EU AI Act Compliance for Construction AI (2026)
Why it matters
The EU AI Act (2024/1689) is binding from 2025–2027 in phases. Construction AI — estimators, computer-vision site monitoring, agent assistants — mostly lands in limited or minimal risk, but misclassification or missing transparency is a compliance failure, and high-risk uses (safety-critical, worker monitoring, access decisions) bring real obligations.
Classification quick check for construction tools
| System |
Likely tier |
Obligations |
| Cost estimation assistant (BOQ, unit prices) |
limited / minimal |
transparency (users know it is AI), basic logging |
| Schedule optimisation |
minimal |
none beyond general law |
| Site camera CV for progress |
limited |
transparency, data minimisation |
| CV for PPE/safety enforcement |
high risk (safety component?) |
full: risk management, data governance, human oversight, logs |
| Worker performance monitoring |
high risk (employment) |
prohibited or high-risk — treat carefully |
| AI agent that signs/submits binding documents |
not allowed without human |
human oversight mandatory |
Rule of thumb: if a human previously had to sign it, an AI must not sign it alone.
Practical compliance checklist (per tool)
- Classify — document the risk tier and reasoning.
- Transparency — users must know they interact with AI output; mark AI-generated estimates as such.
- Human oversight — a qualified human reviews binding estimates, contracts, safety decisions.
- Data governance — minimal personal data in site CV; no biometric identification.
- Logging — record model, inputs, outputs, version (the ERP usage-ledger pattern).
- Instructions for use — document limitations (e.g. early estimates ±30–50%).
Mapping to the DDC stack
oce-estimate-boq, cost-estimation-*, cost-prediction: limited risk — add AI-output labels.progress-monitoring-cv, defect-detection-ai: limited — no personal data, transparency.safety-compliance-checker enforcing PPE via CV: reclassify — keep a human in the loop, document.ai-agent-orchestration: supervisor must be a human at decision gates.
Resources
1---2name: ai-act-compliance3description: Check construction AI systems against the EU AI Act (Regulation 2024/1689): classify risk of estimation, scheduling, CV and agent tools, document transparency, keep human oversight. Use when deploying or auditing AI in construction.4---5
6# EU AI Act Compliance for Construction AI (2026)
7
8## Why it matters
9
10The EU AI Act (2024/1689) is binding from 2025–2027 in phases. Construction AI — estimators, computer-vision site monitoring, agent assistants — mostly lands in **limited** or **minimal** risk, but misclassification or missing transparency is a compliance failure, and high-risk uses (safety-critical, worker monitoring, access decisions) bring real obligations.
11
12## Classification quick check for construction tools
13
14| System | Likely tier | Obligations |
15|---|---|---|
16| Cost estimation assistant (BOQ, unit prices) | limited / minimal | transparency (users know it is AI), basic logging |
17| Schedule optimisation | minimal | none beyond general law |
18| Site camera CV for progress | limited | transparency, data minimisation |
19| CV for PPE/safety enforcement | **high risk (safety component?)** | full: risk management, data governance, human oversight, logs |
20| Worker performance monitoring | **high risk (employment)** | prohibited or high-risk — treat carefully |
21| AI agent that signs/submits binding documents | not allowed without human | human oversight mandatory |
22
23Rule of thumb: **if a human previously had to sign it, an AI must not sign it alone.**
24
25## Practical compliance checklist (per tool)
26
271. **Classify** — document the risk tier and reasoning.
282. **Transparency** — users must know they interact with AI output; mark AI-generated estimates as such.
293. **Human oversight** — a qualified human reviews binding estimates, contracts, safety decisions.
304. **Data governance** — minimal personal data in site CV; no biometric identification.
315. **Logging** — record model, inputs, outputs, version (the ERP usage-ledger pattern).
326. **Instructions for use** — document limitations (e.g. early estimates ±30–50%).
33
34## Mapping to the DDC stack
35
36- `oce-estimate-boq`, `cost-estimation-*`, `cost-prediction`: limited risk — add AI-output labels.
37- `progress-monitoring-cv`, `defect-detection-ai`: limited — no personal data, transparency.
38- `safety-compliance-checker` enforcing PPE via CV: reclassify — keep a human in the loop, document.
39- `ai-agent-orchestration`: supervisor must be a human at decision gates.
40
41## Resources
42
43- EU AI Act text: https://artificialintelligenceact.eu
44- Commission guidance & codes of practice: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
gabrielmoreira