AI for customer email: triage, drafts and where a human stays in the loop
The shared inbox is the best first place for AI. How triage, extraction and drafting work together, and the three points where a person must decide.
The short answer
The shared inbox is where a small company loses the most time to reading and retyping, and it is where AI helps most with the least risk, provided one rule holds: the system prepares, a person sends. Three AI steps do the work. Classify each message. Extract the facts from it. Draft a reply from your own material. People approve, handle exceptions and review samples.
The three steps
| Step | What the model does | What it produces |
|---|---|---|
| Classify | Reads the message and decides what it is: quote request, support question, invoice, complaint, partner, spam, other. Also notes tone and urgency | A category, a priority, a route to the right person or queue |
| Extract | Pulls out the facts: name, company, order or reference number, what is being asked, deadlines mentioned | Structured fields, attached to the ticket or CRM record |
| Draft | Writes a first reply using your policies, pages and previous answers, in your tone | A draft in the inbox, marked as a draft, never sent |
Where a person stays in the loop
- Approving every customer-facing reply. The draft appears, a person reads it, edits if needed, sends. One click for the good ones. This is the rule that makes the system safe from week one.
- Handling uncertainty. When the model is not confident about the category or cannot draft from the available material, it says so and routes to a person without a draft. A system that never says “not sure” is not calibrated.
- Reviewing samples weekly. A short look at a random sample of classifications and drafts, plus every case where a person changed the category or rewrote the draft heavily. Those corrections improve the instructions and the source material.
What to measure
- Minutes from arrival to first human touch, per category. Triage collapses this from hours to minutes, because routing happens on arrival.
- Share of drafts sent without edits, per category. Rising over the first weeks as the source material improves; a category stuck low needs better material or should stay manual.
- Corrections per week. Falling. Each one is a lesson applied.
What this means for you
Map the categories your inbox actually receives, gather the material a good reply would draw on, and turn on classify, extract and draft with a person approving every send. In the first week nothing waits unnoticed; within a month most standard replies are one click. The rule that the system prepares and a person sends is what lets you start now instead of after a long pilot.
Frequently asked questions
Can the AI reply to customers by itself?
It can, and we advise against it for anything customer-facing at the start. Let it classify, extract and draft; let a person approve with one click. After a month of logs showing which categories are consistently right, unattended replies for a narrow, low-stakes category, such as opening hours, can be considered. Beyond that narrow category the case for going further is weak.
Where does the AI get the answers for its drafts?
From material you control: your pages, your policies, your previous replies, your product data. It is instructed to draft only from that material and to flag when it cannot. Drafts from general knowledge are how wrong information reaches customers.
What about angry or unusual emails?
Classification catches them: tone and topic are part of what the model reads. Complaints, legal matters, anything emotional or unusual is routed straight to a person with no draft, or with a draft clearly marked as needing care. The system should be better at spotting these than a rushed human skimming an inbox.
How long until it pays off?
Triage and routing return time in the first week, because nothing waits unnoticed. Drafting returns time once the source material is complete, typically within a month. Measure the two numbers above before and after; they make the case for the next process.