To automatically sort the email coming into a company, the order is this: first define the real categories by looking at a hundred actual emails, decide what action each category leads to (label, assign, reply with a template), start by solving the obvious with simple rules and reserve AI for the ambiguous, and always leave an exit to a human for what the system can't classify. Sorting isn't the goal: the goal is for each email to reach whoever has to handle it, fast.
The problem isn't the email, it's the shared inbox where everything lands mixed: an urgent order next to an invoice, a serious incident beside a rep trying to sell you toner. Someone has to open each one, work out what it's about and forward it. That manual sorting is slow, done reluctantly, and leaves important emails unseen for hours.
What do you need before you start?
Two things before touching anything: knowing which bins you want to sort the email into, and what happens when it lands in each.
The bins first. Take a hundred real emails from your inbox and group them by hand. You'll see almost everything falls into a few categories: orders, issues, billing, sales enquiries, spam. Don't invent twenty fine-grained categories: few, and told apart by what has to be done with each. If two categories end up in the same tray with the same person, they're the same category.
The action second. Sorting without acting saves nothing; it just adds pretty labels. Each category has to lead to something concrete: "issue" goes to the support person and flags priority; "invoice" goes to the admin folder; "sales" waits until someone has time; "spam" gets filed. If you don't know what action sits behind a category, that category isn't ready yet.
How do you do it, step by step?
With the categories and their actions clear, building is five steps.
- Define the real categories. The ones that came out of looking at a hundred emails. Few, distinct from each other, each with its action. This list is the skeleton of everything else.
- Decide what happens with each category. Label, assignment to a person, folder or template reply. The routing (who it reaches) matters as much as the label.
- Start with rules, not AI. A good share of email is sorted with simple rules: by known sender, by domain, by words in the subject. Solve everything a rule catches without hesitation first. It's cheap, transparent, and clears the bulk out of the way.
- Use AI only for the ambiguous, with confidence. What no rule catches —a client telling a problem in their own words— you pass to a classifier that reads the text and proposes a category with a confidence level. What it classifies confidently, through it goes; the doubtful isn't filed on its own: it's flagged for review.
- Always leave an exit to a human. Have a "don't know" category and a person who empties it every day. That valve is what prevents the disaster of an important misfiled email vanishing without anyone noticing.
Rules or artificial intelligence?
It's not one or the other: it's each for its own job. This table sums it up:
| Rules | AI (text classifier) | |
|---|---|---|
| Good for | Sender, domain, exact words | Understanding what free text is about |
| Cost | Very low, depends on no one | Higher; worth checking it pays off |
| Transparency | Total: you know why it sorted so | Lower; that's why confidence is needed |
| Risk | Misses nuance | Errs on the odd cases if not reviewed |
Most companies solve 70 or 80% of the email with rules and use AI only for the rest. Starting the other way round —AI for everything from day one— is expensive and opaque with no need.
What about email privacy?
A company's emails carry personal data, so the sorting has to meet GDPR. If external AI models are used, personal data is anonymised before sending it out. For sensitive material there are two paths that keep the email at home: sort with rules only, or use models that run inside your own infrastructure. It's a design decision, not a detail: worth being clear about before building anything, not after.
When is it NOT worth it?
Don't build this if:
- Little email comes in. If one person sorts the inbox in five minutes each morning, automating it costs more than it saves.
- Almost everything is the same type. If 95% of the email goes to the same person, there's nothing to route: sorting is redundant.
- There's no clear action behind it. Sorting to have labels no one uses is decorative work. A category with no action saves not a minute.
I say this because "sorting email with AI" sounds modern and sometimes gets built without asking whether it solves anything. It solves a lot when there's volume and several destinations; nothing when there aren't.
What mistakes do people usually make?
The usual three.
The first, too many categories. Fifteen ultra-fine bins get made and the system spends its day hesitating between two similar ones. Fewer, clearer categories get it right more often.
The second, sorting and not routing. The label gets applied but the email stays in the same shared inbox, so no one knows it was theirs. A label with no recipient moves nothing.
The third, leaving no exit to a human. The system auto-files what it doesn't understand, and that's exactly where the important email gets lost. Everything doubtful has to end up in front of a person, not in a folder no one opens.
Frequently asked questions
Do I need AI to sort emails or are rules enough?
For many cases rules are enough: by sender, domain or keywords you sort a good share of email without AI. AI helps when the email has to be understood (a client describing a problem in their own words). The sensible thing is rules for the obvious and AI only for the ambiguous.
Is it safe for a system to read the company's emails?
It depends on where they're processed. If external models are used, personal data must be anonymised before sending it and GDPR must be met. For sensitive data you can sort with rules or with models that run inside your own infrastructure, without the email leaving.
How long does it take to set up email sorting?
A first version with rules and a few categories takes a few days. Tuning it —adding AI for the ambiguous, adjusting routing, measuring accuracy— is ongoing work over weeks, because real email always brings cases you didn't foresee.
When is it NOT worth automating email sorting?
When little email comes in and one person sorts it in five minutes, or when almost everything is the same type and there's nothing to route. Also if there's no clear action behind each category: sorting for the sake of it saves nothing.
Route, don't just label
Sorting email works when each category leads to an action and a recipient. Start with rules for the obvious, add AI only for the ambiguous, and always leave an exit to a person for what the system doesn't know. That's how you gain speed without losing a single email on the way.
And don't build it to look modern: measure first how much email comes in and to how many places it goes. With volume and several destinations, it's one of the most profitable moves; without them, typing a forward each morning is still the sensible thing.
Is a shared inbox where everything lands clogging up on you?
If your team loses time sorting email by hand and some important message sits unseen for hours, a good chunk can be sorted and routed automatically. Tell me how the email comes in and we'll look at it.
See the automation service Let's talk about your case