AI for construction: estimating, tenders and project knowledge

Quantities are most reliable from an IFC model: every element has a type, a size and a GUID, so a count is traceable and repeatable. The difficulty is practice: every client models differently, architecture and services come from separate models, and a penetration is only a penetration if the element really crosses a fire-rated separation. That takes logic that follows your own counting rules, not a general-purpose model.

PDF drawings are harder. Symbol recognition on installation drawings (sockets, luminaires, penetrations) now works well enough to save time, but scanned sets, phone photos and incomplete drawings produce errors. That is why we build every count with fixed checkpoints: the AI counts and classifies, the estimator sees the source for every line, corrects where needed, and only then does the file go to the estimate. The output is your own Excel format, with type, size, wall or floor type, fire-resistance class and the reference on every line.

An example of such a request. A passive fire protection specialist has to count and classify every pipe, cable and duct penetration through fire-rated walls and floors for each tender. A BIM modeller does that by hand today from IFC models and PDF drawings, days per project. The solution asked for: detect that an element crosses a fire-rated separation, apply the firm's counting rules, review at fixed points and export an Excel list in the firm's own estimating format with a reference per row. That is a bounded project with a measurable outcome: hours per tender before and after.

AI does not replace your estimating package or ERP; it feeds them. 4PS Construct runs on Microsoft Dynamics 365 Business Central and has standard interfaces for estimating packages such as Ibis. AFAS Profit imports estimates from a construction estimating package as a work budget. Ibis Calculeren voor Bouw (formerly IBIS-TRAD) is the estimating package many Dutch contractors use; in civil engineering, Bakker&Spees CIVIEL and comparable packages work with the RAW system of CROW. An AI count or draft estimate should land in that format, so the estimator carries on where they always work.

The same applies to work orders, hours and purchasing: the AI fills in and prepares, the posting happens in the system your accountant knows. What such an integration looks like in practice is in AI for installation companies.

Since 1 January 2024 the Dutch Wet kwaliteitsborging voor het bouwen (Wkb) applies to consequence class 1: an independent quality assurer checks the work, and the completion notice carries a dossier for the competent authority with the declaration that the building complies. The contractor also hands the client a consumer dossier. The Bouwbesluit merged into the Besluit bouwwerken leefomgeving (Bbl) on the same day. For fire-rated penetrations, NEN 6069 sets the requirements (integrity E and insulation I, usually 30 or 60 minutes), tested to EN 1366-3.

For an AI application this means the evidence has to arise during the work, not afterwards. A count with a GUID per line, a photo with location and time, a report with its source: those are exactly the documents the dossier needs. We design the application so the dossier is a by-product of daily actions. On the AI Act itself: an AI that counts quantities or drafts reports does not fall in the high-risk categories; the Article 4 literacy duty does apply to everyone who works with it. Our free check shows in three minutes where your team stands.