AI email triage: what the agent does
By Philipp Kant 4 min read
AI email triage is an agent that reads every incoming business email, decides what it is, extracts the data it carries, routes it into the system that owns the work (CRM, ERP, support desk, accounting), and drafts the reply. A fixed rule decides which mails a person sees first.
The model is the smallest part of it. The value sits in the extraction, the routing, and the gate.
A triage agent does six things with every email
Every mail runs through the same steps, in the same order:
- Read. Subject, body, signature, quoted thread, attachment names.
- Classify. Intent (order, quote request, complaint, invoice, cancellation, data request, spam), urgency, and the sender’s mood.
- Extract. The values the next system needs: customer and order numbers, amounts, dates, products, addresses.
- Route. The system and queue that own the follow-up.
- Draft. A reply in the sender’s language that promises nothing the email does not support.
- Gate. Act alone, or hand to a person first.
You can watch all six run on your own example in our AI Inbox Triage tool.
Extraction decides the value, not classification
Telling an order from a complaint is the easy part. Current models get that right almost every time. The money is in the fields: an order that lands in the ERP with the right customer number, quantity, and delivery date saves the five minutes someone spends copying it over.
Two rules keep extraction safe:
- Only literal values. The agent takes what is written in the mail, never what it infers. “Same as last time” is not an address.
- A confidence per field. Each value carries a 0 to 100 score. Low confidence on a field that matters (an amount, an IBAN, an order number) sends the mail to a person instead of into the system.
This is where most inbox automations fail in practice. The input stays free text, and a model guessing at fields only moves bad data faster. That is the argument for structured intake as the first automation step: the cleaner the input, the less the agent has to guess.
Routing means writing into the system that owns the work
A triage agent that only adds a label has done half the job. Routing means the follow-up starts in the right place:
| Intent | Owning system | First action |
|---|---|---|
| Order | ERP | Create the order draft with the extracted lines |
| Quote request | CRM | Create the deal, attach the request, assign sales |
| Complaint | Support desk | Open a ticket linked to the order |
| Invoice, payment reminder | Accounting | Match against open items |
| Job application | HR | File it, send the acknowledgement |
| Spam | None | Archive |
Writing into CRM and ERP is ordinary integration work, with all of its usual cost: field mapping, idempotency, error handling. We broke those costs down in the cost behind the button.
The human gate is a rule, not a model judgment
The model may suggest that a mail needs review. It never gets to decide that one does not. The gate is plain code, and any of these sends a mail to a person:
- Critical urgency: deadlines within a day, outages, legal threats.
- An angry sender. A fast, wrong automated answer costs more than a slow right one.
- Any risk sign: changed bank details, payment pressure, a GDPR data request, phishing patterns.
- Sensitive intents: complaints, cancellations, applications, data requests, every time.
- Shaky extraction: average field confidence below 70.
The changed-bank-details mail is the reason this rule exists. It reads like an ordinary invoice, it is often urgent, and paying it is irreversible. No model gets to wave it through.
What to automate first
Start where volume meets structure:
- Orders and reorders from known customers: high volume, clear fields, measurable time saved.
- Quote requests: speed decides who wins them, and the first reply can go out in minutes.
- Invoices and reminders: matching is mechanical once the fields are clean.
Leave complaints and anything legal on the human side of the gate, with the agent preparing the case: summary, order history, a draft. Measure two numbers from day one: the share of mails the agent handles alone, and how often a person changes what it prepared.
FAQ
How does AI email triage work? The agent reads each incoming mail, classifies intent and urgency, extracts the fields the next system needs, routes the mail into CRM, ERP, support, or accounting, drafts a reply, and applies a fixed rule that decides whether a person reviews first.
Can AI answer business emails automatically? Routine ones, yes: order confirmations, quote acknowledgements, status replies. Complaints, cancellations, legal and payment topics should go to a person, with the agent’s draft ready.
When should an AI email agent hand off to a human? On critical urgency, angry senders, any risk sign such as changed bank details, sensitive intents like complaints or data requests, and whenever its extraction confidence is low.
Is AI email triage GDPR-compliant? It can be: process mail in the EU or with a provider under a data processing agreement, store only what the workflow needs, and log what the agent did. The design decides compliance, not the model.
If your team still copies orders and requests out of the inbox by hand, that is exactly the agent we build.
