How to train staff to answer customers who ask whether an AI decided

Staff should answer from the record, not from a script: say plainly whether an agent drafted, decided or refused, name the role of the person who approved, and offer a review by a person. Train this with the decision record open, role-play on the hardest cases, and a short list of situations where a person is offered without being asked.

Why does this question need training?

Customers ask "Was this decided by AI?" when something has gone against them: a refund declined, a claim queried, an application refused, an account restricted. A vague answer sounds evasive. A wrong answer, such as "No, a person decided" when an agent executed at depth 3, is worse, because the record will show otherwise.

Front-line staff rarely know the answer by instinct. The same team may handle work the agent executed alone, work it drafted for a person to send, and work it held for a named decision-maker. Training must teach staff to look it up and say it in plain words.

The head of customer operations owns the programme. Compliance agrees the wording standards. Team leads run the practice.

What do the rules expect?

Several regimes bear on this. Under the EU AI Act, Article 50 requires that people are told when they interact with an AI system. GDPR Article 22 restricts decisions based solely on automated processing that have legal or similarly significant effects, and gives people routes to human involvement; UK GDPR is the UK equivalent. In the US, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used. In the UK, the FCA's Consumer Duty expects firms to support good outcomes and customer understanding.

None of these is satisfied by a front-line script alone, but each is undermined by an unclear or inaccurate answer. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

The steps

1. Person (compliance and head of customer operations): agree three plain answers, one for each pattern in the table below, and a list of situations where staff offer a person without being asked.
2. Person (staff member): learns to read the record: the verdict, the depth at which the agent acted, the bound that decided, and the name or role of any person who approved.
3. Agent, depth 1: retrieves the decision record for the customer's case on request, showing each step and who took it.
4. Person (staff member): answers the customer from the record, in their own words, using the agreed plain answer as a base.
5. Person (staff member): offers a person when the customer asks, when the outcome is significant, or when the customer disputes the facts.
6. Agent, depth 2: drafts a written explanation if the customer wants one; the staff member checks and sends it.
7. Person (team lead): runs weekly role-play on recent difficult calls.
8. Person (team lead): samples answers from recorded calls and messages for accuracy against the record.

What do good answers sound like?

Teach plain sentences built from the record. For an executed refund decision: "Yes. Our automated system applied our refund policy to your order. I can see why it declined, and I can ask a colleague to review it." For a drafted letter: "Our system drafted that letter and one of our claims handlers checked and sent it." For a held decision: "Our system prepared it, and a member of our credit team made the decision."

Avoid three things: denying an agent was involved when it was; implying a person reviewed something that was not reviewed; and hiding behind "the system" without offering a person. The goal is an answer the customer can rely on and the record can support.

When should staff offer a person without being asked?

• The outcome has a significant effect: credit, insurance cover, account access, a benefit or a large refund.
• The customer says a fact the agent relied on is wrong.
• The customer seems vulnerable, distressed or confused about what happened.
• The agent's verdict was a refusal and the customer has not been told how to challenge it.

Which controls matter, and what evidence do they leave?

• Agreed plain answers. Evidence: the approved wording and its version.
• Record lookup before answering. Evidence: the record access logged against the contact.
• Offer of a person. Evidence: the offer and the outcome noted on the case.
• Sampling. Evidence: the team lead's accuracy checks.

What should you measure?

Answers that matched the record, from the weekly sample. Offers of a person made when required. Complaints alleging that the firm was unclear about automation. And how often a person's review changed the outcome, which tells the builder where the agent's bounds may need work. Measuring training impact more broadly is covered in how to measure whether AI agent training worked.

Where does it usually go wrong?

Firms write one generic script. It is wrong for at least one of the three patterns.

Staff are not given access to the record, so they guess.

Managers discourage offering a person to protect handling time. That is a decision for leaders to own openly, and why middle managers decide whether agents work explains why they matter here. Boards should ask about it too; see what a board should ask about AI agents. For teams serving customers in several languages, the plain answers must be agreed in each; see answering customers in several languages with an agent. The guide to training teams for AI agents covers the wider programme.

Where the Academy fits

Pryme Intelligence Academy's Foundations track explains the decision chain, depth and the four verdicts in about four hours, which is what front-line staff need to answer this question accurately. It is free and earns the Foundations badge. Reviewers who decide held items continue to the Operator track.