Yes, within clear limits. An AI agent can tell customers their balance, explain a charge, trace a payment and handle routine service requests well. It should not give personal advice, make credit decisions, discuss anything touching a financial crime investigation, or carry on when a customer shows signs of vulnerability. Customers must know they are talking to an AI.
What are customers asking about?
Most money conversations with a bank are factual. Where is my payment? Why was I charged this fee? When does my fixed rate end? What is the balance on my loan? These questions have answers in the bank's own systems, and an agent that reads those systems can answer them accurately and at any hour.
A smaller share are judgement calls. Should I move my savings? Can I afford this? What should I do about my debt? Those carry a duty of care the bank cannot hand to an agent, and the agent's job there is to recognise them and connect the customer with a person.
What may the agent say, and where must it stop?
| Conversation | The agent may | Where it stops |
|---|---|---|
| Balances, transactions, statements | Answer from the ledger | Nothing routine |
| Fees and charges | Explain what was charged and why; refund within policy | Refunds above the bank's limit |
| Payment status and tracing | Trace and explain; start a recall request | Releasing or recalling funds |
| Product features | Describe terms and rates as published | Recommending a product for this customer |
| Affordability or debt worries | Acknowledge, offer support routes | Assessing affordability or agreeing forbearance |
| Account restrictions or blocked payments | Say a check is under way, in approved wording | Explaining why, if it touches a financial crime review |
| Credit applications | Collect information, explain the process | Any decision or reason for decline |
The restricted-account row needs care. Telling a customer their account is under review for suspected money laundering can be an offence in many markets; see tipping-off risk when agents talk to customers.
Why must the customer know it is an AI?
Because it changes how they weigh what they are told, and because the rules increasingly require it. In the EU, Article 50 of the AI Act requires that people are told when they interact with an AI system, unless it is obvious. The bank's obligations for its wider use of agents are covered in the EU AI Act for banks deploying AI agents.
Disclosure should be plain and early, with a visible way to reach a person at any point.
How does the agent spot when to stop?
It checks for named signals, not vague unease. A customer who mentions bereavement, illness, job loss or confusion; a customer who says they are being told what to do by someone else; a request to move all their money urgently; repeated failed attempts to understand an answer. Each of these hands the conversation to a person, with a summary so the customer does not repeat themselves.
The agent should also stop when it cannot ground an answer. If the ledger does not say, the agent says it does not know and offers a person. A confident wrong answer about someone's money is worse than no answer.
An illustrative example
Take an e-money firm whose customers ask about delayed transfers. A customer writes that a payment of 600 to their landlord has not arrived and the rent is due tomorrow.
The agent checks the payment record: sent yesterday, accepted by the receiving bank, with no return. It tells the customer the payment left, gives the reference, and explains what the landlord's bank may need. The customer then asks whether their money is safe if the firm fails. The agent gives the published explanation of how customer funds are protected, which rests on the firm's daily work to reconcile e-money safeguarding accounts. When the customer adds that they are worried about paying rent at all this month, the agent offers to connect them with a person who can talk through support, and passes a short summary.
What controls make this safe?
- Answers grounded in the bank's systems, with the source recorded for each fact.
- No money moves from a conversation without the bounds in how banks stop AI agents from moving money without approval.
- Approved wording for restricted accounts and anything near a financial crime review.
- Vulnerability signals that hand over to a person, tested before go-live.
- Validation on the bank's own past conversations, scored for grounding, refusal and escalation; see how banks validate an AI agent.
- A full record of every conversation and every handover.
The bank-wide limits are in AI agents in banking: what they may and may not do.
Questions people ask
Can the agent tell a customer their loan application was declined?
It can deliver a decision a person or an approved process has made, using approved wording. Explaining the reasons, and handling the customer's response, should follow the bank's credit rules and usually involves a person.
What if the customer refuses to talk to an AI?
Offer a person straight away. Forcing customers through an agent they do not want erodes trust and may conflict with duties to treat customers fairly.
Should the agent remember past conversations?
It can use the bank's own record of past contacts, the same as staff would. It should not keep a separate memory of its own outside the bank's systems.