AI triage for the info@ mailbox: how to take three hours a day off a team in four weeks

· 8 min read · SilverData team

The shared info@ mailbox in a typical 30-person company is still one of the largest, silent drains on productivity. Three people spend an hour a day each switching between the machine and Outlook just to conclude "this is for accounting", "this is for sales rep X", "this is spam". AI is good enough at this today — and cheaper than a new hire.

Over the past year we've done a dozen or so rollouts like this. It isn't about "the AI assistant that answers customers on its own" — that still doesn't work well enough to be handed off without human oversight. It's about something much less impressive that actually works today: automatic classification of incoming mail and routing it to the right person, folder or system. Boring. But this is exactly the layer of AI that pays off in an SME.

Where the problem comes from

A company grows from 8 to 30 people. The info@ mailbox was logical at the start — one person read it and forwarded. At 30 people, that mailbox now gets 200-500 messages a day: quote requests, invoices, complaints, supplier offers, spam, newsletters, confirmations, customer attachments, mail from the accountant, system notifications, CVs. Nobody planned it — it just accreted.

In practice this means every email passes through a human twice: once to be sorted, once to be handled. That's waste. AI is perfectly suited to remove that first pass.

A real example — spare-parts distributor, 30 people

A client we worked with: a distributor of agricultural machinery parts, 30 people, Comarch Optima ERP, mail on Microsoft 365. The info@ mailbox took an average of 380 messages a day. Four people (two office assistants, two in the warehouse) together lost about 3-4 hours a day just sorting — opening, classifying, forwarding to the right person or folder.

After rolling out a simple AI classifier (17 categories + routing rules) and a three-week pilot with a human in the loop:

This isn't a "the company cut jobs" story. It's a "four people stopped hating Monday mornings and went back to the work you actually pay them for" story.

What such triage actually does

In the first-rollout version the model gets the email content (subject + first 2000 characters + attachment list) and returns:

The rest is regular automation: Power Automate, n8n, Make or Exchange rules — whatever your company already has and maintains. The model is only a classifier. The plumbing stays in tools the admin already knows.

What NOT to do in the first rollout

The biggest mistake we see: "since we now have AI, let it answer customers on its own". No. In the first rollout the model never sends anything outside the company. It classifies, routes, drafts — but a human sends. Two reasons: cost of error (one badly worded reply to a complaint can cost more than the annual subscription) and team trust (if week one shows "the AI writes nonsense", you kill the project at the start).

Second mistake: "let's give the model 60 categories to cover everything". No. Start with 12-20. If you start with 60, the model will confuse the close ones, people will lose trust, and you'll spend months tuning. Better a folder other — manual with 5% of traffic than 60 folders where each has 3% mistakes.

Third mistake: "let's throw all the company's mailboxes into it". Start with one mailbox, ideally info@ or orders@. If it works, the team will ask for the next one themselves.

What this looks like technically in an SME

You don't need a EUR 200k AI platform. A real stack from recent rollouts:

GDPR and sensitive data

This is the most frequent question from management. The short answer: yes, you can, but you need to be able to state three things:

  1. Which provider and where they process data. Pick one that offers EU processing (Azure OpenAI EU region, Google Vertex EU, Anthropic via AWS EU). Sign a Data Processing Agreement (DPA).
  2. What you do NOT send to the model. National IDs, medical data, passwords — filter before sending. That's one regex rule in Power Automate.
  3. How much logging the provider keeps. Choose "no training", "zero retention" where available (that's a standard option in enterprise tiers of most providers from 2025 onwards).

That's it. You don't need to build your own model in the basement to be compliant.

When it does NOT make sense

To be honest. AI triage pays off when three conditions hold at once: (1) the mailbox gets more than about 100 emails a day, (2) more than one person handles it, and (3) someone is already doing manually something you can describe in 20 sentences ("this goes to accounting, this goes to sales rep Y"). If you have 30 emails a day and one assistant — don't bother. Fix your Outlook folders and rules, and go do something more important.

What next

This is one of those AI rollouts that doesn't change the company's strategy, but genuinely gives back a few hours a day and typically pays back in 1-2 months. If it sounds like something that fits your company — instead of guessing at the deployment, we always start with a 30-minute conversation: which mailboxes you have, who reads them today, how much time is sunk in them.

AI in an SME pays off not when it answers instead of a human. It pays off when it takes off a human the work no one ever wanted to do.

← Back to blog