Chris M.29 August 2026

Competence, clients and the first ninety days

The industry does not need thousands of prompt engineers. It needs competent construction professionals who can tell when the machine is wrong, and firms that can explain to a client exactly how they would know.

What were the answers to the previous five questions?

1. What is the actual objective of AI training in a construction business, and how would you know it had been met?

That people can use AI while retaining control of work they are responsible for. You would know because they demand sources, notice missing context, recognise when consequence requires escalation, and can say why an output is wrong rather than that it feels wrong. Course completions measure none of that.

2. Why should training deliberately include AI answers that are wrong?

Because people who have only seen demonstrations where everything works will overestimate reliability. Give the system an old drawing revision, a contract with a known amendment, correspondence where one email changes the position, and let people find the failure. That builds the instinct no policy can.

3. Which AI uses should be disclosed to a client, and which need not be?

For RICS members and regulated firms this is not a judgement call. Where an AI system has a material impact on the delivery of a surveying service, the professional standard requires you to tell the client in writing and in advance when and for what purpose AI will be used, and requires the terms of engagement to set out where AI is involved, the extent of any professional indemnity cover, how the client can contest its use, how they can seek redress, and how they can opt out. Outside RICS regulation the graduated approach holds: Green use such as formatting and internal drafting needs no conversation, Amber use that materially informs advice a client relies on should be disclosed, and Red use always is, with the disclosure naming who checked the output and how. Check the contract as well, because a growing number now make disclosure a term.

4. What are professional indemnity insurers now asking about AI, and what document answers all of it?

Whether AI is used in delivering professional services, what controls exist over output before it is relied upon, who is accountable for AI influenced decisions, whether the policy has been updated, and whether there have been incidents. The register, with its classification and human control entries, answers every one.

5. If you did one thing in the next ninety days, what should it be?

The discovery exercise. It needs no procurement, no budget approval and no technology, it produces the register everything else depends on, and four weeks of conversations will do it.

What should training actually cover?

Not prompting. Models change, interfaces change, products disappear. The durable skills are the ones construction already values: critical thinking, source verification, evidence discipline, risk assessment, and knowing when something does not look right.

Role	What the training has to cover	The specific failure it prevents
Project managers	Programme data, correspondence, instructions, contractual terminology, decision logs, escalation	Wording that concedes a position, or an instruction issued outside authority
Quantity surveyors	CVRs, payments, variations, forecasting, subcontracts, evidence, delegated financial authority	A confident assessment built on a superseded or unamended contract
Designers and technical	Source documents, standards, revisions, specifications, design responsibility, compliance	A query answered from a revision superseded a fortnight ago
Site managers	Site records, photographs, progress, quality, safety, contemporaneous reporting	A generated record entered as fact without confirmation
Bid and pre-construction	Client information, tender requirements, assumptions, IP, capability claims	A plausible but fabricated project reference or statistic in a submission
Directors	Risk, governance, liability, data, supplier assurance, registers, oversight, dependency	A culture where speed is admired and nobody is rewarded for checking

The director row is the one most often skipped and the most consequential. A leader impressed by a demonstration that reviewed five thousand documents in thirty seconds, who then instructs the team to use it for everything, has created automation bias as company policy. People accept recommendations because leadership visibly trusts the system, and the human in the loop becomes a formality.

What do you say when a client asks?

Name the task, name the control, name the person. What separates a client who relaxes from a client who escalates is how much of that you can say without checking.

The version that creates doubt: we use some AI tools to help with the work, we wanted to flag it in case you had concerns. That puts the client in the position of deciding whether to be worried, and signals uncertainty whether or not any exists.

The version that builds confidence: we use AI assisted tools for this specific task, it does this specific thing, every output goes through this specific control before it reaches you, and here is how we govern it. Same underlying facts. The difference is whether you are explaining a system or confessing a habit.

Three things not to say. Do not overclaim control you do not have, because a statement that every output is checked by a senior professional becomes a misrepresentation the moment it is not true of your Green use. Do not imply AI itself assures accuracy; the control is the human review and the framework. And do not promise AI free delivery if your whole team uses Copilot for email, because you cannot keep it.

What about insurers and contracts?

Insurers are asking at renewal and contracts are starting to specify, and one document answers both. An increasing number of professional indemnity proposal forms now ask whether AI is used in delivering professional services, what controls exist over output before it is relied upon, who is accountable for AI influenced decisions, whether policies have been updated, and whether there have been incidents. Yours may not have asked yet, which is not the reassurance it sounds like: the duty to present the risk fairly applies whether or not a question is asked.

A contractor's position is not a consultant's. If you carry design liability under design and build, professional indemnity is where AI influenced design sits. If you do not carry meaningful cover, the exposure has moved rather than gone: an AI informed safety arrangement that failed reaches employers' and public liability, and reaches enforcement, where no policy responds at all. Take the register to your broker and ask which of your policies this touches. It is a thirty minute conversation.

Bring the register, or a summary derived from it. Insurers respond considerably better to here is our system than to a verbal assurance that nothing has gone wrong.

The legal point is worth getting right, because it is also the strongest argument for building the register at all. For business insurance the Insurance Act 2015 replaced the old duty to disclose material facts with a duty to make a fair presentation of the risk, and that duty expressly extends to what a reasonable search of your own business would reveal. Which is the register, exactly. A firm that has run discovery has made a reasonable search and can show it. A firm that has not is presenting a risk it has never looked at. The remedies are proportionate rather than all or nothing, so an insurer can only avoid outright where the failure was deliberate or reckless, but a proportionate reduction still arrives at the worst possible moment. If your presentation did not cover AI, raise it with your broker before renewal rather than firing an unadvised notification at the underwriter.

Contract clauses are appearing too, and they are worth having a position on before negotiating one live. Disclosure obligations. Restrictions on putting project or client information into AI systems without an enterprise data agreement. Liability allocation, some of which is drafted to place enhanced responsibility on the AI using party. Restrictions on client information being used to train models. Your register and framework are the evidence base for negotiating reasonable terms rather than either overpromising or refusing outright. Related: choosing an AI automation partner and what an AI admin system actually includes.

What does the first ninety days look like?

Discovery, then classification, then control, then one pilot. Nothing on the plan requires a platform purchase and the first thirty days require no technology at all.

A ninety day plan sized for a firm doing this alongside delivery. Nothing on it requires a platform purchase, and the first thirty days require no technology at all.

The sequence matters more than the dates. Discovery produces the register. The register makes classification possible. Classification makes proportionate control possible. Only then does a pilot make sense, because only then can you say which use case it is for, who reviews it, and what would count as success.

Two things run alongside from the start. A baseline on one process, captured before anyone knows a tool is coming, because that is the only window it exists in. And the disclosure position, drafted early, because the first client to ask will not schedule it.

What are the five questions you now answer for your own business?

These are the only five in the series with no answer arriving tomorrow. Answer them about your own business, in writing, and compare.

1. Where is AI currently influencing decisions in your business, and can you produce the list in writing today?

If the list is empty, the honest reading is that you have not looked rather than that there is nothing. Discovery closes this in four weeks.

2. For your three highest consequence AI uses, who is the named decision owner, and do they know?

A named role, not a named individual, so the control survives a resignation. Write it as a job title with a deputy, and tell the holder in the same sentence you tell them the classification. If nobody is named, the control is theatre, and it will be read as theatre by the first insurer who asks.

3. What does "human must verify" say for your most exposed use case, and would a tired reviewer know what to look at?

It should name the artefact rather than the quality: the signed and amended contract, the current revision, the approval status. Write it so the reviewer knows which document to open, not that they should feel careful.

4. What is your baseline on the one process you would automate first?

Elapsed time, labour time, systems touched, re-keying count, error rate and rework rate, captured on the process itself rather than from memory. If it was captured after the announcement it is not a baseline, it is a recollection with numbers attached, and every comparison built on it inherits the flaw.

5. If your largest client asked tomorrow how you govern AI, what would you send them?

A register with classifications, named owners, human control entries and a review cycle. If the answer is a policy PDF, you have described intentions rather than a system.

What does the whole argument come down to?

Thirteen taught parts, one argument: the technology is the easy part, and has been for some time. The hard parts are the knowledge layer, the authority mapping and the trust, and none of the three can be bought. Firms that build them will put AI into work that firms banning it will not go near, and will be able to defend the result to a client, an insurer and an adjudicator. Firms that build neither will find out what their people were relying on at the moment somebody asks them to prove it.

Automate the work. Augment the professional. Keep the accountability human.

Start with the register. It costs four weeks of elapsed time and something in the order of fifteen to twenty days of effort across the people involved, and it is the only step that reliably tells a firm something it did not already believe. AI training for construction teams and workflow automation are there if you want help with the rest. The four weeks you can start on Monday without them.

Which sources is this part built on?

Every figure quoted above resolves to one of these. Each was checked before publication.