SMEs win 63% of local public contracts, but only 35% of the money. How buyers really score a technical proposal

In local government, SMEs win 63% of public contracts. By value, they receive only 35%. These are the 2024 figures from the Observatoire économique de la commande publique (OECP, the French public procurement observatory), published in March, covering contracts above €40,000 excl. VAT: 159,435 contracts awarded by the local public sector, for around a hundred billion euros.

Read together, these two numbers say something simple: small companies win small contracts. When the contract grows, they vanish from the picture. Not necessarily because they are worse. Because the file grows with the contract.

I built a tool for exactly this problem, BidForge. Here is what I learned about how a public buyer scores a bid, and why the hardest part was not getting an AI to write, but stopping it from making things up.

A tender file is a scoring grid in disguise

A call for tenders rarely arrives alone. There is the règlement de la consultation (the tender rules, RC), the cahier des charges (the technical specifications, CCTP or CCP), the acte d'engagement (the signed bid form), the bordereau des prix (the price schedule), and sometimes a template for the technical proposal. For a ten-person SME with no lawyer and no bid manager, that is an evening of reading before you even know whether the contract is worth chasing.

The document that matters most is the tender rules, because it contains the grid. The award criteria are announced there, and in a formalized procedure their weighting is mandatory: that is article R2152-12 of the Code de la commande publique (France's public procurement code). Often, the technical value is itself split into sub-criteria, each with its own points.

Take the sample file I wrote to test the tool: cleaning four buildings for a municipality (a town hall, a school, a village hall, a media library). The tender rules give 60 points to technical value and 40 to price. The 60 technical points are split across four sub-criteria: human resources (20), methodology per site (20), quality control (10), environmental approach (10). And it contains two traps you miss when reading fast: a mandatory site visit, without which the bid is inadmissible, and an eliminatory score: below 25 points out of 60 on the technical side, the bid is rejected, whatever its price.

Once you see the file as a grid, the way you answer it changes.

The four rules of a technical proposal that scores points

A technical proposal is not read, it is scored. The evaluator has the grid in front of them and looks, sub-criterion by sub-criterion, for reasons to award points. Anything that forces them to search costs you.

1. One heading per sub-criterion, in the order of the tender rules, with its exact title. If the tender rules say "Organisation of quality control and handling of complaints", your heading says exactly that. The evaluator must find the answer without hunting for it.

2. Under each heading, take the specifications' requirements one by one, citing the article. The specifications in the sample file set frequencies per site: the town hall Monday to Friday after 5:30 pm, the school on class days between 4:45 pm and 7 pm, the media library three times a week before 9 am. One requested frequency, one promised frequency, site by site. A generic proposal, copied from one file to the next, loses points on every sub-criterion because it answers none of these lines.

3. Verifiable commitments. Who does what, when, and how the buyer can check it: a monthly self-inspection grid sent to the municipality, a quantified corrective response time, a named contact. "A quality service" earns no points.

4. Never claim what you cannot prove. A certification you do not hold, a reference you did not deliver: that is misleading information, and article L2141-8 of the Code de la commande publique lets the buyer exclude a candidate who supplies it. This is not a formality, the entire bid can be thrown out.

The fourth rule is the one that took me the most time.

The real problem with an AI that writes: it fills the gaps

Getting a large language model to write a technical proposal is easy. The result is fluent, well structured, and it lies. Not out of malice: when a piece of information is missing (the staff assigned to the site, the name of the cleaning product, a reference from a similar contract), the model fills the hole with whatever is plausible. It will happily write "our team of eight certified agents" when nothing, neither the eight nor the certification, comes from the company.

In a blog post, that kind of invention slips through. In a public bid, it is precisely what can get you rejected.

So I built BidForge on a separation I already apply elsewhere: the big model writes, a smaller judge reviews, and the code counts. In practice:

• the proposal is written section by section, one per sub-criterion, from the file and from what the company has declared about itself (headcount, equipment, references, real certifications);
• each section goes before an AI judge that looks for three things: claims that come neither from the file nor from the company, hollow passages that answer no requirement, and specification requirements left unanswered;
• a section that fails is rewritten once with the judge's remarks;
• everything that is missing stays visible, highlighted in yellow, in the form "[À compléter : effectif affecté à l'école]" (to complete: staff assigned to the school), instead of being invented.

On the sample file, the proposal comes out in six sections, with 18 yellow markers for the company to fill in. That is 18 places where a less careful tool would have confidently written something false. The result is online, untouched.

What BidForge does, and what it costs

The tool covers the whole path, from finding the contract to a proposal ready for review:

• Alerts: you choose your trade and your départements, and every morning you receive the open calls for tenders that concern you, drawn from the BOAMP, France's official tender bulletin. Free.
• Free analysis of the tender rules: upload the règlement de la consultation, and you get the criteria and their weightings, the deadline and the documents to provide. Also free.
• Express file analysis (€29 excl. VAT): on top of the criteria, the key requirements, the dates, the points of attention, the questions to ask the buyer, and a clear recommendation on whether to bid or not.
• The technical proposal (€149 excl. VAT, €120 excl. VAT if the analysis has already been purchased): written sub-criterion by sub-criterion, reviewed by the judge, delivered in Word.

In parallel I wrote a complete guide to writing a technical proposal, free, which covers these rules with more examples.

If you run a construction or services SME and public contracts seem out of reach, start with the free analysis on your next file. In about thirty seconds you will know what the buyer is going to score.

Sources: Observatoire économique de la commande publique, 2024 census (figures relayed by Weka, March 26, 2026); Code de la commande publique, articles R2152-12 and L2141-8.