Winning public tenders with AI: a Belgian SME guide

Why public tenders stay a blind spot for Belgian SMEs

Every year, Belgian contracting authorities — municipalities, provinces, inter-municipal companies, public social welfare centres, public hospitals, federal and regional services — buy billions of euros' worth of supplies, services and works. And every year, most Walloon SMEs watch those opportunities go by without ever bidding. Not for lack of competence: for lack of time. Using AI to bid for public tenders changes precisely that equation, because the real cost of a tender for a small firm isn't the price of the work — it's the reading, screening and writing time you have to commit with no guarantee of winning.

The calculation every owner makes is simple and rational. A tender specification runs to 60–120 pages. Screening it, understanding it, checking you're eligible, gathering certificates, writing a methodology note, pricing it up: easily three to five person-days. With a hit rate that, for a first-time bidder, rarely tops one in five. Five days for a one-in-five shot, when those same five days spent on an existing client are billable today — the choice makes itself, and it makes itself against the public tender.

AI won't win a tender for you. It attacks the one factor that blocks everything: it cuts those five days down to one or two, which makes the bet statistically defensible. That's a shift in threshold, not a miracle, and it's exactly what a five-to-twenty-person SME needs to open a sales channel it had been ruling out.

The Belgian framework in 2026: what to know before you start

Before we talk tools, we need to frame the ground, because the Belgian mechanics shape the whole strategy.

Belgian public procurement is governed by the Act of 17 June 2016. Above certain thresholds, European publication is mandatory: the notice goes out on TED (Tenders Electronic Daily) and in the Bulletin of Awards. Below them, the contracting authority can use a negotiated procedure without prior publication — it contacts a handful of suppliers of its own choosing, and the contract never appears publicly.

The thresholds were revised downwards on 1 January 2026, for the 2026–2027 period:

| Contract type | 2024–2025 threshold | 2026–2027 threshold | |---|---|---| | Supplies and services (classic sectors) | €221,000 excl. VAT | €216,000 excl. VAT | | Supplies and services (special sectors) | €443,000 excl. VAT | €432,000 excl. VAT | | Works and concessions | €5,538,000 excl. VAT | €5,404,000 excl. VAT | | Negotiated procedure without prior publication | €143,000 excl. VAT | €140,000 excl. VAT |

These figures are published by the Walloon public procurement portal and confirmed by Delegated Regulation (EU) 2025/2152.

Two practical consequences for an SME. First, lower thresholds mean more contracts get pushed into European publication: mechanically, more opportunities become visible on TED and on the federal e-Procurement platform. Second — and this is what most owners miss — if your average deal size sits below €140,000, a large share of your addressable market will never be published at all. It's awarded through negotiated procedures, from supplier lists drawn up in advance. No AI will find you a notice that doesn't exist. In that case the strategy is commercial: get known by the public buyers in your region before the need arises. Using AI to monitor a platform when your contracts are awarded off-platform is an expensive misdiagnosis.

Step 1: automated monitoring of tender notices

The first pool of wasted time is monitoring. Checking e-Procurement and TED by hand every week, reading dozens of titles, opening notices that don't concern you: it's repetitive, it's thankless, and it's exactly what a machine does better than a tired human on a Friday evening.

The basic mechanism runs on CPV codes (Common Procurement Vocabulary), the European nomenclature that classifies every contract. Every published notice carries one or more CPV codes. An alert set on the right codes will therefore push you the notices in your field automatically. That's the foundation, and it's free.

The problem is that CPV is coarse. An "IT services" code will send you a network cabling contract just as happily as an application development one. This is where AI adds a layer of judgement: instead of thirty raw notices a week, you set up a screen that reads each notice and rates it against your real criteria — are you eligible, is the value in your range, is the deadline workable, is the required reference in your portfolio. You now receive three notices, each with a two-line note explaining why it made the cut.

The logic is exactly that of AI-assisted competitive monitoring: collection has been a solved problem for years, qualified screening is the real one. And AI is excellent at qualified screening — provided you give it criteria written down in black and white rather than a vague "whatever might interest me".

Step 2: taking apart the specification in an hour

Once you've kept a notice, you have to face the tender specification. It's the document that decides everything: access conditions, qualitative selection criteria, award criteria and their weighting, the exact form of the bid, technical clauses, penalties.

It's also where the knockout clauses hide. A technically excellent bid gets thrown out unread because a social security certificate was missing, because a required reference covered three years rather than two, because a form wasn't in the right format. Documentary rigour isn't an administrative detail in public procurement: it's the first test, and it's binary.

AI is very effective at this kind of reading, on one condition: make it produce a structured extraction, not a summary. A summary of a tender specification is useless and dangerous — it smooths over precisely the details that eliminate you. What you want is a checklist:

• Every access and exclusion condition, with the required document alongside each.
• Every qualitative selection criterion, with its exact numeric threshold.
• Every award criterion, with its weighting in points or percent.
• Every date: question deadline, submission deadline, any mandatory site visit, bid validity period.
• Every formal requirement: number of copies, annex format, mandated structure of the methodology note.

This pass exists to let you decide fast. On an hour's reading, you know whether you're eligible and whether the game is worth the candle. The method is a cousin of the one you apply when structuring an AI project brief, with one difference that changes everything: here you don't write the requirements, you're subject to them.

The point to watch, and it's a serious one. AI hallucinates. On a tender specification, an invented weighting or a date shifted by a week costs you the contract. So the rule is non-negotiable: every extracted item must point back to the article and page number in the source document, and a human checks the knockout points — dates, certificates, thresholds — against the original text. AI takes your reading from five hours to one; it doesn't take it to zero, and pretending otherwise is the surest way to lose a bid on a technicality.

Step 3: the ESPD and a reusable documentary base

The ESPD (European Single Procurement Document) is the self-declaration that replaces, at the application stage, the production of every certificate. You only supply the supporting documents if you're the prospective winner.

It's well designed, but it's mind-numbingly repetitive: the same company data, the same declarations, bid after bid. It's assembly work, not thinking — and therefore an ideal candidate for automation.

The approach that works is building a living documentary base: a single, maintained folder holding your company identification, social security and VAT data, financial capacity, a structured CV for each team member, and above all a library of client references, each with its context, value, duration, measured outcome and referee contact. From that base, AI assembles the ESPD and annexes for a given bid in minutes, picking the three references closest to the contract at hand rather than copying the three most recent.

The real gain isn't assembly speed. It's that the base exists and is current. Most SMEs waste their time reconstituting, every single time, information they already hold — scattered between the accountant, an old USB stick and the owner's memory. That documentary discipline is the same reflex described in our piece on automating quotes: the value comes from structuring the material up front, the tool merely cashes it in.

One point that matters: the base holds company data, CVs, client contact details. Before handing it to the first tool that comes along, apply the rules set out in our guide to data security when using AI. A CV and a client reference are personal data under the GDPR.

Step 4: writing the methodology note — where AI helps least

This is the heart of the bid, and it's where you have to be honest about the tool's limits.

In most service contracts, the award hinges on a weighting between price and technical quality. The methodology note carries that technical quality: how you understand the need, how you organise delivery, who does what, what risks you anticipate.

AI writes a plausible text fast. It also writes, unprompted, exactly the kind of hollow prose a tender panel has read four hundred times: "a tailored approach", "an experienced team", "close support". That text costs nothing to write — and it earns no points, because it says nothing verifiable.

What earns points is the specific: the name of the council officer who'll be your single point of contact, the response time in figures, the week-by-week schedule, the identified risk and its fallback plan, the figure from last year's comparable project. None of that is in the model. It's in your head and in your files.

So the right use comes down to three roles, none of which is "write it for me".

The structurer. You give the AI the award criteria and their weighting; it proposes an outline that covers them in order, with page counts proportional to the points. Mundane — and yet it's the first mistake first-time bidders make: three pages on a 10-point criterion, one paragraph on a 40-point one.

The challenger. You write, the AI reviews with a hard instruction: "Here are the award criteria. Here is my note. For each criterion, tell me what an evaluator will not find in my text." This is where the return is highest, because a tireless reviewer with no ego in the text sees the holes you no longer see.

The translator. Many Belgian tenders run in Dutch or require a bilingual version. AI genuinely helps here, provided a native speaker checks the technical and legal vocabulary — our article on AI-assisted multilingual translation sets out the guardrails.

What it looks like in practice, and how to measure it

Let's put the numbers in their place. A public tender bid for an SME, without tooling: three to five person-days. With the chain described here — screened alerts, structured extraction of the specification, a current documentary base, AI as structurer and challenger — you can reasonably get to one or two days for a bid of comparable complexity. The first bid will cost you more than it would have without AI, because that's when you build the base.

That ratio isn't a promise, it's an order of magnitude to validate in your own shop. The way to do that is to measure, bid after bid: time spent, submitted yes/no, rejected at selection or evaluated, points scored per criterion. Three bids are enough to know whether your chain holds. The method is the one in our guide to the ROI of an AI project in an SME: without measurement, you'll never know whether the tool saved you time or merely changed what you were busy with.

And stay clear-eyed about the substance: going from five days to two doesn't change your hit rate. It changes how many bets you can afford. One in five stays one in five — but five bids a year instead of one means one expected win instead of a fifth of one. That's where all the value sits, and it's arithmetic, not magic.

Two caveats to close, because advice that doesn't state its limits is worthless. First, if your business mostly sits below €140,000, the bulk of your market runs through negotiated procedures and is therefore invisible: invest in the relationship with public buyers first, tooling later. Second, AI doesn't compensate for missing references. If a contract demands three similar references above €100,000 and you have none, no tool makes you eligible. Start with contracts your size, often as a subcontractor or on a separate lot — the 2026 European reform is precisely pushing buyers to split contracts into more lots to open access to SMEs.

Where to start

Three actions, in this order, with no sophisticated tooling.

This week. Create an account on the federal e-Procurement platform and set an alert on your CPV codes. Free, an hour's work. Watch what comes through for a month without bidding. You'll learn whether your market is published or awarded through negotiated procedures — that's the diagnosis everything else depends on.

This month. Build the documentary base. One folder, current, with company data, CVs and a library of quantified references. It's the most thankless and most profitable task in this whole guide, and it needs no AI at all.

Next quarter. Take one real contract, your size. Run it end to end with the chain described here, and measure. One bid made and measured will teach you more than six months of technology watching.

If you'd like to frame this for your own business — assessing whether your contracts are genuinely addressable, structuring your documentary base, tooling up monitoring and specification reading — let's talk directly. Half an hour is usually enough to know whether the game is worth the candle in your case, and I'd rather tell you plainly when it isn't.