The problem

Construction projects rarely fail because of structural calculations. They fail because up to 70% of project costs reside in materials, while procurement awards and gate deliveries are handled via spreadsheets, WhatsApp messages, and paper receipts.

Historical disasters

When the supply chain fractures.

€900M
Fyra HSL (2013). Written off entirely after a public tender awarded strictly on lowest purchase price. Physical winter durability and material performance were never verified; the fleet was permanently grounded within six weeks.
€435M
Afsluitdijk (2024). Massive cost overrun caused by material specifications diverging from contract terms without an auditable verification trail. What was delivered to the site failed to match what was contracted.

Parliamentary Inquiry Fyra 2015 · Netherlands Court of Audit 2024

The triple crisis

Office, gate, and memory.

up to 50 mingatepaper ticket
The gate today: 70% of trucks wait up to 50 minutes, 95% of tickets fail data quality checksAldeeb 2024 · Qflow 2025

The breakdown of traditional construction procurement unfolds across three consecutive operational stages:

1. Pre-construction clerical bottleneck: Estimators spend 31.7% of working hours on administrative data entry instead of value analysis. Calculating a single formal tender quotation takes a median of 42 labor hours. While contractor margins remain compressed between 1.0% and 2.0%, construction failure costs run between 5% and 11% of turnover.

2. Jobsite and gate breakdown: 70% of delivery trucks wait up to 50 minutes at the site gate due to manual checking. 95% of delivery tickets failed data quality criteria in an audit of 831,206 records. Delivery failures halt 8% to 25% of scheduled tasks on site.

3. Institutional amnesia: 78% of material procurement relies on informal emails and WhatsApp chats. Less than 50% of field purchases comply with negotiated master agreements. At one contractor studied, cleansing more than 1,000 historical projects left 54 usable, while on 500 to 600 million euro projects fewer than 5% of staff stay from tender to completion, one interviewee estimates.

Donyavi et al. 2024 · Laryea & Hughes 2011 · EIB 2014 · Aldeeb 2024 · Qflow 2025 · Rockx 2023, TU Delft · Steuer et al. 2023

Why now

Regulation demands auditable proof.

Informal processes are legally closed. Four binding frameworks make spreadsheets and paper clipboards indefensible:
Wkb
Dutch Quality Assurance Act (Wkb). The Dossier Bevoegd Gezag demands as-built proof of all installed materials and inspections. Missing certificates cause 4 to 6 week handover delays before municipal occupancy permits are granted.
CPR 2024
Construction Products Regulation (EU 2024/3110). Digital Product Passports (DPP) and Environmental Product Declarations (EPD) are mandatory at the gate for supply chain traceability.
Art. 2.130
Public Procurement Act 2012. Every public award decision requires an objective, verifiable, and reproducible audit trail for municipal review and regulatory scrutiny.
AI Act
EU AI Act (High-Risk). Systems supporting procurement and supplier selection face strict transparency, documentation, and determinism mandates by December 2027.
The spreadsheet and clipboard are no longer legally compliant.

Precision

Generic AI fails to calculate reliably.

€8,354

swing on one subcontractor total when a generic enterprise AI tool got the same input in a new session

Hamppi 2025, Aalto University. Case study based on an internal Skanska experiment.

In an internal Skanska experiment, a generic enterprise language model returned different numerical outputs for identical procurement inputs, with deviations up to 28%. This is the inherent nature of probabilistic language models (Atil et al. 2025).

A language model predicts the most probable next word; an optimization engine calculates the mathematical optimum. Procurement decisions involving contract liabilities require exact mathematics, not statistical approximations.

Dundir decouples language from computation: the AI layer structures specifications, while the deterministic solver calculates the optimal allocation.

What this does not prove The experiment proves that generic language models are unsuitable for reliable numerical comparisons. It does not prove that our solver buys more cheaply in practice than an experienced chief estimator. That validation requires empirical pilot data gathered on active projects.

Four ways in

Where you come in depends on what you need.

Open now

Become a research partner

Construction SMEs join the research track with Windesheim. No cost, and a say in what gets built.

What it involves Way in 02

Pilot as a design partner

The system is not finished. You decide what it gets calibrated on, and you own the result.

Terms Way in 03

Procurement scan

Ten questions about your current process. Report on screen straight away. No email address needed to read it.

Start the scan Way in 04

A 30 minute conversation

No deck, no pitch. We ask the questions.

Request a call