Can an AI agent handle complaints about itself?

Partly. An agent can take a complaint about itself, record it in the customer's words, correct a plain factual error from the record, and pass it to a person. It should not investigate its own conduct or decide whether it was at fault. That is a conflict of interest, and a person reviews the conversation and the filed record instead.

Why can't the agent judge its own conduct?

For the same reason a member of staff does not investigate a complaint about themselves. The agent will read its own earlier reply through the same instructions, sources and blind spots that produced it. If it misread a policy the first time, it will likely misread it again and conclude it was right.

There is also a tone problem. An agent asked "why did you tell me that?" tends to explain and justify. To an upset customer, that reads as the machine defending itself. A short acknowledgement and a handover serves them better.

What can the agent still do?

Some things are safe and useful, because they do not involve judging the earlier conduct.

• Record the complaint accurately, quoting the customer and linking the conversation they are complaining about.
• Correct a fact. If it gave a wrong date or figure, it can give the right one from the record now. Correcting is not the same as deciding what the error cost the customer.
• Preserve the evidence. The earlier conversation and the verdicts behind it must be kept.
• Offer a person, straight away, without trying to win the customer back first.

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

The full process is set out in how to handle complaints about an AI agent's own decisions with an AI agent. The general rules for complaints are in whether an AI agent can handle customer complaints.

What does this look like in practice?

Take a consumer lender. A customer asks the support agent for a settlement figure to pay off their loan, pays it, and a week later finds a balance still on the account. They write: "Your chatbot gave me the wrong number."

The current agent checks the ledger and gives the correct remaining balance, with the date it was calculated. It logs a complaint flagged as concerning the agent, links the earlier conversation and its filed verdict, and tells the customer a person will review what happened and respond. It does not say the earlier figure was "correct at the time" or speculate about why.

The complaints handler reads the record and finds the agent quoted a figure that excluded interest accruing up to the payment date. The lender waives the difference, the handler replies, and the agent's owner pauses its settlement-figure capability until it is retested.

Why are these complaints worth so much?

Because they show exactly where the agent fails, with the evidence attached. A complaint about the agent is a free test case from the real world. Each one should feed back into testing, as argued in complaints about the agent are the most useful complaints.

Which controls make it safe?

• A complaint category for the agent itself, routed only to people.
• Automatic linking of the conversation complained about, with its filed verdicts.
• A rule against the agent explaining or defending its earlier answer.
• A review of the agent's capability whenever a complaint about it is upheld.
• Every upheld complaint added to the agent's test set.

Where Pryme fits

In Pryme every verdict the agent reached is filed with its sources and the bound that decided it, so a person can reconstruct exactly what happened. An upheld complaint can trigger a fresh Attest run, and a run that clears no depth drops the agent to depth 1 the same day. The wider approach is in the guide to AI customer support agents.