AI in construction: find the right project answer before the site moves on

By Kirill Naumov · 2026-10-24

The first useful AI on a construction project may not be a robot on site. It may be a project manager finding the approved drawing revision before a crew acts on the wrong one. A commercial lead needs the clause behind a variation; a supervisor needs to know whether an instruction is approved or still being reviewed. The questions are ordinary. The cost of a confident answer from an obsolete file is not.

Adoption is real, but still early

RICS’s 2025 global survey of more than 2,200 professionals found that 45% reported no AI implementation in their organisations and 34% described early pilot phases. Respondents saw potential across design, progress monitoring, scheduling and document review. These are reported views, not evidence that a particular product improves project outcomes. The practical signal is simpler: many teams are exploring AI, while the path from a promising demo to a dependable workflow is still being worked out.

RICS also identifies skills, integration, data availability and implementation costs as barriers. That is a good reason to begin with a bounded process—not a company-wide transformation plan. Pick one recurring information task, one document set and a team that can check the answer against an authoritative source.

Make every answer point back to evidence

An assistant searching approved specifications, drawings, meeting notes and change records should return more than a polished paragraph. Ask it to show the source file, revision, page or section and date. If two records conflict, the system should surface the disagreement and route it for review. Quietly choosing the answer that sounds most plausible is not coordination; it is a new way to hide uncertainty.

Set the information boundary before uploading anything: which folders are in scope, who marks a file current, how superseded versions are labelled, and which information is confidential or restricted to a project team. Check the provider’s access, retention and data-use terms before project documents go into an AI service. If the approved record is missing, the assistant should say ‘not found’ rather than improvise an instruction.

RICS’s responsible-use standard calls for data and system governance, risk management, output assurance and professional judgement. Its construction case material also makes clear that AI outputs can depend on incomplete context and must be assessed by the responsible professional. The standard is not a call to avoid tools; it is a reminder to design the review into the work.

Keep consequential decisions with named people

AI can help locate a contract clause, compare revisions, summarise meeting actions or prepare a draft response. It should not approve a variation, certify a quantity, change a safety procedure or instruct a team to proceed on disputed information. Those decisions carry technical, contractual or safety responsibility. A qualified person checks the cited material, considers site conditions and records the decision through the project’s normal process.

A useful pilot rule is: AI can help answer ‘Where is the evidence?’ and ‘What does this source appear to say?’ It cannot turn an ambiguous document into formal approval. NIST’s Generative AI Profile recommends governance, role clarity, human oversight and evaluation against known information. In practice, that means recording who owns the workflow, what the tool may do, what requires escalation, and how errors or overrides are reviewed.

Test a real delay, not a flashy demo

Choose a recurring task that costs time: finding an approved detail, tracing a change, preparing a handover note or checking an action against the latest record. Start with historical questions whose correct answers are already known. Require citations and a clear ‘not found / needs review’ response. Ask project staff to score answer accuracy, source quality and the time needed to verify it. Record misses and near misses as carefully as successful searches.

Measure the problem you set out to solve. If the goal is faster retrieval, compare the time to locate and verify a correct answer. If the goal is fewer coordination errors, define what counts as an error and whether the pilot can influence it. A chatbot producing a paragraph in seconds is not, by itself, a productivity result. Expand only when the team trusts the process—not just the demo.

Before asking ‘Which AI platform?’, ask where project teams lose time, which record is authoritative and who is accountable when documents disagree. AI will not magically make a fragmented project record coherent. It can make a governed information set easier to navigate and give professionals more room for judgement. That is a useful first step: fewer people searching folders while the site waits.

Kirill Naumov is a consultant and entrepreneur with 16 years of experience across retail, construction, hospitality and other sectors, including 11 years at Apple. He taught marketing and sales to MBA participants for seven years. He helps organisations connect AI adoption with customer experience and the realities of frontline work.

Explore the AI for Retail workshop → · Discuss an AI diagnostic

Further reading: RICS Artificial intelligence in construction report 2025 · RICS responsible-use standard · NIST Generative AI Profile.

Start a conversation →