AI and public tenders in construction: what to automate, what stays yours
Monitoring, reading the tender file, go/no-go, technical proposal, compliance documents: where AI helps on a public works tender, and what stays in your hands.
The short answer. On a public works tender, AI is already useful at four steps: spotting the notices that concern you, reading the tender file (the French DCE) to pull out requirements and deadlines, drafting an outline of the technical proposal (mémoire technique) from your own references, and checking that the bid is complete and your certificates are up to date. It does not decide whether to bid, it does not set your price and it signs nothing. Used well, it mostly gives you back the time you would have spent reading files for nothing.
Why this matters for a small construction firm
Public contracts are not just for large groups. According to the 2024 census of the French public procurement observatory (OECP; contracts above €40,000 excl. VAT, published March 2026), 223,383 contracts worth €233 billion were awarded in 2024, and in the local public sector 63% of contracts by number went to SMEs (35% by value). Figures as reported by Weka and Seban Avocats from the OECP report. As the two summaries disagree on the overall SME share, we only keep the local-sector figure, which both give.
But bidding costs a lot of time. A works tender file means bidding rules, administrative and technical specifications, a price breakdown, drawings: often hundreds of pages. In the small construction firms we meet, it is the owner, sometimes with an assistant, who monitors notices, reads the file, rewrites the technical proposal from older ones and chases certificates. The worst part is the half-day spent reading a file only to conclude it was not worth bidding.
Worth knowing: below €100,000 excl. VAT, a French public buyer can award a works contract without advertising or competition. This threshold, initially temporary, was made permanent by decree no. 2025-1386 of 29 December 2025 (FFB, January 2026). Small public jobs are therefore also won through relationships and local reputation, not only through published notices.
What we learned building our own tool
Since this summer, we have been building a tool for our own public tender responses. It follows the steps of a bid: monitoring, go/no-go, reading the tender file, technical proposal, compliance check. Three lessons apply to any small business, with or without a tool.
1. Most of the sorting needs no AI. Monitoring collects notices from open sources several times a day (BOAMP, buyer platforms, the EU Official Journal). The first filter, which costs nothing, keeps what falls within your area and trades; simple rules then discard the rest. In a trial with the profile of a construction firm (its trades, its area), on the first fill of the monitoring, about 5 notices out of nearly 1,800 were left to read in detail. An AI model only reads at the very end, on those few notices, and it costs us less than one euro cent per notice, measured on a real run.
2. A good tool says “I don’t know”. On a real file, the cover page of the bidding rules was an unreadable scan. Instead of inventing a deadline, the tool flagged the line as uncertain and asked for a human check. Another trap we hit: a deadline stored in universal time, shown two hours earlier than Paris time. Nothing was wrong, but without the time zone on screen, the page seemed to contradict itself. On a submission deadline, that kind of detail matters.
3. Compliance documents are the most cost-effective check. Company registration extract, tax and social security certificates, liability and decennial liability insurance: each has an expiry date. An expired document can get a bid rejected, however good the proposal. We chose a simple rule: any document that expires and whose date could not be read stays “to be checked”. For subcontracting, the French social security “vigilance” certificate must be requested again every 6 months for any contract of €5,000 excl. VAT or more (Urssaf): exactly the kind of deadline people forget.
Step by step: what AI does well
Monitoring
A tool (or a simple filter on free BOAMP alerts) collects notices and sorts them by your trade codes, your area and the lot sizes you can handle. AI then helps summarise each retained notice in three lines: subject, location, deadline, estimated value when published.
Go/no-go
AI can list the reasons to bid or not based on criteria you set: distance, required qualification, completion time, references asked for, team workload over the period. The decision stays yours: it depends on your order book, your relationship with the buyer and your appetite for risk.
Reading the tender file
This is where the gain is clearest. An AI model reads the bidding rules and technical specifications, then lists the requirements, the award criteria and their weighting, the documents to provide, the dates (mandatory site visit, questions, submission). Each line should point back to the passage it comes from, so you can check it in one click.
The technical proposal
AI drafts an outline that follows the criteria in the bidding rules, then fills it from your material: past proposals, project references, staff and equipment, methods. It must invent nothing: a reference you do not have or a commitment you cannot keep ends in a dispute. You review, adapt to the project and sign.
The check before submission
Last pass: does every requirement have an answer? Is every document present, signed and current? AI does this checklist well, provided it flags what it could not verify rather than ticking by default.
What stays human
- The price. The price breakdown depends on your costs, your suppliers and your strategy. No model knows it.
- The decision to bid. AI informs, you decide.
- The relationship with the buyer: site visit, questions, negotiation where the procedure allows it.
- Signature and liability. What the proposal says commits you.
One more limit: tender files sometimes contain sensitive information (drawings of public buildings, personal data). Before handing them to a tool, check where the data is processed and whether it is used to train a model. It is one of the first things we check in an assessment.
Where to start this week
Take the last three tenders you looked at and note, for each: time spent reading it, time spent answering, and the final decision. If reading weighs heavily on files you dropped, start with monitoring and reading the tender file. If the technical proposal wears you out, gather your past proposals and references in one clean folder: it is the material without which no AI can help you.
To put euros on that time, our calculator Hours freed, euros gained helps you do the maths honestly, and our article How much does an AI agent cost a small business gives the orders of magnitude. Our packages are on the pricing page, and the free 30-minute assessment is held by video call. ROI-first approach: we first check whether the time saved justifies the tool.
Frequently asked questions
Can AI answer a public tender for me?
No. It can read the tender file, draft a technical proposal outline from your documents and check the bid. The price, the decision to bid, adapting to the project and the signature remain yours.
Can I use ChatGPT to analyse a tender file?
Yes for a first summary, with two precautions: check every piece of information in the original document, and do not upload sensitive documents without knowing where the data is processed. A dedicated tool mainly adds traceability (each requirement points to its page) and deadline tracking.
Is a technical proposal written with AI penalised?
What is assessed is the content against the criteria in the bidding rules. A generic proposal, whether written by AI or copied from an old file, shows. A proposal that talks about the actual project, with your real references, holds up.
