Before appointing anyone, ask for five things in writing: the current workflow, the baseline, the data and approval map, the work that must remain human, and the measures that will decide whether the first build continues or stops.
The gap between an impressive trial and a system that works inside a company is still wide. The RICS artificial intelligence in construction report 2025 surveyed more than 2,200 professionals globally, 48% of them in the UK. It found that 45% reported no AI implementation, 34% were still in early pilots and fewer than 1% had AI fully embedded across their organisation.
Those figures do not show a lack of interest. They show how hard it is to move from a useful demonstration to repeatable work, clear ownership and measurable value. That is the part your selection process needs to test.
1. Give them a workflow before they give you a platform
Start with work your business already performs. Enquiry to quote. Tender receipt to submission. Variation raised to approval. Valuation prepared to payment received. Compliance evidence requested to pack issued.
Ask the consultant to trace one route from beginning to end. They need to see the source documents, the people involved, the handovers, the rekeying, the exceptions and the point where the work usually waits. A diagram based on assumptions is not enough. The map needs to reflect what happens on a busy Tuesday when information arrives late or in the wrong format.
Construction fluency appears in the detail. An M&E contractor may lose time between supplier quotations, revisions and the final tender return. A quantity surveying practice may have a clean report template but a weak audit trail behind the figures. A property or FM team may hold the right compliance evidence across five disconnected locations. The consultant needs to identify the constraint in the actual process, not merely recognise the vocabulary.
If the first meeting revolves around agents, dashboards or a preferred technology stack before anyone has followed the work, the order is wrong. The starting point is the workflow, whether the answer becomes automation, a simpler process change or nothing at all.
2. Make them establish the baseline from your records
A credible proposal states what will change and where the current number came from. Depending on the workflow, that could mean:
- elapsed time from the trigger to the finished output;
- senior and administrative hours used each month;
- the number of revisions, exceptions or missing-information chases;
- the delay to a quotation, application, approval or payment; and
- the source record and person responsible for confirming the measure.
Estimates are sometimes unavoidable at the start. Their inputs still need to come from your business, and the uncertainty needs to be visible. A consultant who promises a percentage saving without inspecting the source records has supplied a sales number, not a baseline.
A useful readiness assessment leaves you with a short written case for the first workflow, the assumptions behind it and the evidence needed to confirm or reject those assumptions.
3. Ask for the data and approval map
Construction and built-environment companies run on commercially sensitive documents: quotations, drawings, contracts, applications, cost reports, RAMS, O&Ms, supplier emails, asset records and personnel information. Before any build, you need to know:
- which files and systems the workflow will read or write;
- whether personal or commercially sensitive data is involved;
- where the information is processed and retained;
- who can access it and how that access is removed;
- what is logged so errors can be traced; and
- which output requires a named person to approve it.
The Information Commissioner's Office guidance on AI and data protection covers lawfulness, transparency, data minimisation, security and accountability where personal data is processed. The UK government's AI Cyber Security Code of Practice also calls for human responsibility, documented data, models and prompts, appropriate testing, monitoring and proper disposal.
Ask how those obligations become controls in your workflow. “The vendor says the data is secure” is not a control. For some companies, the most valuable first step will be better document control and access discipline rather than a new assistant.
4. Find out what they refuse to automate
Ask the consultant to write down what the first version will not do. This exposes judgement faster than a list of features.
A tender price, contract notice, payment recommendation, safety decision or client-facing commitment may use automated preparation without allowing automatic issue. The right boundary depends on the work, the contractual position and the consequence of an error. The important distinction is whether the system drafts, recommends or acts.
You also need the fallback. If a source document is missing, a confidence check fails or the underlying service is unavailable, does the work stop visibly, route to a person or continue with a guess? A safe system fails loudly enough for the team to act.
5. Make them price the payback without double counting it
The commercial case needs to separate four things:
- Time returned: work genuinely removed from someone's week.
- Delay reduced: quotations, approvals, applications or evidence issued sooner.
- Value protected or won: margin, cash or revenue with a defensible link to the workflow.
- Total cost: implementation, licences, integration, training, support and ongoing monitoring.
Do not count the same benefit twice. Faster quotation preparation and extra senior capacity may describe the same returned hours. A payment arriving earlier improves cash timing, but it is not automatically a cost saving. The consultant needs to keep those effects separate and agree the review point before the build starts.
The numbers can be ranges. The inputs must be yours. You should leave with an expected result, a minimum acceptable result and a point at which the company stops spending because the evidence is not strong enough.
6. Check the plan for adoption and ownership
A working build still fails if it sits outside the way people do their jobs. Ask who owns the workflow after launch, which users are involved before it is finished, how exceptions are handled and who maintains it when a form, supplier format or internal process changes.
The first ninety days need named reviews. At thirty days, is the workflow being used on live work? At sixty, are the exceptions understood and the measures moving? At ninety, has it earned a permanent place, does it need another iteration or should it stop?
RICS recommends cross-functional leadership, practical training, data governance, human oversight and formal evaluation as companies move beyond pilots. A consultant who treats training and operating documentation as handover admin has left out the part that determines whether the investment survives.
7. Ask for evidence, including what went wrong
Request one relevant example with the starting position, the intervention, the measured result and what happened after the first few weeks. Confidential work can be anonymised without becoming vague. “We improved efficiency” is not evidence.
Then ask for a failure or refusal. What did they decide not to build? Which assumption proved wrong? What needed more human review than expected? A consultant with no example of a stopped or narrowed project may not have done enough difficult implementation work, or may not be willing to show you the full picture.
Finally, ask whether they receive commission or other benefit from any product they recommend. Vendor relationships are not automatically a problem. Undisclosed incentives are.
A practical selection scorecard
Score each line from zero to two. Ask the consultant to provide the evidence behind the score rather than scoring the presentation.
| Test | 0 | 1 | 2 |
|---|---|---|---|
| Workflow understanding | Generic demonstration | Basic map, no exceptions | Live process, sources, owners and exceptions traced |
| Baseline and payback | Broad benefit claims | Estimate with weak sourcing | Business records, assumptions, costs and review point |
| Data and security | Vendor reassurance | Partial controls | Documented access, retention, logging and risk ownership |
| Human approval | Unclear or absent | Review mentioned | Named gates, failure route and fallback agreed |
| Refusal and stop rules | Everything is possible | General cautions | Written no-go list and measurable stop conditions |
| Adoption and ownership | Technical handover only | Training promised | Named owner, user testing, documentation and reviews |
| Evidence and independence | Logos, features or tool claims | Relevant but incomplete example | Traceable result, failure lesson and incentives disclosed |
11 to 14: credible enough to progress to detailed due diligence. 7 to 10: ask for the missing evidence before appointing. 0 to 6: do not appoint on the current case.
This is a practical decision aid, not a guarantee of delivery. Procurement, legal, data-protection and professional obligations still need to be checked against the actual work.
When the right decision is not to hire anyone yet
Pause the engagement if the problem cannot be stated clearly, the source data is inaccessible, no one inside the business can own the workflow or the company is unwilling to measure the result. Also check whether a straightforward process change, an existing software feature or conventional automation solves the issue without introducing AI.
That is not a failed assessment. Avoiding an unnecessary build protects management time, reduces risk and leaves the budget for work that can earn its place.
What you should have after the first month
You should have one current-state workflow map, a baseline tied to source records, a data and approval map, a written decision on the first build and named measures for thirty, sixty and ninety days. You should also know what will stop the work.
The consultant's value is not making AI sound bigger. It is making the first decision smaller, testable and owned.