Everyone in construction is talking about AI. Almost nobody knows where to start.
Construction does not have an AI awareness problem. It has an implementation problem.
RICS surveyed more than 2,200 construction professionals globally in the first quarter of 2025. 45% had no AI implementation at all. 34% were piloting. Just under 12% used it regularly in one process. Under 1% had it embedded organisation wide.
Now look at the profession furthest ahead. RIBA's 2026 AI report found 74% of architectural practices using AI on at least some projects, up from 59% in 2025 and 41% in 2024. Three quarters report productivity gains. 57% report positive return on investment.
And 17% say their designs are actually better.
That is the whole story in one number. Faster, cheaper, no better. It is what happens when you buy the tool before you understand the problem.
Do we actually need AI?
The question sounds heretical in 2026. Ask it anyway.
Sometimes a business needs AI. Sometimes it needs automation. Sometimes it needs better data. Sometimes it needs its SharePoint sorting out. Sometimes it just needs to stop paying someone to copy figures from one spreadsheet into another. The gap between firms that have adopted AI well and firms that have not is rarely nerve or budget: more often it is whether the information a tool would need even exists anywhere a system can read it.
Put Copilot licences into a broken process and you get a faster broken process.
What eight questions should you ask before buying anything?
1. Where are we losing time, money, information or certainty right now?
2. Where is AI already being used without us knowing? ChatGPT. Copilot. Email drafting. Meeting summaries. Tender responses. That is shadow AI and it is already in your business.
3. What problem are we actually solving? Not "should we use AI" but "why does this report take four hours?"
4. Is AI the answer? It might be an API, Power Automate, a database or a better workflow.
5. What information would it need? Drawings, contracts, emails, the CDE, cost data, programme, site photographs.
6. What happens when it gets something wrong? Who carries that.
7. How will we know it worked? Set the baseline first. Four hours to forty minutes. Three days preparing a CVR to half a day.
8. Who owns AI in this business? Not IT. Someone accountable for process, governance, training and measurable outcomes.
That is a shortened version of a longer method. The full construction automation audit walks each task in the business against six criteria and a judgement test, and asks the question most audits skip: what happens to the quality of decisions, not just the speed of them.
How do the answers turn into a plan?
Sort every candidate into four bands, and be honest about which one each idea actually belongs in.
| Band | What it means | Typical candidate |
|---|---|---|
| Now | Low risk, measurable, quick to implement | Drafting first-pass correspondence, summarising long document trails, meeting notes into actions |
| Next | Needs integration, workflow redesign or better data first | Automated reporting that pulls from several systems, a structured site record |
| Later | Agentic systems, predictive tools, deeper document intelligence | Programme risk prediction, autonomous document review at scale |
| Don't | Liability, data quality or lack of oversight make AI wrong for now | Anything touching a regulated sign-off with no named reviewer, anything built on records nobody trusts |
Across all four bands, the same six things need deciding: governance, data, security, human oversight, training and measurement. Being willing to put something in the Don't column, and mean it, is what makes the other three columns credible. A thirty day pilot is the right size for testing a Now candidate; it is the wrong size for finding out whether a Later candidate was a good idea.
What have the professional bodies already told the industry?
RICS made its Responsible Use of AI in Surveying Practice standard mandatory on 9 March 2026, for work where AI has a material impact on the service. The surveyor stays accountable for the output regardless of the tool that produced it.
CIOB's AI Playbook set out a twelve step implementation strategy in June 2024. It starts with what AI can do, what the business needs, what capability exists, what can be bought, and what has to be built.
McKinsey's State of AI research found nearly nine in ten organisations using AI in at least one function, with most reporting no significant effect on enterprise profit. The organisations seeing real impact redesigned how the work happens. They did not bolt AI onto what was already there.
So where should we start?
Don't start with the tool.
Start with your business. Start with the work people hate doing. Start with information that keeps getting lost. Start with the process that takes three hours because nobody has questioned it in ten years. Start with duplicated admin, commercial leakage, slow decisions and repetitive reporting.
Then ask whether AI helps.
The first AI decision your business makes should not be which AI to buy. It should be which problem is worth solving.