Vulnerable customers need a person sooner

Support agents should recognise the signs that a customer may be vulnerable and bring in a person early, before the conversation reaches a decision. Bereavement, illness, financial distress, confusion or any mention of harm change what good service looks like. In those moments an agent's job is to notice, respond kindly and hand over quickly to someone trained to help.

What signs should an agent recognise?

A vulnerable customer is someone whose circumstances make them especially susceptible to harm, particularly when a firm does not act with appropriate care. Vulnerability is often temporary and rarely announced. Customers do not usually say "I am vulnerable". They say things that suggest it.

Signals an agent should be trained and certified to notice include:

• Life events: a death in the family, a divorce, a job loss, a serious diagnosis.
• Financial strain: mentions of not being able to pay, choosing between bills, borrowing to make repayments.
• Health: references to illness, hospital, medication, mental health or disability.
• Capacity: confusion about basic facts, repeated questions, difficulty following explanations.
• Distress: expressions of panic, hopelessness or being overwhelmed.
• Third parties: someone writing on behalf of another person who cannot manage their own affairs.
• Harm: any mention of self-harm, abuse or danger.

The last category needs the fastest response of all, and firms should have a clear, tested route for it.

Why sooner rather than at the usual escalation point?

Because the usual escalation points are built around transactions, not people. A support agent typically escalates when a request exceeds its authority: a refund too large, an exception to policy, a complaint. A vulnerable customer may reach none of those points. They might ask a simple question about a payment date, and the agent could answer it perfectly, while missing that the customer has just said they cannot afford to eat this week.

Waiting for a transactional trigger means the customer may be given a standard answer, a standard payment plan or a standard script at the moment they needed flexibility and a human voice. By the time the conversation reaches a person, the harm may already be done: a customer agreed to a repayment they cannot sustain, or gave up and stopped engaging.

Bringing a person in early, at the first credible signal, avoids that. It also tends to be cheaper in the end, because problems handled with care at the start are less likely to return as complaints, defaults or disputes. Whether an AI agent can help a vulnerable customer at all is a fair question; the answer is that it can help most by noticing and handing over well.

What should the agent do in the moment?

Four things, in order:

1. Acknowledge what the customer has said, briefly and humanly. "I'm sorry to hear about your father" is better than ignoring it or launching into a script.
2. Pause the routine process. Do not continue with the payment, the upsell or the standard next step.
3. Offer a person, explaining that a colleague who can help with their situation will take over, and roughly when.
4. Pass on what matters, so the customer does not have to explain their circumstances again. The handover note should record the signal, in the customer's words, and anything already discussed.

The agent should keep its language simple and avoid jargon. It should not rush the customer, and it should not try to solve the underlying problem itself.

What should the agent avoid?

Some well-meant behaviours cause harm:

• Diagnosing or labelling. The agent should not tell the customer they are vulnerable or suggest what their condition might be.
• Intrusive questions. It should not probe for detail about health or circumstances beyond what is needed to route the case.
• Recording more than necessary. Health information is special-category data under GDPR, and the agent should not store detail it does not need. A note such as "customer mentioned a health condition affecting their finances; handed to specialist team" is usually enough.
• Continuing to sell. No retention offers, cross-sells or upgrades once a vulnerability signal appears.
• False comfort. It should not promise outcomes, such as that charges will be waived, that a person has not yet decided.

A vulnerable customer handling checklist helps make these consistent across the agent's brief, its certification cases and the receiving team's process.

What do the rules expect?

In the UK, the FCA's Consumer Duty expects firms to deliver good outcomes for retail customers, including those with characteristics of vulnerability, and to make sure those customers are not systematically worse off. That expectation applies whether a customer deals with a person or an agent. The specific implications for agents are discussed in FCA Consumer Duty, vulnerable customers and AI agents. Other markets have their own consumer protection and fairness rules with similar effect.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

How do you test whether the agent recognises vulnerability?

Through certification cases built from real conversations, with personal details removed, where vulnerability signals appear in different forms: stated outright, mentioned in passing, implied by behaviour, or raised by a third party. Each case is labelled with the correct outcome, which is almost always an early handover with a good note.

Score the agent on whether it noticed, how it responded and whether it stopped the routine process. Include cases that look similar but are not vulnerability, such as a customer mentioning a holiday, so the agent does not over-escalate. And watch live conversations: a fall in vulnerability handovers without a reason is a signal to retest. The general guidance on when an AI support agent should escalate to a person applies, but vulnerability triggers should be tested as their own slice.

Where Pryme fits

In Pryme, vulnerability handovers are tested in Attest through escalation and harm cases built from the customer's own conversations, and each gate must clear on its own. Once a signal is detected, the customer's own limits can hold any further action for a named person. This reflects the stance of the AI support agent that knows when to say no.