How to automate a shared company inbox: rules, a helpdesk, Make or AI?

A hundred messages a day land in info@, orders@ or claims@, and someone has to read them, sort them, forward them and retype data from half of them into a system. Email automation can be done in four quite different ways — from Outlook rules you already pay nothing for, to AI that understands what a message says. Which one is right depends on a single question: how do you know where a message belongs?

The short answer: if you can tell by sender or subject, Outlook or Gmail rules are enough. If a team answers customers’ questions, you need a helpdesk. If messages come from a few systems in a fixed format and you need to move data out of them, Make or n8n will do. And if the content decides what a message is and there are hundreds a day, it is time for AI sorting with a person in control.

The rest of this article covers where each route breaks, what to watch out for and what to prepare.

What email automation actually means

Four jobs of email automation: sort, route, extract data, draft a reply

“Automating the inbox” covers four different jobs, and it pays to keep them apart:

Sorting — is this an order, a claim, an invoice, an enquiry, a question or spam, and how urgent is it?
Routing — getting the message to the right person or department.
Extracting data — order, invoice or claim number, amount, customer — and writing it into the ERP, CRM or a spreadsheet.
Replying — acknowledging receipt, answering a routine question, asking for a missing detail.

Most companies start with the fourth, because it is the most visible. But most of the time is usually lost in the first three.

Route 1: Outlook and Gmail rules

Rules (filters in Gmail) move and label messages by sender, subject or keyword. They are free, reliable, and any IT admin can set them up in an afternoon.

They break wherever the content decides the type. A customer writes “the goods arrived damaged” without ever using the word claim. An order arrives as an attachment with the subject “RE: RE: today’s”. An invoice comes from a supplier’s new address. A rule has no chance — and the more exceptions you add, the harder it becomes for anyone to follow.

When it is enough: you can tell messages apart by sender or subject, or you have a few regular partners who always send the same way.

Route 2: a helpdesk

A helpdesk (Freshdesk, Zendesk, Help Scout, Daktela) turns emails into tickets with a status, an owner, a history and reply templates. For a team that answers customers, it is often the best first investment — and many tools now have their own AI features for drafting replies.

Where it falls short: a helpdesk is built for conversations with customers, not for processing documents. It will not move a supplier invoice or an order spreadsheet into your ERP. And if the inbox also carries quotes, partner enquiries and internal matters, you will not want those as tickets.

When it is enough: the inbox is primarily customer support. If the problem is mostly replying, have a look at customer support automation.

Route 3: Make, Zapier or n8n

Integration tools can react to a new email: pull out the attachment, save it to a drive, add a row to a spreadsheet, post a notification to Teams. These days you can also add a “classify with AI” step to the scenario.

For a pilot, that is excellent. The trouble starts with exceptions: two topics in one email, a correction sent an hour later, an attachment you have to open to know what the message is about. Every exception is another branch, and eventually nobody wants to maintain the scenario. On top of that, the content of your emails flows through yet another third party, which has to be reflected in your GDPR records.

When it is enough: dozens of messages a day, clear types, and you want to test the idea cheaply first.

Route 4: AI email sorting

Processing emails with AI differs from the previous routes in that it does not look for a pattern — it reads the content, body and attachments alike. For every message it does what a person does today:

sets the type and urgency according to your categories,
routes it to a department or a named person, following rules only you know,
extracts the data and writes it into your system,
drafts a reply to routine questions,
and hands to a person anything it cannot classify with enough confidence.

That last point matters most. A system that is forced to decide every time will start guessing. A good one says “I don’t know” for 10–20% of messages, and those go to manual review.

People keep working in their own Outlook or Gmail — they just find the messages already sorted, with the data filled in and a draft reply waiting. We describe how this works in practice on the page email automation and shared inbox triage.

When it makes sense: a hundred or more messages a day, many types, frequent retyping of data into systems, or an inbox that money flows through — orders and invoices.

Side-by-side comparison

	Rules	Helpdesk	Make / n8n	AI sorting
Sorts by content	no	partly	with an AI step	yes
Extracts data from attachments	no	no	fixed formats only	yes
Draft replies	no	yes	limited	yes
Handles exceptions	no	a person does	with difficulty	hands to a person
Cost	free	per-seat licence	low	project + operations

When AI is not worth it

To be honest: in most small companies the first or second route is enough. AI does not pay off if:

the inbox gets a few dozen messages a day and sorting takes half an hour,
you can tell messages apart by sender — well-configured rules will do the same for free,
the problem is not sorting but that nobody has time to reply — that is solved by organising the work, not by software.

If you do not even have rules today, start there. You will see what is left over, and that becomes a precise brief.

Risks to watch out for

Sensitive emails do not go out without human approval

Automatic email replies without review. The most common mistake, and the most expensive. One reply with the wrong price, or a promise the company cannot keep, is enough. Start in “AI drafts, a person sends” mode and enable automatic sending only for narrow message types — acknowledging receipt of an invoice, say — and only once measurements on real emails back it up.

GDPR. Emails are full of personal data. You need to know where they are processed (EU region, a data processing agreement, a ban on training on your data), who has access and how long they are kept. For sensitive data, a smaller model running on your premises is an option.

Overly broad access. The system should access only the mailboxes it processes — not every mailbox in the company, and certainly not management’s personal ones.

Sensitive topics. Claims, legal matters and payments should always go to a person, however confident the AI is.

What to prepare

A list of the message types you distinguish today and who handles each. A simple table is enough.
A sample of 200–500 real emails from recent months, including the ugly ones — accuracy is measured against it.
Access to the mailbox (Microsoft 365, Google Workspace or IMAP) and to the systems the data should go into.
A person to review the set-aside messages who has time for it during the working day.

If it is mainly about orders or invoices, see also our articles on orders from email and invoice processing — those are specialised cases that go all the way into the ERP.

What it costs

Indicatively, excluding VAT, for a small or medium company: inbox analysis and sample €900–1,800, pilot on real emails €3,500–7,000, production rollout connected to your systems €5,000–12,000, operations from €390 a month plus model usage (tens of euros at normal volumes). More detail in how much an AI solution costs.

In a model case where one person spends two hours a day sorting the inbox, that is roughly 40 hours a month. Add the retyping of data from attachments and, at hundreds of messages a day, payback is measured in months rather than years. At a few dozen messages it never pays back — and in that case we will tell you.

Summary

You can tell messages apart by sender or subject: Outlook or Gmail rules.
The inbox is mainly customer support: a helpdesk, possibly with draft replies.
Dozens of messages a day and you want to test cheaply: Make or n8n, knowing you will hit a wall on exceptions.
Hundreds of messages a day, the content decides the type, and data gets retyped into systems: AI sorting that hands uncertain cases to a person and sends nothing without review.

How many messages a day arrive in your shared inboxes, and who sorts them today? If you can answer that, you have a number to start from. We are happy to work it through with you — and look at what could be automated in your company.