In November and December a Brussels firm, or a Walloon firm selling across the language border, turns a 2027 automation idea into a budget number. This is for the owner or operations lead writing that number. The window matches a calendar financial year. It is not a filing deadline. Fund one workflow, with an owner, a human approval and a baseline you can defend. The later example is fictional, not a client result and not a market average.
On 31 December 2024 the FPS Economy, using Statbel's VAT register, counted 1,186,099 VAT SMEs in Belgium: 122,370 in Brussels-Capital (10.3%) and 299,562 in Wallonia (25.3%). The 2023 to 2024 rise was 3.1% in Brussels-Capital, 1.4% in Wallonia and 1.7% nationally. On that page a VAT SME is a Crossroads Bank enterprise, liable for VAT, with fewer than 250 employees. Recommendation 2003/361/EC keeps the staff ceiling and adds turnover of at most €50 million or a balance sheet of at most €43 million, plus group rules. A 15-person Brussels consultancy and a 180-person Walloon plant can both qualify and still need different projects.
Statbel reported on 2 December 2025 that 34.5% of Belgian enterprises with at least 10 employees used an AI technology in 2025. Written-language analysis was the most common named use, at almost 23%. Three in four enterprises above 250 employees used AI. The FPS Economy overview of 9 June 2026 gives 34.54% in 2025, up from 24.71% in 2024, against a European average of 19.95%. It cites text analysis, written or spoken content, machine learning and workflow automation, mainly in administration, finance, and marketing or sales. Small enterprises are at 28.82% and large ones at 76.41%. The release does not restate headcount bands, so that 28.82% is not every firm under 250 staff. The first obstacle named is a lack of internal expertise, then legal uncertainty and data protection.
Four processes worth scoring before you write the line
These four already match the official pattern: language, documents, administration, finance and sales. Score them before you compare tools.
- Inbox triage. A shared mailbox receives French, Dutch and English. Someone guesses the intent and copies a summary into the CRM. Fund classification, a draft reply and a missing-field check. A person still sends the message.
- Quotes. Proposals are rebuilt from an old file, a spreadsheet rate and exceptions in the owner's head. A wrong discount is expensive. Draft from the current price list and flag anything outside written rules. Do not start here while the rules are unwritten.
- Supplier invoices. Since 1 January 2026, a structured invoice has been compulsory for B2B between Belgian VAT-liable enterprises. FPS Finance says a PDF may accompany it, but only the structured invoice is legally required, via Peppol-connected software unless an allowed alternative is agreed. What remains is the exception: a new supplier, a VAT mismatch, a duplicate, or an amount over your threshold.
- Client onboarding. A signed proposal becomes a checklist and chasing emails. Detect what is missing and draft the chase. Skip this first if the pack holds identity documents.
Intyb's workflow automation work starts from that shortlist. None of the four needs an autonomous agent on day one.
Scope one workflow, not an AI programme
Write one sentence: the workflow starts when a named event happens and ends when a named role updates a named record. "Automate the back office in 2027" is not a scope. "Read supplier invoices already in the accounting package, propose a booking or an exception, and wait for the accountant" is a scope.
- Trigger and finish line. Name what arrives, and the system that holds the finished record: CRM, accounting package, proposal folder or onboarding checklist.
- One AI task. Extract fields, classify, draft or flag a gap. Not all four in the first release.
- Written rules. Thresholds, new-supplier stops, discount limits and language routing do not depend on the model. The model suggests. The rule says when a person must look.
- Approval and a log. Keep the source, the suggestion, the edit, the approver and the time. That trail is what you can show a client, an external accountant or an auditor.
- A manual fallback. If the export fails, the language is unclear or the case is outside the rules, the item returns to today's queue.
A wider way to choose the slice is in how to choose the first AI workflow in a Belgian SME. This article adds the budget question: volume, the cost of a mistake, and whether the data is already readable.
What moves the cost and the calendar?
There is no honest universal euro figure without your volumes and your systems, so none is invented here. Price the drivers. They decide whether the December line still holds in March.
- Access. If an external accountant holds the ledger, the project waits on that firm's export and calendar. Budget the conversation. Do not assume a login.
- Systems. A draft on a review screen is smaller than a workflow that writes to accounting software, a CRM and a mailbox.
- Unwritten rules. Discounts, supplier habits and VAT treatments must be written before a model may suggest a price or a booking. That writing is staff time.
- Languages. A Brussels mailbox often mixes French, Dutch and English. A Walloon team may work in French and still receive a Dutch order from Brussels. Test each language you actually get.
- Review time. Someone must accept, edit or reject every suggestion in the pilot. Unnamed, the workflow will sit unused.
- Running cost. Model calls, the automation platform and a safe copy of real documents are their own line, usually smaller than the staff time.
Confirm access, write the stop rules, try a sample with the person who does the work, then compare it with the baseline. The date slips if the export is blocked, the price list is missing, or the design answers clients itself. Transparency duties for direct interaction have applied since 2 August 2026, where the system itself speaks to the person. A draft a colleague sends is a different purchase. Say which one the budget buys. A live assistant also needs a notice and a human route.
A worked example: three scores, one 2027 project
The table is a worked example for a fictional 22-person firm, office in Brussels, clients in Brussels and Wallonia, ledger with an external accountant. Volumes are assumptions, not a survey: about 90 messages a week, 12 quotes and 35 supplier invoices. Score 1 to 5. Volume is 5 when the queue is a weekly management topic. Error cost is 5 when a mistake changes a price, a tax position or a client promise. Data readiness is 5 when a system you can read already holds the fields.
| Process | Volume | Error cost | Data readiness | Total |
|---|---|---|---|---|
| Shared inbox | 5 | 3 | 4 | 12 |
| Quotes | 3 | 5 | 2 | 10 |
| Supplier invoices | 4 | 4 | 5 | 13 |
The inbox is close: one mailbox, a CRM summary, and the usual mistake is delay. Quotes hurt the margin and score worst on data, because exceptions are unwritten. Supplier invoices win. The fields are already in the accounts package, and a duplicate or a VAT mismatch is costly. This example's 2027 project is an exception desk: propose "book" or "send back", and let the accountant approve a new supplier, an amount over the threshold, or a VAT check the system cannot explain.
If that export is unavailable, the inbox wins, because the firm already controls it. If a price list and a one-page discount rule exist in January, quotes can move to the second half of 2027. Score it with the people who do the work. Do not paste these totals into your budget.
Where this budget line fails
It fails if a chat window summarises an invoice while someone still retypes the booking, or if the first release can send a quote or accept an invoice alone. Do not add identity documents because they were easy to export. The European Data Protection Board's guide for small business asks for minimum data, narrow access and a record of what happened. Names and addresses on an invoice are already enough.
It also fails when Brussels and Wallonia are treated as one operation. A firm in Liège or Charleroi may work in French and still have to read a Dutch order from a Brussels buyer. A Brussels firm may need the same flow tried in three languages before a pilot means anything. The Brussels AI automation agency page is the local context. The score still uses your queue, not a generic SME template.
Measure the baseline before you defend the number
Count two to four weeks before you approve the line, or a full month if you only handle about a dozen items a week. Keep the same definitions after the pilot.
- Items entering the queue each week.
- Median time to the first human review.
- Share of items missing a required field.
- Rework from duplicates, wrong routing and later corrections.
- Minutes spent retyping, chasing or rebuilding the record.
After launch, record suggestions accepted, edited and rejected, and why a person overrode the system. It is not a savings percentage.
To build the figure from your own queue, book a free 30-minute call with Intyb's Brussels team and price your first 2027 workflow. Bring one process, a rough weekly volume and the system that holds the record. The call turns those facts into a scope and cost drivers. It does not promise an unmeasured saving.
Key takeaways
- Fund one workflow with an owner, approval and a baseline.
- Score inbox, quotes, invoices and onboarding before comparing tools.
- The example is fictional, not a savings claim.
