tall window onto a rainy street.

Banks want two things from a phone line at the same time: speed for the customer and control for the bank. A caller who has lost a card wants it blocked now. A customer checking a loan status wants a clear answer without a queue. An AI voice agent for banking can handle those calls within the rules, provided the rules are written down, enforced on every call and reviewed by the people who own them.

An AI voice agent for banking verifies the caller in proportion to the request, handles balance, card and loan-status calls, and can block a lost card through the bank's own systems where that action is permitted. Every call is recorded and scored. Vulnerable customers, complaints and anything involving advice or exceptions go to a person with a full summary.

How much verification is enough?

Verification should match the risk of the request. A question about branch opening hours needs no checks at all. Reading out an account balance needs more. Blocking a card, changing contact details or moving money needs the strongest check, or it goes to a person. A single rule for everything either frustrates customers or leaves the bank exposed.

The verification use case covers the methods in more depth. A workable set of tiers looks like this:

  1. Information only. Opening hours, product descriptions, general fees from published tariffs. No verification.
  2. Account information. Balances, loan status, the next payment date. The caller confirms identity with details your bank chooses, and a second factor if your policy requires it.
  3. Changes and sensitive actions. Card blocks, address or phone changes, payment changes. A stronger check, such as a one-time code sent to the registered number, before the action runs.
  4. Anything else. Disputes, exceptions and complaints. These go to a person.

Whatever tier applies, the agent never asks for a full card number, a PIN, a password or a one-time code read out by the caller. Put that list in writing before go-live, and test that the agent holds to it when a caller pushes.

The common call types

Calls to a bank's service line usually fall into a few groups. The agent can handle the first four end to end and should pass the rest on:

  • Balance and recent payments, after verification.
  • Card questions: a lost or stolen card, a new card that has not arrived, a card that has been declined.
  • Loan status, repayment dates and what a payment will cover.
  • Forgotten details, such as a sort code or a customer reference.
  • Complaints, disputes, hardship and anything that needs an exception.

The lost-card flow, step by step

A lost card is the call where speed matters most, so the flow should be short and clear. This is a typical sequence, and each bank will set its own details:

  1. The caller says the card is lost or stolen. The agent runs the sensitive-action check before it does anything else.
  2. It blocks the card through the bank's systems, if the bank allows the agent to take that action, and confirms the last four digits only.
  3. It explains what happens next in plain words: when a replacement will arrive comes from the bank's policy, not from the agent's guess.
  4. It asks whether any payment the caller does not recognise has been made. It records the answer and passes any suspected fraud to the bank's fraud team at once, with the summary.
  5. It sends a written confirmation with a reference number by SMS or email, so the caller has a record.

The agent does not decide whether a transaction is fraudulent. That judgement belongs to the fraud team, and the agent's job is to capture the facts quickly and accurately.

Spotting vulnerability and handing over

Some customers need more than the standard flow. They may be dealing with bereavement, illness, money worries, confusion or distress. The FCA's guidance FG21/1 describes vulnerability as a spectrum of risk that personal circumstances can increase, and it expects firms to help customers say what they need and to support frontline staff who respond. The FCA's guidance on fair treatment of vulnerable customers is the primary text, and it was updated in July 2026 to point firms to the Consumer Duty pages for current expectations.

This is general information, not legal advice.

In practice, the agent listens for what the customer says, not for labels. If a caller mentions a recent bereavement, says they cannot keep up with payments, or sounds confused about what they are agreeing to, the agent slows down, offers a person and passes the summary. It does not diagnose the customer or record a judgement about their state of mind. It records what they said, in their words, so the person who calls back can respond with care.

Complaints go to people

A complaint is not a service request. The agent records it in the customer's own words, reads it back to confirm, and passes it to the complaints team with the full record. It does not argue the point, offer compensation or promise an outcome. If a call-back is promised, the promise becomes a tracked task that stays open until someone completes it.

The Consumer Duty guide explains what firms are expected to show about how these calls turn out for customers, and the evidence that supports it.

Recording, quality checks and evidence

Tell callers at the start of the call that it is recorded and why. Every call should then be transcribed, scored against your quality rubric, and available for replay. Sample a share of calls for human review each week, and treat the scores as a guide, not a verdict. Recordings also give the bank an audit trail if a customer disputes what was said.

Keep the rules outside the model. Permissions, limits, consent, redaction and the audit trail should be enforced by the platform, so the agent cannot talk its way past them.

What does the agent never do?

Some things stay with trained staff, whatever the caller asks for:

  • Financial advice, such as which product suits a customer or whether they should take a loan.
  • Exceptions, such as fee waivers, overdraft increases or payment holidays.
  • Credit decisions or changes to a credit limit.
  • Hardship arrangements, which need a person who can assess the customer's position.

The agent can explain published terms and point to the right page or team. It should say clearly when a decision is not its to make.

How VoiFlow handles this

VoiFlow agents answer on the first ring, 24 hours a day, with local numbers and voices in the UK. They speak 30+ languages and regional dialects, and they follow the rules you set outside the model, including permissions, limits, consent, redaction and audit trail. Every call is recorded, transcribed, scored and replayable. Supervisors can listen in, whisper to the agent or take over, and a warm handoff sends the full summary to the person who takes the call. Certifications and security details are on the trust centre, and the financial services page shows how this fits a lender's service line.

What to do next

Choose the three call types that generate the most volume and the least risk, usually balance, loan status and card questions. Work with your fraud, security and vulnerability teams to write the verification tiers, the never-ask list and the signals that trigger a handover. Test the agent with staff who play difficult callers, then review the first hundred calls with the team before you widen the scope.

Frequently asked questions

Can an AI voice agent block a lost card?

Yes, if the bank allows the agent to take that action in its systems and the caller passes the sensitive-action check. The agent blocks the card, explains what happens next and sends a written confirmation. Any transaction the caller does not recognise goes to the fraud team with the summary.

What should a banking voice agent never ask for?

It should never ask for a full card number, a PIN, a password, or a one-time code read out by the caller. Keep that list written down, and test the agent against it before go-live.

How does the agent spot a vulnerable customer?

It listens for what the customer says, such as bereavement, illness, money worries or confusion, and it slows down and offers a person. It records the customer's words rather than attaching a label. Your vulnerability policy sets the signals and the handover route.

Are the calls recorded?

Yes, and the agent says so at the start of the call and explains why. Recordings, transcripts and scores support quality checks and any later dispute. To hear a balance or card call sounds like in practice, call the UK demo agent.

Call it now · United States line+1 (831) 282-9226

Hear it run one of your calls.

Call the agent yourself, start a free trial, or bring us your hardest workflow. Most teams are live in days.