Noticing is the agent's job; diagnosing is not. It listens for signs that a customer's circumstances make them more likely to be harmed, adjusts how it is helping at once, records the customer's own words, and offers a trained colleague. It never labels the customer. A named person decides whether to record a lasting marker and what support follows.
What is the process, and who owns it?
A customer in vulnerable circumstances is someone who, because of their health, a life event, their financial resilience or their ability to deal with you, is especially likely to be harmed if things go wrong. The circumstances can be lasting or temporary: a bereavement, a new diagnosis, a job loss, a language barrier, someone else putting pressure on them. Most do not say "I am vulnerable". They mention it in passing, or it shows in how they write.
The vulnerability lead, or the customer outcomes lead in smaller firms, owns the process, the signal list and the support options. The data protection owner owns how vulnerability information is recorded, because much of it is special-category data. A trained team takes the customers the agent refers. The wider support position is in the AI support agent that knows when to say no.
What does the agent prepare, step by step?
- Agent, depth 1: notices signals in the conversation. Mentions of illness, bereavement, separation, job loss, caring duties, debt worries; confusion about something already explained; someone else answering for the customer; distress. Evidence: the passage that raised the signal.
- Agent, depth 1: adapts immediately. Slows down, uses shorter sentences, stops any sales or retention step, and checks understanding. This happens whether or not anything is recorded.
- Agent, depth 1: offers a person. "It sounds like a lot is going on. Would it help to speak to a colleague who can take more time with this?" Evidence: the offer and the answer.
- Agent, depth 2: asks before recording. With the customer's agreement, it notes their own words and what they said would help. Without agreement, it records only that support was offered. Evidence: the consent answer.
- Agent, depth 2: drafts a referral to the trained team: what was said, what the customer asked for, what is urgent. No labels or guesses about conditions.
- Person, trained team member: decides what follows. A support plan, a lasting marker on the account, a change of communication method, a pause on collections. Evidence: the decision and its basis.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Notice | Identifies signals in the conversation | 1 | Nothing |
| Adapt | Changes pace and stops sales steps | 1 | Nothing |
| Offer | Offers a trained colleague | 1 | Taking the conversation |
| Consent | Asks before recording | 2 | Nothing |
| Referral | Drafts a factual summary | 2 | Deciding support |
| Marker | Nothing | None | Recording a lasting marker |
What does a named person approve, and why?
The lasting record. A marker on an account follows the customer into every future contact and every system that reads it. It can protect them, and it can also lead to them being treated differently in ways they did not choose. That decision belongs to a trained person who has spoken with the customer and understands the data protection position.
The signal list itself is also approved by the vulnerability lead, together with the wording the agent uses to offer help. A clumsy phrase at the wrong moment does harm. Two signals need their own routes. Possible financial abuse, where someone else seems to be controlling the customer's money, is covered in can an AI agent spot signs of financial abuse. Financial difficulty has its own options and is covered in can an AI agent spot a customer in financial difficulty.
Which controls matter, and what evidence do they leave?
Adapt first, record second. The change in how the agent helps does not wait for consent or a referral. Evidence: transcripts sampled by the vulnerability lead.
Customer's words, not labels. The agent records "customer said their father died last month", never a category. Evidence: the note text.
Consent before recording. Evidence: the consent answer stored with the note.
Sales and retention steps stopped. Once a signal is raised, offers and upsells are refused by a platform check for the rest of the conversation. Evidence: refused verdicts on any attempted offer.
Referral reaches a person. Evidence: time from referral to first contact by the trained team.
Conduct duties. In the UK, the FCA's Consumer Duty expects firms to deliver good outcomes for retail customers, including those with characteristics of vulnerability. Other markets have their own expectations. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where does it usually go wrong?
Keyword lists. The agent reacts to "died" in "my phone died" and misses a quiet "things have been hard since my husband passed". Signals are about meaning, and Attest should test them on your own past conversations.
Over-flagging. Every mention of a cold produces a referral, the trained team drowns, and real cases wait. Calibrate the list with the team that takes the referrals.
Labels harden. A temporary circumstance becomes a permanent marker. Markers need a review date.
Third parties take over. Someone on the line answers every question for the customer. The agent should ask to hear from the customer directly, gently, and refer if that does not happen.
The wrong measure. Satisfaction scores do not show whether a vulnerable customer was protected. Use outcomes. A satisfaction score has its uses, explained in what is a customer satisfaction score, but this is not one of them.
What should you measure?
- Signals raised that the trained team confirms as relevant, and signals they find in transcripts that the agent missed.
- Time from referral to contact.
- Offers refused after a signal. Should be zero sent.
- Markers with a review date, and reviews completed on time.
- Complaints from customers who were referred.
Questions people ask
Can the agent ask customers directly whether they are vulnerable?
No. The question is intrusive and few people would describe themselves that way. Asking whether there is anything that would make the conversation easier is a better route.
What if the customer refuses a referral?
The agent respects that, keeps the adapted approach for the rest of the conversation, and records only that support was offered and declined.
Does the agent see markers from earlier contacts?
If a trained person has recorded one with the customer's agreement, yes, so it can adapt from the start without asking the customer to explain again.