AI tender screening for a construction firm in Amsterdam (2026)

The verdict in three sentences

For a construction contractor turning over 90 million euros, manual tender screening is more expensive than it looks: roughly 400 notices read each month on TenderNed and TED to shortlist 20. A custom AI tender scoring tool turns that pile into a short, reasoned list with the award criteria already extracted from the tender documents. Expect 30,000 to 60,000 EUR to build in the Amsterdam market and 400 to 900 EUR per month to run, with a payback driven mainly by contracts won rather than hours saved.

What the tool actually does, notice by notice

The commercial director does not want another chatbot. He wants a dashboard every Monday showing which notices deserve a bid / no bid meeting. The tool runs four steps on every notice published on TenderNed, TED and the procurement portals used by Dutch municipalities and water boards.

First, collection: notices pulled by CPV code (45000000 and its building and civil works sub-codes), by province (Noord-Holland, Utrecht, Zuid-Holland, Flevoland) and by estimated value. Second, extraction from the tender documents: selection guidelines, technical specifications, bill of quantities, schedule. The language model identifies the EMVI / BPKV award criteria and their weighting, deadlines, required certifications (VCA, ISO 9001, CO2 Performance Ladder level), mandatory site visits and social return clauses.

Third, scoring: each notice is compared with the firm's reference database (delivered projects, values, building types, contracting authorities). Fourth, summary: a one-page digest of the specifications, technical red flags and a score from 0 to 100.

The score is never an automatic decision. It ranks, it explains, and the director decides. That is what makes the tool acceptable to a sales team that knows its clients better than any algorithm.

2026 budget: build, run and alternatives

Tender alert services in the Netherlands cost a few hundred euros a month but stop at detection. They do not read your documents and do not know your references. A custom tool goes further, at the cost of a 10 week project.

At Kolonell, this type of tool falls under our custom application grid: medium complexity 4,000,000 to 7,000,000 FCFA (6,100 to 10,670 EUR), complex 7,000,000 to 15,000,000+ FCFA (10,670 to 22,870+ EUR). The level depends on the number of sources and on integration with your estimating software. Maintenance is then 35,000, 75,000 or 120,000 FCFA per month (53, 114 or 183 EUR).

What makes the project succeed or fail

The first factor is the quality of the reference database. A fit score only means something if the tool knows you delivered three schools worth 6 to 12 million euros in Noord-Holland. That means structuring 5 to 10 years of projects before tuning anything.

The second is traceability. Every extracted criterion must link back to the page of the tender document it came from. When something looks wrong, the bid manager checks in one click instead of rereading 80 pages.

The third is confidentiality. Tender documents are public, your references, prices and win rates are not. The model must run on an offer that does not train on your data, hosted in the EU, with GDPR-compliant access logs.

The fourth is calibration. For two weeks, the team scores 60 to 100 past notices itself. Results are compared with the tool's score and weights are adjusted: building type, value, distance to site, price weighting, whether variants are allowed.

Mini case study

Sarah, commercial director of a construction firm in Amsterdam, spent around 4.5 days a month on screening with her bid assistant. Estimated loaded cost: 4.5 days x 750 EUR = 3,375 EUR per month, or 40,500 EUR per year.

After go-live, screening drops to 1.5 days a month: 3 days saved, worth 2,250 EUR per month or 27,000 EUR per year. Usage and hosting cost 800 EUR per month, 9,600 EUR per year. Net time saving: 17,400 EUR per year. On a 45,000 EUR build, payback on time alone takes about 2.6 years.

The real lever lies elsewhere. With a 4 % net margin in construction, one extra 3 million euro contract won thanks to a notice that would have been missed brings 120,000 EUR of margin. Sarah therefore tracks two metrics: screening time and relevant bids submitted per quarter, up from 9 to 13.

FAQ

Can the AI write our tender response for us?

It can draft a structure and prefill 30 to 50 % of sections from your past submissions. Final wording, delivery commitments and technical choices stay human, because they bind the company contractually.

Do we need paid access to procurement portals?

Most documents on TenderNed and Negometrix are free after registration. The tool manages company accounts on 10 to 20 portals, but some require manual download, which stays with the bid assistant.

How reliable is criteria extraction?

On standard tender documents, expect in order of magnitude 90 to 95 % of criteria correctly extracted with their weighting. The remaining 5 to 10 % involve scanned or poorly structured files, flagged for review.

How long before the score is reliable?

Allow two weeks of calibration on 60 to 100 historical notices, then one month of real use. After that, the team adjusts weights once a quarter in about one hour.

Can the tool connect to our CRM or estimating software?

Yes, through an API or export. A CRM integration usually adds 2 to 3 weeks and moves the project to the complex level of the grid.

Let's scope your project. We define with you the tender sources, the reference database and the integration level, for go-live in about ten weeks. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.