Most businesses have one mailbox that everything runs into: the address from the website footer, the contact form, purchasing, accounts. In the morning somebody opens it and sorts by hand what is an enquiry, what is an invoice, what is an appointment request and what is advertising.
Per message that costs little time. Across the day it costs a lot. And there is a second price: until somebody has sorted the mail, every message sits there equally long, the urgent one and the optional one.
Can AI route incoming email to the right person automatically?
Yes, with the sorting being the part that works reliably today: a language model reads the text, assigns it to a category and states how certain it is. The routing itself is then a setting you decide on, not a capability of the model.
The distinction matters because the errors land in different places. A message sorted wrongly that stays visible in the shared inbox costs one click. A message routed wrongly that disappears from the shared inbox can cost an order. So the sorting starts out as a suggestion.
What is email classification?
Email classification is the automatic assignment of an incoming message to categories defined in advance, together with a value for how certain that assignment is.
The categories come from your business, not from the tool. Three to eight are enough, and each one needs a person and a deadline. Set up fourteen and you get fourteen folders that are opened at fourteen different rates.
When are plain mailbox rules enough?
As long as a sender or a word in the subject line repeats reliably. Rules cost nothing and everyone on the team understands them, which makes them the first attempt. Their limits are documented by the vendors and arrive sooner than people expect.
- According to Microsoft, inbox rules in Exchange Online are capped at 256 kilobytes per mailbox. Once that is reached, the next rule cannot be saved.
- According to the same source, one rule may redirect to at most ten recipients, and a message is redirected only once.
- According to Microsoft, automatic forwarding to external addresses has been off by default for new organisations since 2021. Rules that redirect outside then do not take effect.
- According to Microsoft, a shared mailbox in Microsoft 365 is intended for at most 25 users; if more of them work in it at the same time, connection failures and duplicated messages can follow.
- In Google Workspace, according to the administrator help, a limit of 1,000 settings and a combined 5 megabytes per account applies. Beyond that, Gmail saves no further setting.
Above all, a rule only answers whether a word occurs. It does not answer what the message is about. A message with the subject line Quick question can be a complaint, a change of appointment or a payment reminder, and no filter tells those apart.
Sorting only beats the person doing it in the morning once every category has a named owner and a deadline attached. Without that, AI merely moves mail faster into folders nobody opens.
How do you introduce it?
In four steps, and the first two need no technology at all.
- 1.Keep a tally for two weeks: how many messages arrive, which categories they fall into, how long they sit. That is the figure everything later gets measured against.
- 2.Fix the categories and their owners, eight at most, each with a deadline for the first response.
- 3.Shadow mode: the sorting runs along and writes its suggestion into the message without moving anything. As a guide two to four weeks, depending on daily mail volume.
- 4.Only then deliver automatically, and only for the categories that were right in shadow mode. The rest stays a suggestion.
A residual basket belongs to the design, because every classification produces cases that fit no category. That basket needs a name on it, otherwise it grows quietly. How we plan such work, bound it and connect it to existing systems is described on our page about AI automation.
What has to be settled legally first?
Three points, and none of them is technical.
- Monitoring: according to WKO, introducing control measures and technical systems for monitoring employees requires the consent of the works council under the Austrian Labour Relations Act, as far as those measures affect human dignity; without a works council, the consent of the affected employees is needed. Whether sorting mail falls under this depends on what happens with the evaluations, and that is a legal question.
- Private messages: according to WKO, an employer may generally not simply access private mail, because of the secrecy of correspondence and data protection law. Where private use is allowed, settle in advance what the sorting does with such messages.
- Disclosure: once sorting turns into an automatic reply going outside, the transparency duties of the AI Act come with it. Article 50 has been applicable since 2 August 2026 and provides that people be informed when they are dealing with an AI system.
A fourth point concerns job applications and the allocation of work. Annex III of the AI Act lists AI systems for evaluating job applications and for allocating tasks based on individual behaviour or personal traits among the high risk uses; under Regulation (EU) 2026/1744 the duties attached to them apply from 2 December 2027. Which back office work is worth automating in the first place is covered in a separate article, and what has to be documented when AI is in use is on our page about GDPR and the AI Act.
According to Statistics Austria, 30 percent of Austrian companies with ten or more employees used AI technologies in 2025, against 20 percent across the EU. The inbox is where the benefit shows up fastest, because the volume is known and every error stays visible. So start with two weeks of tallying rather than with a tool.