A mailbox everyone can see and nobody owns
A shared address like info@, orders@ or support@ usually starts as a practical idea: several people can see it, so nothing gets lost. Over time it becomes a place where enquiries, orders, supplier invoices, complaints, newsletters and spam all pile up together. Everyone opens it, everyone assumes someone else is replying, and an important message can sit for a week among a hundred others.
A typical picture: the owner goes through the inbox in the morning, answers whatever catches the eye and leaves the rest for later. The bookkeeper searches for an invoice that arrived at info@ instead of her own address. A customer’s complaint sits unnoticed beneath a cold sales pitch. This isn’t carelessness. The inbox simply has no structure, and human attention is the only filter you have.
This article is about the internal workings of that inbox: sorting messages, getting them to the right person and not forgetting them. It is not about a chatbot that talks to customers instead of you — we covered that in the piece on AI agents on the front line.
Sorting, extracting details, routing
An AI model can read a message and say what it is: an enquiry, an order, a complaint, an invoice, spam or something else. It copes with emails that have no subject line, were written in a hurry, lack diacritics or raise two topics at once. That is exactly where classic rules such as “if the subject contains invoice” fall apart, because people write however they feel like writing.
The second step is pulling out the essentials: who is writing, which product or order it concerns, what document number they quote, by when they need it, what amount is mentioned. These details land in a tidy card or a row in a spreadsheet, so nobody has to read the whole thread to see what it is about.
The third step is routing. Invoices go to the bookkeeper, complaints to the operations manager, enquiries to the salesperson, spam to quarantine. When the model is unsure, the message goes into an “unclear” pile where a person decides. Simply giving every message an owner solves most of what is wrong with a shared inbox.
Draft replies a human reads and sends
Once a message is sorted and its details extracted, AI can prepare a draft reply: thank the sender for the enquiry, ask for missing information, confirm an order was received, point to the complaints procedure. The draft appears next to the message, and the responsible person reads it, edits it and sends it. You save the blank-page writing; the decision stays with a human.
That should be the default, at least for the first few months. Only when you see a certain type of draft go out unchanged week after week might you consider sending it automatically — and only for simple things, such as confirming that a message was received.
Some things should never go out automatically: replies to complaints, price quotes or promises of a discount, anything involving payments or bank details, legal or contractual matters, a reply to someone who is clearly upset, and anything where the model is unsure. One wrong message sent in the company’s name cannot be taken back.
Reminders when something sits unanswered
The biggest benefit is often not what gets solved but what no longer gets lost. The system remembers when a message arrived and who owns it. If an enquiry sits for two days with no reply, the owner gets a reminder; if nobody reacts after that, the alert goes to their deputy or to the head of the company.
Deadlines can be set by type: an hour or two during working hours for an enquiry, end of day for a complaint, enough days before the due date for an invoice. The reminder can arrive in a tool the team already uses, for example as a chat message or a short morning summary.
One warning: people quickly learn to ignore too many alerts. A few that mean “this one really is urgent” beat a daily list of fifty items. In practice we usually start with two or three deadlines and adjust to what the team actually uses.
Privacy: what goes where and who sees it
Emails contain customers’ personal data, prices, invoices and sometimes sensitive matters. Before you deploy anything, settle three things. First: which service reads the message, and where does it process it? With AI services you can often choose an EU processing region, and you should check the terms on whether your data is used to train models. Business plans often exclude this, but verify it in the contract rather than trusting an impression.
Second: not everything has to go to AI. Attachments with sensitive content, messages from certain addresses or mailboxes full of personal data can be excluded from sorting. Often it is enough for the AI to see the message text rather than the whole archive.
Third: access rights. The tool should reach only the mailboxes it needs, with the smallest rights possible — read and prepare a draft, not delete. Keep a record of who sent what and decide who may change the rules. If you process personal data, a data processing agreement with the provider and an entry in your data-protection records belong here too; when in doubt, talk to a lawyer.
When this is overkill
If twenty messages a day land in your inbox, you do not need AI for sorting. A simple rule will do more: who reads what, two or three filters, and an agreement that every message gets a reply within a day. Setting up and supervising AI has overhead, and at low volume it never pays back.
It starts to make sense when messages arrive in the dozens each day, several people take turns in the inbox, or something important has already slipped through. Another signal is when someone spends an hour a day sorting and forwarding messages.
Even then, AI does not replace agreed responsibilities. If it is unclear who handles complaints, even perfect sorting will not handle them. First decide who owns what, then automate it. If you would like to judge whether it makes sense for your inbox, we are happy to look at it with you — and to say so if the answer is “not yet”.