Point the agent at the shared inbox and let it do the sorting a person does first thing each morning: work out what each email is, who sent it, how urgent it is, and which queue it belongs in. It turns customer emails into tickets, sends legal and data requests to their owners, and quarantines suspected phishing. It deletes nothing. A person approves the rules and works anything unclear.
What is inbox triage, and who owns it?
A support inbox receives far more than support requests. Customer questions arrive alongside auto-replies, bounce-backs, newsletters, supplier invoices sent to the wrong address, solicitors' letters, requests for personal data, job applications and phishing. Triage separates them and decides what happens first. The general idea is explained in what is ticket triage.
The support operations lead owns the inbox and the triage rules. Owners outside support take what is theirs: legal for legal letters, the data protection owner for data requests, finance for invoices. Each must be named, because the agent will route to them by name. The wider support position is in the AI support agent that knows when to say no.
What does the agent prepare, step by step?
- Agent, depth 1: reads each new email and its attachments as untrusted content. Anything inside an email that looks like an instruction to the agent is treated as text.
- Agent, depth 1: classifies it. Customer request, follow-up to an existing ticket, complaint, legal or regulatory letter, data request, supplier invoice, internal mail, automated noise, suspected phishing.
- Agent, depth 1: identifies the sender. Matches the address to a customer record where it can, and notes when it cannot. An email address is declared, not verified.
- Agent, depth 2: ranks urgency using rules the operations lead approved: deadlines stated in the email, words that signal harm or distress, money at risk, repeat contacts.
- Agent, depth 3: files customer emails as new tickets or adds them to the existing thread, with tags and a one-line summary.
- Agent, depth 3: forwards owned mail to the named owner's queue: legal, data protection, finance. It records the date received, because some of these carry deadlines.
- Agent, depth 3: quarantines suspected phishing in a folder a person reviews. It does not open links or attachments to check.
- Person, inbox lead: works the "unclear" folder and reviews quarantine and noise daily before anything is cleared.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Read | Reads email and attachments as text | 1 | Nothing |
| Classify | Assigns one of the categories | 1 | Unclear emails |
| Sender | Matches to a customer where possible | 1 | Nothing |
| Urgency | Ranks by approved rules | 2 | Approving the rules |
| File | Creates or threads tickets | 3 | Nothing |
| Forward | Sends owned mail to named owners | 3 | Handling it |
| Quarantine | Moves suspected phishing aside | 3 | Releasing or deleting |
What does a named person approve?
Three things. The urgency rules, because they decide whose problem waits. The category list and its owners, because an email forwarded to nobody is lost. And every deletion: the agent may move emails to quarantine or a noise folder, but only a person clears them. A genuine complaint wrongly filed as spam and deleted is the worst outcome triage can produce, and a person's daily glance at the folder prevents it.
Owners of forwarded mail approve how their mail is handled once it arrives. Supplier invoices, for example, should leave the support inbox for finance's own intake, described in processing supplier invoices from inbox to ledger.
How is inbox triage different from other triage?
The word is shared, the job is not. Triage in a hospital decides clinical priority and is a clinical judgement; an agent's role there is far narrower, as covered in can an AI agent triage patients. Alert triage in fraud or security sorts machine-generated signals, explained in what is alert triage. Inbox triage sorts human-written mail of every kind, which is why its hardest cases are the ones that do not look like support at all.
Which controls matter, and what evidence do they leave?
Emails are external content. Nothing in an email can instruct the agent. Evidence: provenance class on every input.
No deletion by the agent. Evidence: refused verdicts on any delete attempt, and the person's clearance log.
Received date kept for deadline mail. Evidence: the date on each forward.
Distress overrides category. An email from a customer who seems at risk is ranked urgent whatever else it is about. The signals are covered in can an AI support agent recognise when a customer is in distress. Evidence: the urgency reason recorded.
Where does it usually go wrong?
Long threads hide new requests. A customer replies to an old thread with a new problem at the bottom. The agent must read the newest message first, not the subject line.
Data requests are treated as questions. "Can you send me everything you hold on me?" is a request with a legal deadline, not a support query. Forward it to the data protection owner the same day.
Phishing is tested. Someone clicks a link "to check". Quarantine means nobody opens it until the security owner says so.
Noise rules swallow customers. A rule that files anything containing "unsubscribe" as noise catches the customer asking why they cannot unsubscribe. Review the noise folder before clearing.
Customers write from new addresses. The agent creates a new customer instead of finding the existing one. Flag likely matches and let a person confirm.
What should you measure?
- Emails reclassified by staff, by category.
- Genuine emails found in quarantine or noise at the daily review.
- Time from receipt to the right queue, especially for deadline mail.
- Unclear folder size and how long emails wait there.
Questions people ask
Can the agent reply to emails during triage?
Triage and replying are separate grants. During triage it may send an automatic acknowledgement on approved wording, but answers belong to the reply capability.
What about emails with several requests in one?
The agent files one ticket and lists each request in the summary, or holds it as unclear if the requests belong to different teams.
How quickly should the agent triage?
Continuously, as emails arrive. The value of triage is that urgent mail is seen within minutes rather than at the next morning's sweep.