Using AI to Sort a Shared Customer Inbox
By CodexierPublished 6 min read
The info@ inbox is where a small company's customer service, sales leads, supplier questions and spam all land together, and where someone loses the first hour of every day sorting it. AI can take the sorting, the routing and the first draft of the reply. What it should not take, at least in the first year, is the send button. This guide shows how to set that up in Outlook or Gmail and what to measure after a month.
Audit a week of incoming email
Before touching any tool, scroll through five working days of the shared inbox and put each mail in a bucket: order status, quote request, invoice question, complaint, supplier, newsletter, spam. Note how long each type took to answer and who answered it. This is the specification. Companies that skip it end up with categories the AI invented, which nobody on the team recognises.
Categories and routing rules
Keep the category list short: six to eight is plenty. Each category gets a destination and a rule for urgency. The model reads the subject, body and sender domain and assigns one label; a simple rule engine then acts on the label.
| Category | Routed to | Draft reply? | Urgency rule |
|---|---|---|---|
| Order or delivery status | Customer service queue | Yes, from order data | Same day |
| Quote or new enquiry | Sales, and the CRM as a lead | Yes, acknowledgement plus questions | Within two hours |
| Invoice or payment | Finance | Yes, standard replies only | Next working day |
| Complaint or dispute | Named manager | No, human writes it | Flagged immediately |
| Supplier or partner | Purchasing or owner | No | Weekly review |
Anything the model is unsure about goes to an Unsorted label that a human empties daily. An honest unsure bucket beats a confident wrong label.
Drafts, not auto-sends
A language model writes a plausible answer even when the underlying facts are wrong: a delivery date it guessed, a price from an old list, a promise your terms do not allow. With a person on send, that costs thirty seconds of review. With auto-send, it costs a customer. The mechanism is the same reason we keep humans on send for every workflow automation that talks to customers in the first months.
- Give the model your real sources: order system, price list, delivery terms, opening hours. Drafts grounded in data need far fewer corrections.
- Draft in the sender's language. Swedish enquiries get Swedish drafts; English gets English.
- Include a short reason line in the draft header: why this category, which source was used. Reviewers catch errors faster.
Tools that plug into Outlook and Gmail
Built-in assistants
Copilot in Outlook and Gemini in Gmail summarise and draft for one user. They do not run rules on a shared mailbox for a team, which is what triage needs. See our Copilot vs ChatGPT comparison.
Help desk tools with AI
Shared-inbox products (Front, Freshdesk, Zendesk, HubSpot inbox) add classification and drafts on top of ticketing. Good when you already want a ticket system.
Automation platforms
Make, Zapier or n8n with a Microsoft Graph or Gmail connector plus a model call. Flexible, cheap to run, and yours to change.
A small custom workflow
For Swedish data-handling requirements or an odd mail setup, a purpose-built workflow with an EU-hosted model is often simpler than bending a platform.
Whichever route you pick, check where the mail is processed and get a data processing agreement, because customer mail is personal data under GDPR. If you are unsure which route fits your volume, a 15-minute call is usually enough to say.
What to measure after a month
Four numbers, taken from the mail system itself, tell you whether to continue, tune or stop.
- Time to first reply per category, compared with the audit week. This is the customer-facing result.
- Share of drafts sent with no or minor edits. Below half means the prompt or the sources need work.
- Mails that landed in the wrong queue, counted by the people who received them. Above a handful a week means the categories are unclear.
- Minutes per day the inbox owner spends on sorting. This is the internal result, and the one that pays for the setup.
Decision rule: if reply time and sorting minutes both improved, keep it and add one category to drafting. If only one improved, tune for another month. If neither did, switch it off and go back to the audit; the categories were probably wrong.
When you do not need this
If the shared inbox gets a couple of dozen mails a day and one person already answers them well before lunch, a filter rule in Outlook is enough and an AI layer is a toy. If most mail is complaints or negotiation, AI drafting adds risk without saving much, since a human writes those anyway. And if your team does not yet have an agreed reply standard, fix that first: the AI will faithfully reproduce whatever inconsistency it learns from. Triage pays when volume is high, categories are repetitive and the data the answers need is already in a system.
Frequently asked questions
Can the AI read our customer mail under GDPR?
Yes, provided the processor is contracted with a data processing agreement, the purpose is documented in your records of processing, and the model provider does not train on your data. Mail from EU customers processed on an EU-hosted model keeps the review simplest. Tell your privacy notice that automated sorting is used.
Will it work in Swedish?
Current language models classify and draft in Swedish well. The weak spot is domain vocabulary: your product names, your delivery terms, your abbreviations. Give the model a short glossary and a few example replies and quality improves noticeably.
Should we let it reply automatically to simple questions?
Not in the first months. Run drafts with a human on send, log the edits, and only consider auto-send for a category where drafts have gone out unchanged for weeks and a wrong answer is cheap to correct, such as opening hours. Complaints and anything involving money stay with a person.
Want to know if your inbox is worth automating?
Bring the categories from your audit week to a short call. We tell you which ones AI can sort and draft reliably, which should stay manual, and what the setup would cost as a fixed price.
Book a free 15-minute call