AI on the Building Site: Useful Tools, Human Responsibility

A shared drawing register, a building information model, a digital inspection form and a generative AI assistant do different jobs. A three-dimensional model does not automatically contain verified, current information. A dashboard does not automatically measure what its title suggests.

Before buying or using a tool, define the problem. Is the team losing time finding drawings, recording inspections, interpreting requirements or repeating administrative work? AI will not necessarily repair a missing document-control process.

A useful first trial might involve summarising approved meeting notes, drafting a plain-language explanation for review or organising a list of questions for a designer. Those uses can be tested against known source material. A safety-critical design decision needs a much more demanding evidence and review process.

Using AI does not remove health-and-safety responsibilities. The Health and Safety at Work etc. Act 1974 and CDM 2015 remain relevant in Great Britain; Northern Ireland has its own health-and-safety framework and CDM 2016. Building-regulations responsibilities follow the applicable national system. [1,2]

If a business uses generated text in a risk assessment or method statement, the business still needs to establish that it addresses the real work. “The software said so” does not demonstrate that the hazards, interfaces or controls were properly assessed.

Similarly, an electrical calculation must be checked against the actual design conditions by someone competent for that work. Regulation 16 of the Electricity at Work Regulations 1989 concerns necessary knowledge, experience or appropriate supervision; access to a tool is not equivalent to having those capabilities. [3]

Ask for the source of a claim, then open that source. Confirm that it exists, says what is claimed, applies to the right jurisdiction and remains relevant to the work. A realistic-looking reference can still be wrong.

For project information, verify the drawing revision, status, units, inputs and exclusions. A calculation can be arithmetically consistent but based on an incorrect load or an omitted condition. A summary can accurately repeat a document that has already been superseded.

Separate a draft from an authorised instruction. Decide who may approve the result and how people on site will recognise the approved version. Retain enough of the input, source material, checking and decision to understand what happened if a later concern arises.

Site photographs, accident reports, employee assessments and attendance records can contain personal information. Drawings may also contain confidential, security-sensitive or commercially valuable material even when no individual is identifiable.

Before using an external AI service, establish what the supplier does with uploads, how long information is retained, who can access it and whether the organisation is permitted to share it. Removing a name may not anonymise a record if the person remains identifiable from the other details.

The UK GDPR and Data Protection Act 2018 continue to apply, as amended by the Data (Use and Access) Act 2025. The ICO’s organisational guidance, updated on 19 June 2026, states that all the Act’s data-protection provisions are now in force. An article repeating the old rules without checking those changes could mislead readers. [4]

Where personal data is involved, assess the lawful basis, transparency, necessity, accuracy, security and retention arrangements. High-risk processing may require a data-protection impact assessment. Contracts, confidentiality and copyright require separate consideration: a lawful basis to process personal data does not automatically grant permission to reuse someone else’s drawings. [5,6]

The 2025 Act changed the framework for significant decisions based solely on automated processing. It is inaccurate to repeat a blanket claim that such decisions are always prohibited. Safeguards remain important, including opportunities to challenge decisions and obtain human intervention; special-category information attracts additional restrictions. Check the current framework before using automated scores for consequential workforce decisions. [4,7]

For SKEB assessment, ask what the score actually measures. A text-based quiz may assess some knowledge. A simulated decision may provide evidence of judgement in that scenario. Neither automatically proves practical skill, verified work experience or consistent behaviour on a live site.

A fair assessment should expose uncertainty. Allow a reviewer to inspect the evidence, correct an error and explain why a result changes. If a worker’s language, reading speed or access needs affect how they use the test, examine whether the assessment is measuring the intended ability.

In a small, controlled trial, record the task, time spent, errors found, checking time and rework. Compare like with like. An output generated quickly can still cost more overall if it takes extensive correction or creates confusion downstream.

For example, compare two sets of meeting minutes against the same approved recording or notes. Count omitted decisions, incorrect owners and invented deadlines. A manager can then judge whether the tool improves the complete process, including review, rather than celebrating generation speed alone.

These are suggested evaluation measures, not published claims that AI delivers a particular percentage improvement. There is no reason to invent a productivity graph when a business has not measured its own results.

Agree which uses are permitted, which information may be entered, which decisions require specialist review and who authorises release. Train people to recognise a plausible but unsupported answer and give them a route to report it.

AI can assist a capable team. Its use becomes defensible when the organisation can explain the evidence behind the decision, the limits of the tool and the person responsible for accepting the result.

• Skills Are Not Enough: What Competence Really Means in Construction: Why recruiting more people must go hand in hand with developing and evidencing competence.
• Building Safety After Grenfell: Accountability Beyond the Paperwork: Dutyholders, higher-risk buildings, controlled changes and evidence that matches the work.
• Retrofit That Works: Why the Whole Building Matters: Insulation, ventilation, heating and electrical work have to be designed together.
• Getting Paid in Construction: The Small Print That Controls Cash Flow: Payment notices, due dates, variations and retention: the details that decide cash flow.
• Immersive learning in construction
• Virtual reality training
• Explore Tradefox courses and simulations

1. HSE — CDM 2015 guidance
2. HSENI — CDM (Northern Ireland) 2016
3. HSE — Electricity at Work Regulations guidance
4. ICO — Data (Use and Access) Act: implications for organisations
5. ICO — data-protection impact assessments
6. Government — copyright protection
7. Government — Data (Use and Access) Act data-protection changes

Fact-checked against the linked sources on 8 October 2026. Educational overview: project circumstances, applicable amendments and transitional provisions must be checked before decisions are made.