How do you know if customers trust an AI support agent?

Watch what customers do after the agent answers, not what they say. Customers who trust an agent act on its answers, do not call back to check them, and stop asking for a person in the first message. Customers who do not trust it ask again elsewhere. Those behaviours are measurable and say more than a satisfaction score.

What does distrust look like in the data?

Distrust rarely shows as a complaint. It shows as checking. A customer gets an answer from the agent about when their refund will arrive, then phones the next day to ask a person the same question. A customer types "agent" or "human" as their first message, before the agent has said anything. A customer abandons the chat halfway and emails instead.

Each of these is a signal you can count:

• Verification contacts. The same customer, same topic, a different channel or a person, within a short window after the agent answered.
• Immediate requests for a person. The share of conversations where the customer asks for a human before the agent's second reply. Watch the trend, not the level.
• Abandonment at specific moments. If customers drop out right after the AI disclosure, or right after identity checks, that tells you where trust breaks.
• Action taken. Did the customer follow the agent's instruction: use the returns link, submit the meter reading, keep the appointment? Bookings that are kept are a quiet sign of trust; see how an AI support agent books appointments.

Which steps give you a reliable reading?

1. Split the measures by what the agent did. Trust after a clear answer, after a refusal and after a handover are different things. A customer who was refused and still rates the conversation well has been treated fairly.
2. Compare with your human team on the same contact reasons. Recontact after a person answers is the baseline, not zero.
3. Read the verification contacts. For a sample, have a person read what the agent said and what the customer asked next. Sometimes the agent was right and the customer simply did not believe it. Sometimes the agent was vague.
4. Ask a single plain question after the conversation. "Did you get what you needed?" is more useful than a scale about the AI.
5. Watch complaints that mention the agent. Few in number, high in detail.

What builds trust in the first place?

Mostly three behaviours, all of which you can control.

Saying what it is. Customers who know they are talking to an AI, and know how to reach a person, relax. Customers who discover it later feel tricked. The argument is made in full in customers trust agents that say what they are.

Admitting what it does not know. An agent that says "I can't confirm that; I've passed it to the billing team, who'll reply by tomorrow" is trusted more than one that answers confidently and wrongly. What should happen next is set out in what happens when a support agent does not know the answer.

Being careful with security. Customers trust an agent that checks who they are before discussing their account, and are rightly suspicious of one that does not. The same goes for scams: an agent that handles a scam report calmly and quickly earns trust that lasts, as covered in handling scam reports from customers with an AI agent.

An illustrative example

Take a regional bank that launched a chat agent for card queries. Satisfaction scores looked fine. But the team noticed that many customers who asked the agent whether a card payment had gone through phoned within an hour to ask again.

A person read a sample. The agent's answers were correct, but it said "the payment appears to be pending", copying the system status. Customers read "appears" as uncertainty. The team changed the approved wording to say plainly what pending means and when it would clear. Verification calls on that topic fell over the following weeks.

Nothing about the agent's accuracy changed. Its wording did. That is a common finding: trust problems are often clarity problems.

What controls keep the measure honest?

• Count recontacts across all channels, not just the one the agent works in.
• Do not reward containment. An agent rewarded for keeping customers away from people will learn to make asking harder, and trust falls.
• Keep the human route visible. A customer who knows they can reach a person is more likely to give the agent a chance.

Identity checks deserve a note. Customers sometimes read repeated questions as distrust of them. Explaining why the check exists, in one line, helps; the wider purpose of KYC is protection, and customers respond to that framing.

Where Pryme fits

In Pryme, every verdict is filed, so you can split recontact and satisfaction by whether the agent answered, refused or handed over. That shows whether customers trust its answers, its refusals, or neither. More is in the AI support agent that knows when to say no.