Can an AI agent spot a customer in financial difficulty?

Yes, often earlier than a person would, because it sees every transaction and message. An AI agent can notice the signs, pause collection activity, soften its approach, and pass the case to a trained person. Deciding what support to offer, such as forbearance or a payment plan, stays with that person. The signals must never be used against the customer.

What signs can an agent see?

Some signs are in the account. Salary stops arriving, or arrives smaller. The overdraft becomes permanent rather than occasional. Payments to high-cost short-term lenders appear. Direct debits start bouncing. Gambling spending rises sharply. Essentials such as rent or energy go unpaid while the customer keeps up with credit card minimums.

Other signs are in what the customer says: "I've lost my job", "I can't afford this", "I'm behind on everything", or simply a request to delay a payment. An agent reading conversations can catch these even when a busy person would miss them, and it can catch them in a chat about something unrelated, such as a card replacement.

What decides how far the agent goes?

The deciding factor is purpose. Spotting difficulty is only acceptable if it leads to help. The same signals could be used to cut a credit limit, chase harder, or target a high-cost loan offer, and an agent with access to them could do any of those if its brief allowed it. Its brief must not.

So the agent's actions on spotting difficulty are all protective: it stops pressure and opens a route to support. Anything that changes the customer's terms, positively or negatively, is a decision for a trained person. That includes reductions of limits for risk reasons, which may be legitimate but should never be triggered by the agent noticing someone is struggling.

What may the agent do, and what stays with a person?

• Notice signals in transactions and conversations, at depth 1.
• Pause collection contact it controls, such as reminder messages, until a person reviews.
• Change its approach: slower pace, plainer language, no pressure to pay, and no time-limited offers.
• Ask, gently and in approved wording, whether the customer would like help with their finances.
• Point to free debt advice using the bank's approved list.
• Pass the case to the financial support team with the signals it saw.
• Stays with a person: forbearance, payment holidays, interest freezes, payment plans outside the agent's limits, and any decision to record the customer as needing ongoing support.
• Never: reduce credit, suggest new borrowing, or use the signals for marketing.

The point is argued in collections agents must spot difficulty before they ask for money, and the full process is in identifying customers in financial difficulty with an AI agent.

An example: a card question that was not about the card

Take a customer who messages to ask why a card payment was declined at a supermarket. The agent checks: the account is over its overdraft limit. It also sees that the customer's salary has not arrived for two months, that two direct debits bounced last week, and that the customer has made three small payments to a short-term lender.

The agent explains the decline plainly. Then it says, in the bank's approved wording, that it has noticed some payments have not gone through and asks whether the customer would like to talk to someone about support. The customer says they were made redundant and are waiting for a new job to start. The agent pauses automated overdraft reminders, passes the case to the financial support team with the signals listed, and tells the customer when they will be contacted. It also gives details of a free debt advice service.

A person calls the next day and agrees a temporary interest freeze on the overdraft.

What controls make it safe?

• Signals from the system of record. Missed payments and income changes are read from the core banking system, not guessed.
• Protective actions only. The agent's capabilities on detection are limited to pausing, signposting and referring.
• No credit-reducing triggers. The detection signal is not connected to any limit or pricing change.
• Approved wording. What the agent says to a struggling customer is drafted by the support team and certified.
• Referrals filed. Each referral records the signals, the customer's words and what was paused.

Difficulty often overlaps with wider vulnerability; see spotting signs of customer vulnerability with an AI agent. When someone else is controlling the customer's money, the concern is different again, covered in whether an AI agent can spot signs of financial abuse. The general limits are in AI agents in banking: what they may and may not do.