The site manager has sent the photos. The delivery note is somewhere in the email chain. The office needs a job reference before it can match the paperwork, and the contracts manager is trying to work out what was actually agreed.
That is where I'd start with AI in construction. UK contractors have seen plenty of demonstrations of what a chatbot can write. What they need is a clearer route from work happening on site to the office knowing what to do next.
My view: use AI to help people find, extract and draft information. Keep decisions with the person who understands the job. And check whether the problem needs AI at all.
AI in construction UK: start with the handover
Picture a fictional contractor handling refurbishment and maintenance work. Its supervisors send job updates through WhatsApp, its office keeps an Excel tracker, and its accounts team works in Sage. This is an illustrative setup; the contractor is made up.
Suppose a supervisor reports that access was refused. The office needs to know which job was affected, who needs contacting and whether another visit should be arranged. The outcome that matters is a checked update reaching the right person.
I'd map that handover before choosing software:
- What information starts the work?
- Where should the job reference come from?
- What does the office need to decide?
- Who checks an unclear or incomplete update?
- What record shows that the next action happened?
If the missing piece is a required job-reference field, add the field. If the same notification must always go to the same person, consider a straightforward rule. Save AI for the parts that need interpretation, such as pulling a possible action out of a free-text note.
For a wider look at the underlying problem, read our article on the hidden admin cost inside a construction business.
Turn a long email thread into a checked briefing
A sensible place to test AI is a handover briefing for a contracts manager returning to a job. Ask for the issue, the proposed action, what remains unresolved and the messages supporting each point.
This capability already exists in a mainstream product. Microsoft's Outlook documentation describes Copilot summarising email threads and, in some summaries, providing citations back to the corresponding messages. Check the product and account requirements before buying anything; this is not a promise that every Outlook installation includes it.
My suggested test would use a completed, approved-for-use project thread whose outcome the reviewer already knows. Have the contracts manager compare the summary with the original messages. Look particularly for the difference between a suggestion and an agreement.
A useful briefing might say, “The supplier proposed an alternative finish; approval is not recorded in this thread.” It should not turn that into “The alternative finish was approved.”
Keep the source messages available. Do not let the summary become the only record, or let it send commitments to a client without review. I would measure the time taken to produce a correct briefing, including checking and corrections, rather than how quickly the first paragraph appears.
Extract delivery-note information without handing over payment decisions
Consider another illustrative workflow: a delivery-note PDF arrives by email and an administrator needs its supplier, reference, delivery details and line items in a review queue.
Microsoft's AI Builder documentation describes custom document-processing models that extract fields, tables and checkboxes. That supports testing document extraction. It does not prove a model will read your supplier's notes accurately, match them to the right order or work with your accounts system without additional setup.
I would separate the work into distinct steps: extract candidate information, check it against the original, match it to the relevant order, then let an authorised person decide what happens next. Copying a value and approving a transaction are different jobs.
Keep the original document beside the extracted information. Ask the reviewer to check supplier identity, quantities, units and job allocation. If a reference is missing, the workflow should visibly stop for clarification rather than manufacture a plausible answer.
Also test the awkward documents alongside the clean PDF used in a demonstration. Include a handwritten amendment, an unclear scan and a note covering materials for different jobs. Those are suggested test cases; they say nothing about any product's performance.
If structured supplier data is already available, I'd investigate importing that before teaching a model to read a document. The simplest reliable route wins.
Find project information, but insist on the source
Another candidate is searching a controlled set of project documents. The task might be to find where a particular finish is specified, then take the user to the relevant passage for checking.
Procore's own description of Procore AI says users can search project data and analyse documents, with source citations and proposed changes staged for human review. That is Procore describing its own product, so its accuracy still needs testing, and I'm not recommending a move onto Procore.
If you already use it, ask for a demonstration using documents you are allowed to test. If you do not, I would not buy a broad construction platform just to solve a narrow search problem without assessing the alternatives.
My acceptance test would be practical: can the reviewer open the supporting document, identify its revision and confirm that it answers the actual question? Add a deliberately unanswerable question. The desired result is a clearly flagged gap.
Keep superseded material clearly separated from current working information. I would not use an AI search result as permission to change a design, substitute a material or issue an instruction. Finding evidence is assistance; deciding what it means for the work stays with the responsible person.
Draft updates from notes, without inventing the missing detail
For a contractor, I would test a checked progress update before attempting autonomous estimating or programme decisions. The boundary is easier to explain: reorganise the supplied notes, preserve uncertainty and leave gaps visible.
Take this fictional supervisor's note: “Ceiling opened, leak visible, awaiting instruction, photos attached.” A useful draft would keep those observations separate from whatever the office needs to ask next. It should not add a cause, claim the leak was repaired or promise a return visit.
The person responsible for the update should compare the draft with the note and the relevant records before issuing it. Keep observations, assumptions and proposed actions distinct.
I would use a template with prompts for:
- Work recorded as completed.
- Issues reported and their supporting records.
- Information still missing.
- Decisions requested and their owners.
That is a proposed workflow. I'm making no claim that any named product will produce a reliable report automatically. Test the actual tool against your own approved examples.
If a short digital form captures everything clearly enough, use the form instead. There is no prize for inserting a language model between a clear note and a clear update.
Give the trial a stop rule and an owner
Pick a narrow workflow and write down what would make the trial worth keeping. I'd include the total handling effort, corrections required, unresolved items and whether the next person receives usable information. Compare those with the current process using similar work, rather than a polished demonstration.
Name the operational owner and the reviewer. Decide where rejected outputs go and how the team returns to the existing process if the trial disappoints. Do not make the busiest administrator responsible for quietly repairing everything behind the scenes.
This week, sit with the person who handles an awkward handover. Follow a completed job from the original message to the final record. Mark each point where somebody searched, retyped, chased or guessed. Choose a problem from that map, then decide whether a clearer process, a simple integration or AI is the right response.
That is how I would approach Assess. The public stages are Assess, Design, Implement, and Operate & Optimise; each is agreed separately and has to earn the next.
AI has to earn its place in the working day, and looking useful in a demonstration doesn't count. Start with the operation; use technology only where it helps.
If you want to work through a stubborn site-to-office handover, let's talk.