An AI agent for tenders does four things, and it does them unevenly well. It reads the tender pack and extracts the criteria, the required documents and the deadlines. It scores the pack to help call the go or the no-go. It drafts a first technical proposal from your past responses. Finally it suggests an order of magnitude for the price, based on your history. Extraction works very well today. Pricing, far less so.
This article describes what artificial intelligence can do on each of those four links. The stack behind them, the real cost, and the places where human validation stays mandatory. Construction serves as the example because that is where the packs are heaviest, but the mechanics hold for engineering, services and consulting.
What an AI agent really extracts from a tender pack
Extraction is the most mature link in the chain. An AI agent opens the tender pack, runs OCR on every document, then locates the fields that decide what happens next. The consultation rules give the deadline, the scoring criteria and their weighting, and the documents required on pain of rejection. The administrative clauses give the penalties, the payment terms and the retention. The technical specifications give the lots, the quantities and the mandatory standards.
The orders of magnitude are public at the large construction groups. Vinci Construction built an in-house application, Select, that pre-analyses tenders. It pulls 24 types of data per pack and gets around 19 of them right on 80% of the tenders it processes. More than 5,400 packs have been submitted to it, by 800 employees across 124 branches, and analysis time dropped from about half an hour to around 35 seconds, according to Le Monde Informatique in October 2024.
Eiffage describes the same shift on its internal generative AI platform, built with Google Cloud. The group reports tender pack analysis dropping from two days to twenty minutes. Jean-Philippe Faure, its chief information officer, explains in this published account from April 2025 that one detail matters. The instructions were written by an operational employee, from field experience. The IT department then scaled them up.
How the agent helps call the go or no-go
Once the fields are extracted, the agent compares them against your company. It matches the required references to the ones you hold, and the quantities to your teams' capacity. It checks the deadlines against your workload. It then compares the theoretical margin on the lot with what you actually earned on comparable projects. Out of that comes a score and a list of sticking points. The decision stays yours.
This is where the agent pays off fastest. Across the consultations arriving every week, the opportunity cost sits in two places. The winnable packs nobody saw. And the packs worked on for three days before someone spotted a missing reference. Automatic pre-scoring surfaces both cases within minutes.
- Eliminating criteria: the agent first checks that you are admissible. A missing certification, a minimum turnover you do not reach, a mandatory reference you do not hold.
- Workload and calendar: it cross-checks the start date and announced duration against the projects you have already committed to.
- Expected margin: it maps the nature of the lot onto your past results on the same type of work.
- Cost of responding: it estimates the volume of technical writing required, which prices the bid itself.
Tandem runs the same mechanics on its own commercial watch, over the AI tenders published every week. The finding fits in one sentence. Automatic qualification triggers judgement earlier, and on the right packs.
Drafting a first technical proposal from your archives
For an agent to write, it needs your past responses, your project records, your working methods and your references. Technically, those documents are chunked, vectorised and stored in a database, and the agent searches them by similarity as it drafts. This is called retrieval augmented generation, and I filmed the full setup in my video tutorial. The part that matters for a technical proposal starts at the document base.
The output is a first draft, to be reworked before submission. It reuses your own phrasing, cites the right references and follows the structure imposed by the consultation rules. The writer arrives at a filled page rather than a blank one. They spend their time on the two or three paragraphs that genuinely move the score.
The prerequisite rarely shows up in sales demos. Our audits surface it almost every time. During our audit at Coulon, the pricing office of an energy retrofit contractor, the document base came out as the prerequisite for most of the selected use cases. Eighteen years of records had piled up project after project, scattered and unstructured. Without that tidying, the agent reads the pack and has nothing to compare it against.
Pricing from history, and where AI gets it wrong
Pricing is the most fragile link, and it is also the one vendors promise soonest. An agent can match the lines of a bill of quantities to those you have already priced, suggest a median unit price and flag the outliers. It knows nothing about current pressure on a supplier, difficult site access, or the commercial relationship that justifies a thinner margin on this particular client.
The right way to frame it is to ask the agent for a variance rather than a price. It reports that this line sits 18% above your median over the last three years, and says why. The account manager decides. A submitted bid commits the company for years, and no tool carries that commitment for you.
Buy tender software or build your own agent
The question comes up in almost every engagement, and I raised it in July 2026 in a post on build versus buy. The answer has not changed since. Buy when your process looks like everybody else's. Build when it carries your difference.
For tenders, the border is reasonably clear. Watching the platforms, extracting the pack and preparing the administrative documents are the same for every company in a sector. Off-the-shelf software covers them, quickly and at a known price. Pricing against your own history and writing in your own voice exist nowhere but inside your company.
What an AI tender agent costs
The cost reads on three lines, and the artificial intelligence model comes last. Document processing is counted in cents per document. On our extraction projects, the order of magnitude sits below five cents per processed document, orchestration included. The genuinely heavy item remains integration into your environment, detailed in our article on project cost.
| Item | What you pay for | Price around |
|---|---|---|
| Specialist tender software | Subscription per seat or per pack | From a few tens to a few hundred euros a month |
| Document processing | OCR and model, on usage | Under 5 cents per document |
| Knowledge base | Hosting and indexing your archives | A few tens of euros a month |
| Custom integration | ERP wiring, business rules, acceptance testing | The main item, a few weeks of project |
| Maintenance | Platform and document changes | An ongoing commitment, not a one-off cost |
The payback calculation is easy to set up. Count the time your pricing office spends going through packs, multiply it by a loaded hourly cost, and compare. At Coulon, the audit estimated the gain at more than a thousand hours a year across all the selected use cases. Copying quote lines into architect schedules was one of them, as was querying technical documents instantly.
Should you hand your tenders to an AI agent?
Yes for reading, sorting and preparing. No for deciding and signing. Artificial intelligence is excellent today at document extraction and decent at drafting a first version. It stays mediocre at pricing, because the price of a bid contains information that is written down nowhere.
The movement reaches well beyond the large groups. The French building federation ran a training session in June 2026 on AI as an assistant for responding to tenders. France's business directorate separately launched in early 2026 a call for interest aimed at small trade businesses, construction among them. It covers eleven use cases, drafting quotes among them.
If you receive more packs than you process, start with extraction and pre-scoring. Those two bricks ship in a few weeks and can be measured immediately. Pricing comes afterwards, once your archives are tidy. Tandem runs these projects as part of our Autopilot work, and construction is one of the sectors where AI applied to admin flows pays back fastest.
The questions you ask us about AI and tenders
Three of them come up every time. They all carry the same worry, which is how far the machine can be left alone. The short answer is short. The agent produces, the human commits. For the technical mechanics that let an agent cite your own documents without inventing, we covered them in our article on RAG.
