The agent that made the decision never handles the complaint about it. A separate intake agent acknowledges the complaint, pulls the filed verdict, sources and bounds, and rebuilds what happened. A person reviews the decision afresh. If the complaint is upheld, the customer is put right, similar cases are checked, and the agent's brief or bounds are fixed.
What is the process, and who owns it?
As agents take on more customer-facing work, some customers will complain about what an agent did: refused a request, gave wrong information, held a payment, closed a chat too early, misread a document. These complaints are about a decision made by software under authority the firm granted. They carry two questions at once. Was this customer treated fairly? And is the agent working within the bounds it was given?
The complaints function owns the complaint. The owner of the agent, usually the business team that deployed it, owns the fix. Risk or compliance owns any look-back across other customers. The intake agent is a helper inside that process, under the rules in our guide to the AI support agent that knows when to say no. The case for treating these complaints as a gift is made in complaints about the agent are the most useful complaints.
What are the steps, and at what depth?
- Agent (intake), depth 1: recognises that the complaint concerns an agent's decision, from the customer's words or from the case history.
- Agent (intake), depth 3: acknowledges within your deadline, using the approved template, and tells the customer a person will review the decision. The process in acknowledging formal complaints within deadlines applies.
- Agent (intake), depth 1: rebuilds the decision from the append-only record: what the customer asked, the agent's verdict, the bound that decided it, the sources used and their provenance, the certificate in force.
- Agent (intake), depth 2: drafts a decision review pack: the timeline, what the agent did, what its bounds allowed, and where the decision may have gone wrong.
- Person, complaints handler: reviews the decision afresh, without treating the agent's verdict as correct by default.
- Person, complaints handler: decides the complaint and any redress, and writes to the customer.
- Person, agent owner: fixes the cause if upheld: a source corrected, a bound tightened, the brief changed. The agent is certified again before returning to the same depth.
- Agent (intake), depth 2: proposes a look-back: other conversations where the same decision path was taken.
- Person, risk or compliance: decides the look-back scope and any wider remediation.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Recognise | Spots agent-decision complaints | 1 | Nothing |
| Acknowledge | Sends template | 3 | Wording |
| Rebuild | Reconstructs from the record | 1 | Nothing |
| Pack | Drafts review pack | 2 | Reviewing it |
| Decide | Nothing | none | Outcome and redress |
| Fix | Nothing | none | Agent owner changes the agent |
| Look-back | Proposes similar cases | 2 | Scope and remediation |
Which controls matter, and what evidence does each leave?
Separation. The agent complained about cannot see or act on the complaint. Evidence: case access records showing a different agent and a named handler. Can an AI agent handle complaints about itself? explains why.
Full reconstruction. If the decision cannot be rebuilt, that is a finding in itself. Evidence: the review pack with references to each filed verdict.
Human review where the rules require it. Under GDPR Article 22, people have rights around decisions based solely on automated processing with legal or similarly significant effects, including obtaining human intervention. Evidence: the handler's review and decision.
Fix before reinstatement. Evidence: the change to the agent and a new certificate.
Look-back decided by a person. Evidence: scope, cases checked and outcomes.
What should you measure?
- Complaints about agent decisions, by agent, capability and verdict type.
- Upheld rate by capability. A capability with several upheld complaints should be re-examined, even if the agent's certificate is current.
- Decisions that could not be fully rebuilt. The target is none.
- Time from upheld complaint to fix and re-certification.
- Customers remediated through look-backs.
Where does it usually go wrong?
The decision is defended, not reviewed. Handlers assume the agent followed the rules because it usually does. Review the decision as if a new colleague had made it.
Only the customer is fixed. Redress is paid, but the agent keeps making the same decision. Every upheld complaint needs an owner for the cause.
Refusals are treated as correct by definition. A refusal can be wrong too, if the bound was set too tight or the source was stale.
The record shows the wrong version. The agent's brief or sources changed after the decision, and the reviewer judges the old decision against the new rules. The review pack must show the brief, sources and certificate as they were at the time.
Complaints are not labelled. Agent-related complaints are mixed into general themes and nobody sees the pattern. Tag them at intake.
First-contact resolution is attempted. Resolving at first contact, as in resolving complaints at first contact with an AI agent, suits many complaints, but not those about an agent's own decision, which need a person's review. The wider limits are in can an AI agent handle customer complaints?
Where Pryme fits
Every verdict in Pryme is filed with the bound that decided it, the sources and their provenance, and the certificate in force, so the intake agent can rebuild any decision. When a cause is fixed, Attest certifies the agent again on your cases, including the one complained about, before it returns to its previous depth.
Questions people ask
Should the customer be told an AI made the original decision?
Yes. They should already have known when they dealt with the agent, and the complaint response should be clear about how the decision was made and that a person reviewed it.
Does an upheld complaint automatically suspend the agent?
Not automatically, but it should prompt a check. If a certification run clears no depth as a result, the agent drops to depth 1 the same day.
Who owns the agent's mistakes?
The firm. Accountability sits with the named people who own the agent and the capability, not with the software or its vendor.