Agents are good at the mechanical part of complaints handling: spotting that a message is a complaint, starting the clock, acknowledging it and chasing deadlines. That is the easy part. Deciding whether the customer was treated fairly, what went wrong and what redress puts it right needs a person with the authority and judgement to answer for it.
Which parts of complaints handling suit an agent?
Complaint handling has a large administrative core. Each complaint has to be recognised, logged, acknowledged, assigned, tracked against regulatory and internal deadlines, and closed with a written response. Most firms that struggle with complaints struggle here first: cases sit unassigned, acknowledgements go out late, deadlines pass without anyone noticing.
Agents handle this work well. They can read every incoming message and flag those that express dissatisfaction, open a case, send an acknowledgement in the firm's approved wording, set the deadlines that apply, and remind the handler as each one approaches. The methods are set out in how to track complaint deadlines with an AI agent and how to acknowledge formal complaints within deadlines with an AI agent.
Acknowledgement in fixed wording is routine work within bounds, so it can run at depth 3 once certified. Tracking is depth 1. None of it requires the agent to form a view about the complaint itself.
Why are outcomes and redress different?
Because they are judgements about fairness, and fairness depends on facts and context an agent can assemble but should not weigh. Was the fee correctly charged under the terms? Even if it was, was the customer properly told? Did the firm's delay cause a loss, and how large? Is the customer in circumstances that change what a fair outcome looks like?
Answering those questions means deciding between competing accounts, interpreting terms, judging whether the firm met its own standards, and choosing a remedy. In the UK, the FCA's Consumer Duty expects firms to deliver good outcomes for retail customers, and complaints are one of the main places that duty is tested. A redress decision is the firm saying what it owes, and that statement needs someone accountable behind it.
The specific limits are set out in FCA complaints handling: what an AI agent may do.
Where does the line sit?
| Task | The agent | Depth | Stays with a person |
|---|---|---|---|
| Recognise a complaint in an incoming message | Flags and opens a case | 1 | Reviewing borderline messages |
| Acknowledge receipt | Sends approved wording | 3 | Changes to the wording |
| Track deadlines | Sets dates and reminds | 1 | Acting before they pass |
| Gather the facts | Pulls account history, call notes, terms in force | 1 | Deciding what the facts show |
| Draft the investigation summary | Drafts with sources | 2 | Accepting or correcting it |
| Decide the outcome | Not permitted | Complaint handler | |
| Decide redress | Not permitted | Handler, within their authority | |
| Draft the final response | Drafts for review | 2 | Approving and sending |
Is recognising a complaint really easy?
It is the hardest of the easy parts, and the one most worth testing. Many complaints do not use the word. A customer who writes "this is the third time I've had to chase this and I'm losing money" has complained, even if the message ends with a question about a payment date. An agent tuned to answer questions may answer the question and miss the complaint.
Under-recognition is the costly direction. A missed complaint never starts its clock, never gets investigated and never appears in root-cause analysis. Certification for this capability should include plenty of indirect complaints, labelled as complaints, and score the agent on catching them. Over-recognition costs handler time; under-recognition costs customers their rights.
What about the complaints the agent itself causes?
Some complaints will be about the agent: a wrong answer, an unhelpful refusal, a promise it should not have made. Those cases need particular care, because the agent has an obvious conflict in handling them. They should go straight to a person.
They are also some of the most useful information a firm has. As argued in complaints about the agent are the most useful complaints, each one points to a specific gap in the agent's sources, bounds or brief, and should feed directly into its next certification.
What should the record show?
For each complaint: when it arrived, when it was recognised and by whom or what, every deadline and whether it was met, the facts gathered with their sources, the person who decided the outcome and redress, the reasoning, and the final response. Where the agent drafted the summary or response, the record should show the draft and the person's changes.
That last part helps with quality assurance. If handlers routinely accept the agent's investigation summaries without change, either the agent is very good or the handlers are not reading closely. Sampling the cases where nothing was changed tells you which.
What should you do differently?
Hand the administration to an agent and hold it to a high standard on recognition, timeliness and wording. Keep outcomes and redress firmly with handlers, and use the time saved to improve the quality of those decisions rather than to cut the team. Complaints are where customers test whether the firm is fair, and that test is answered by people.
The same principle runs through all compliance work with agents: AI compliance agents are second line, never the last word.
Questions people ask
Can an agent resolve simple complaints by itself?
It can resolve the underlying issue where that issue is routine and within its bounds, such as correcting an address. Whether the complaint was upheld, and whether redress is due, is still a decision a handler should confirm.
Should customers know an agent acknowledged their complaint?
Yes. Be clear that an AI agent has logged and acknowledged it and that a person will investigate. Customers react badly to discovering later that the reply they took as personal was automatic.
What deadline should the agent work to?
Whatever the rules in your market and your own policy set. Configure them as the firm's own limits and have a person check the configuration, rather than letting the agent infer deadlines from regulatory text.