Write down what "low risk" means as tests a case must pass, not a judgement the agent makes. The agent closes only cases that meet every test, with nothing owed, no complaint and no sign of vulnerability. It records why each case qualified, the customer can reopen it easily, and a person samples closures and owns the criteria.
What does closing a case mean, and who owns it?
Closing a case declares that nothing more is owed: no reply, no refund, no follow-up. Each closure on its own is small and usually reversible, because a case can be reopened. The risk is in what closure hides. A complaint buried in a long thread, a customer who said they were struggling, a promised call-back nobody made: closing the case removes it from every queue a person looks at. This is the reversal question at the centre of autonomy in regulated industries: a closure can only be undone if someone notices it should be.
The operations manager for the queue owns the closure criteria. The complaints lead owns the definition of a complaint, which the agent must apply before closing anything. A quality reviewer samples closed cases. Each of these is a named person, not a team mailbox.
For a narrower view of the same question in customer service, see whether an AI agent can close a support ticket on its own.
How do you set it up, step by step?
- Person, operations manager: writes closure criteria as tests. For example: the customer's question was answered; the customer confirmed, or did not reply within the set period after the answer; no money, document or call-back is outstanding; the case type is on the approved list. Evidence: versioned criteria.
- Person, complaints lead: defines complaint signals. Expressions of dissatisfaction, requests to escalate, mentions of a regulator or ombudsman. A case with any of these is never closed by the agent. Evidence: the complaint definition.
- Person, vulnerability lead: defines vulnerability signals. Bereavement, illness, financial difficulty, confusion. Same rule: these cases go to a person. Evidence: the signal list.
- Agent, depth 1: replays past closed cases. For each, it says whether it would have closed it and why. Compare against what happened. The method is set out in scoring an agent against your own filed cases. Evidence: a scored run.
- Agent, depth 2: proposes closures on live cases. A person clicks close or reopens. Evidence: proposals and overrides.
- Person, operations manager: approves the move to depth 3 for named case types, once overrides are rare and understood. Evidence: a signed decision with an expiry.
- Agent, depth 3: closes qualifying cases. It writes a closure note stating which tests passed, and sends the customer a message saying the case is closed and how to reopen it. Evidence: closure note per case.
- Person, quality reviewer: samples closures weekly. Evidence: sample findings.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Criteria | Nothing | none | Operations manager writes tests |
| Complaint and vulnerability signals | Nothing | none | Complaints and vulnerability leads define them |
| Replay | Judges past cases | 1 | Owner compares and reviews failures |
| Propose | Recommends closure with reasons | 2 | Agent handler closes or keeps open |
| Close | Closes qualifying cases, notifies customer | 3 | Any case with a complaint or vulnerability signal |
| Sample | Nothing | none | Quality reviewer checks closures |
Which decisions does a named person keep?
- The criteria and the approved case types. The agent does not add case types because they look similar.
- Every case with a complaint signal. Closing a complaint without handling it is a conduct failure, not an efficiency gain.
- Every case with a vulnerability signal.
- Every case with something outstanding, however small: a refund, a document, a promised call.
- Moving a case type to unattended closure, and taking it back.
Closing a claim or an expense is a different act: it may settle money. See whether an AI agent can close an insurance claim and how to let an agent approve low-value expense claims for those.
Which controls matter, and what record do they leave?
Criteria enforced as bounds. The platform checks the case type and the outstanding-items flag at the moment of closure, not only the agent's reasoning. Evidence: verdict citing the criteria version.
Signal screening before closure. Evidence: each closure note records that complaint and vulnerability screening ran and found nothing.
Customer notice with a reopen route. Closure becomes reversible in practice only if the customer knows how to reopen. Evidence: the message sent, linked to the case.
Closure note. Which tests passed, which facts were used. Evidence: one note per closure.
Sampling. Evidence: findings and actions taken.
What should you measure?
- Reopen rate: closed cases reopened by the customer, compared with cases closed by people. This is the single most useful number.
- Complaints that follow a closure.
- Sample findings, grouped by which test was wrongly passed.
- Cases the agent declined to close, and why. If it declines most, the criteria may be too narrow to be worth it.
- Time cases spend open after the last real action.
Where does it usually go wrong?
"Low risk" is left to the model. The agent closes what feels finished. Write the tests.
Silence is read as satisfaction. A customer who stopped replying may have given up. Set a waiting period, and treat a second contact on the same issue as a reopen, not a new case.
Complaint wording is narrow. Customers rarely write "I wish to complain". Define signals broadly and review missed ones.
Closure messages discourage reopening. A message that sounds final reduces reopens and hides problems.
Bulk closure of a backlog. An agent is pointed at an old queue and closes hundreds of cases in an hour. Volume limits apply here as much as anywhere.
Questions people ask
What waiting period should apply before closing an unanswered case?
Set it by case type and customer expectation, and write it into the criteria. Longer for anything involving money or documents, shorter for simple information requests.
Should closures by the agent be labelled?
Yes, internally. Every closure note should show that the agent closed it and under which criteria version, so reviewers can find agent closures quickly and trace any error to the rule that allowed it.
Can the agent reopen cases too?
It should be able to reopen on any new contact or signal, at depth 3. Reopening is the safe direction, and making it easy reduces the cost of a wrong closure.