Five tasks every SME can hand over to AI this quarter

Most AI news is about frontier models and billion-euro data centres. For a company with twenty or two hundred people, the useful question is much simpler: which hours of the week could we get back?

Here are the five places where small and mid-sized companies tend to get the fastest and safest return. None needs a data team, and each keeps a person in control of the decisions that matter.

1. Invoices and receipts

Since 1 January 2026, invoices between Belgian companies arrive through Peppol as structured data, so nobody has to type them in. That covers only part of the pile. Receipts for a client dinner, parking or fuel, and invoices from suppliers abroad, still arrive as PDFs, scans and phone photos.

For each one, someone types the supplier, amount, VAT and date into the accounting package and matches it to an order or an expense claim. It's repetitive, rule-based and high-volume, which makes it an ideal first candidate.

Today's AI models read receipts, invoices and scans reliably. Combine that with simple checks (is the VAT number valid? does the amount match the purchase order?) and a one-click approval, and the work shrinks from typing to reviewing.

Start small: one inbox for receipts and foreign invoices, one accounting package, one approval step.

2. The shared inbox

Mailboxes like info@ and sales@ are where time disappears: reading, forwarding, looking up an order, writing the same answer for the tenth time that week.

An AI workflow can classify each message by intent and urgency, pull up the customer and their latest order, and draft a reply in your tone of voice, in Dutch, French or English. A colleague still reviews and presses send. Nobody sorts mail any more.

3. Orders that arrive by e-mail

Customers send orders as PDFs, Excel sheets or a few lines in an e-mail, each in their own format and with their own names for your products. Someone retypes them into the ERP, looks up the article numbers and checks prices and delivery dates.

An AI workflow reads the order in whatever format it arrives, matches the customer's wording to your article numbers and creates a draft sales order in your ERP. Anything it can't match with confidence is flagged, and a colleague confirms the order before it reaches the warehouse.

4. Answers from your own documents

Manuals, procedures, price lists, past tickets: the knowledge is there, but finding it takes a senior colleague. An assistant built on your own documents (technically, retrieval-augmented generation, or RAG) answers questions and shows the sources it used.

Start internally with your own team before putting it in front of customers. Configure it to say "I'm not sure" rather than guess, and to hand over to a person when that happens.

5. The Monday report

Every week someone exports figures from the ERP, the CRM and the webshop, pastes them into a spreadsheet and writes a summary for the management meeting. That can run by itself: collect the numbers, compare them with targets and last week, explain in plain language what changed, and deliver it to your inbox before 8 a.m.

Pick the most boring task that happens most often. It won't impress anyone, and it will pay back first.

How to choose your first one

Score each candidate on four questions:

• Frequency × time: how often does it happen, and how long does it take each time?
• Clear input and output: can you describe what goes in and what should come out?
• Low cost of a mistake: can a person review the result before it matters?
• Reachable data: do your tools have an export, an API or at least a mailbox the workflow can read?

The process with the highest score is your first workflow.

What not to automate first

• Decisions with legal or financial consequences and no human review.
• Processes that nobody agrees on yet. Fix the process first, then automate it.
• One-off tasks. The payback comes from repetition.

A word on data and privacy

Automation doesn't have to mean sending your data around the world. Choose EU-hosted infrastructure where possible, use AI providers and settings that don't train on your data, and document which data each workflow touches. It keeps you on the right side of the GDPR, and it makes the EU AI Act conversation much easier.