The Construction AI Syllabus

Most writing about AI in construction is either a vendor pitch or a warning. This is neither. It is a course, with a sequence, an evidence base, and an assessment at the end of every part.

Why does construction need a syllabus rather than another opinion piece?

Because the industry has moved past the stage where the interesting question is whether AI can read a subcontract. It can. The interesting questions now are who signs the assessment it produced, what happens when it worked from a superseded revision, and how a firm proves any of that eighteen months later in an adjudication.

Those are governance and competence questions, and they are answerable. They are also cumulative: you cannot sensibly classify AI use by consequence until you know where it is being used, and you cannot design human oversight until you have classified it. A sequence of essays published in whatever order they were written does not teach that. A syllabus does.

The audience is deliberately wide. A graduate quantity surveyor, a site manager, a design lead, a commercial director, and an owner of a twelve person subcontracting business all have a stake in this, and each needs a different part of it most. Read out of order and each part still stands up. Read in order and the matrix in part 6 is built on the classification in part 5, which is built on the register in part 4.

What does the evidence say about where the industry actually is?

It says the gap is not between construction and technology. It is between talking about AI and doing anything durable with it. Surveyed separately, 69 per cent of project managers and 67 per cent of quantity surveying professionals expect AI to help surveyors deliver greater value. On separate evidence drawn from construction specifically, almost nobody works somewhere that has built the conditions for that to happen.

That combination is unusual and it is the reason for the sequence below. Where the constraint is enthusiasm you run a demonstration. Where the constraint is that nobody can say which decisions currently depend on a model output, a demonstration achieves nothing at all.

How is the syllabus structured?

Four modules across thirteen taught parts. The order is a dependency chain, not a preference.

How should you work through it?

Read the executive summary, then the parts, then attempt the five questions before the next part arrives, and mark yourself honestly against the answers at the top of the following part. The questions are not comprehension checks. Each one has a defensible answer that the following part sets out, and several have answers that will be specific to your business rather than general.

Two habits make the difference. First, answer the questions against your own organisation rather than in the abstract: "who approves a variation over fifty thousand pounds here" is a better use of five minutes than agreeing that someone should. Second, keep a note of every place where you cannot answer. That list is the beginning of the register built in part four.

If your firm is already further along than most, the syllabus still assumes nothing. It is common to find a firm with an AI policy, no register, and no idea which of its people are pasting subcontract clauses into a consumer tool. Related reading while you wait: what a proper AI audit looks like, controlled adoption rather than blanket bans, and what a good AI pilot looks like.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. What proportion of UK construction businesses use any AI at all, and how does that compare with the economy as a whole?
2. What is the difference between the breadth of AI adoption and its depth, and which one is construction short of?
3. Which three barriers to adoption do construction professionals themselves rank highest?
4. Why is enthusiasm among professionals a poor predictor of whether a firm will get value from AI?
5. If a firm wanted one honest measure of how far it has actually got, what would it be?

Which sources is this part built on?

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

• ONS, Artificial intelligence in UK businesses: 2023 to 2026
• RICS, Artificial Intelligence in Construction (2025)
  Caution: Three separate traps. First, the adoption depth figures and the sentiment figures come from two different surveys; do not present them as one dataset, and do not round 67 and 69 into a single "about 70 per cent of the profession". Second, the skills survey is global and covers all surveying disciplines, so its figures are not UK construction figures. Third, these are cross-sectional categories from one survey, not a cohort, so nothing in them shows a pilot "becoming" production.
• RICS, Construction Productivity Report 2026
  Caution: The 0.4 per cent figure is McKinsey 2024 quoted by RICS, not a RICS finding. Attribute it to McKinsey via RICS or cite McKinsey directly. The 17 per cent attaches to automation specifically, not to automation and digitalisation jointly.
• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.

Almost every UK contractor of any size has now tried AI. Almost none have put it anywhere near a decision that matters. That distinction, not the adoption rate, is the thing worth measuring.

What were the answers to the previous five questions?

1. What proportion of UK construction businesses use any AI at all, and how does that compare with the economy as a whole?

Around 13 per cent of construction businesses with ten or more employees, against around 35 per cent across all UK businesses of that size and 58 per cent in information and communication. Construction is under a quarter as likely to use AI as the leading sector.

2. What is the difference between the breadth of AI adoption and its depth, and which one is construction short of?

Breadth is how many firms have touched it. Depth is how much of the work actually runs through it. Across UK businesses breadth has roughly tripled since 2023, and ONS publishes no construction time series, so the 13 per cent is a single point estimate rather than a trend. Depth has barely moved: under 1 per cent of construction organisations report AI embedded organisation wide. Construction is short of depth.

3. Which three barriers to adoption do construction professionals themselves rank highest?

Shortage of skilled personnel at 46 per cent, systems integration at 37 per cent, and data quality or availability at 30 per cent. Notably, none of the three is solved by buying software.

4. Why is enthusiasm among professionals a poor predictor of whether a firm will get value from AI?

Because enthusiasm sits with individuals and value requires organisational plumbing: a controlled knowledge source, a defined reviewer, a delegated authority table, and a record. Two thirds of RICS members surveyed expect AI to help surveyors deliver greater value. Separately, and among construction professionals specifically, almost none work somewhere that has built any of that.

5. If a firm wanted one honest measure of how far it has actually got, what would it be?

Count the live business decisions that currently depend on an AI output, and name the accountable role for each, with the person currently holding it. Zero means the firm is experimenting, whatever its policy says. More than zero with no names is worse than zero, because the exposure exists and the accountability does not.

What do the adoption numbers actually show?

They show an industry that has tried AI widely and deployed it almost nowhere. The Office for National Statistics puts construction at 13 per cent, well behind the 35 per cent average for UK businesses with ten or more employees (ONS, 2026).

That figure alone invites the usual conclusion, which is that construction is slow. The RICS data, drawn from the profession rather than from businesses generally, suggests something more specific and more useful.

Read down that funnel. Forty five per cent report no implementation at all. Thirty four per cent are piloting. Just under twelve per cent use AI regularly in one specific process. One and a half per cent use it across more than one. Under one per cent have it embedded organisation wide (RICS, 2025).

The interesting cliff is between 34 and 12. A third of the industry is piloting, and the group using AI routinely in even one process is barely a third of that size. These are snapshot categories rather than a cohort tracked over time, so nothing here shows a particular pilot failing. The shape is still unambiguous: the piloting population is far larger than the producing one, and pilots stall for reasons that have nothing to do with the model.

Why do the pilots stall?

Because the barriers professionals report are structural, and a pilot is specifically designed not to encounter them.

The ranking matters. Regulatory uncertainty comes last, at 11 per cent. The three that lead are all about the state of the organisation rather than the state of the technology, which means the constraint is internal and therefore fixable without waiting for anyone.

What is the cost of standing still in this particular industry?

Higher than in most, because the underlying margins are thinner and the failure modes are expensive.

Construction recorded 3,805 company insolvencies in England and Wales in the twelve months to June 2026, the highest of any industry (Insolvency Service). The Get It Right Initiative puts directly recorded rework at around 5 per cent of project value, and total cost of error, once latent defects and process waste are counted, at 10 to 25 per cent (GIRI). CITB forecasts a need for roughly 41,200 additional workers a year to 2030 (CITB).

Those three facts describe a sector with no spare margin, a large avoidable error bill, and no prospect of hiring its way out. That is the case for doing this properly, and it is a stronger case than any efficiency claim. It is also why the real cost of doing nothing is rarely on anyone's line item, and why the admin tax on small contractors is paid without ever being costed.

One further number explains why the argument stalls. RICS found that roughly one in five UK construction firms never measure productivity at all, and that UK respondents were the most sceptical of five regions about automation, with 17 per cent rating it a high impact intervention (RICS, 2026). A sector that does not measure has no way to be persuaded by evidence, which explains a good deal about how these conversations go.

How do you place your own firm honestly?

Answer four questions in writing, today, without consulting the policy document.

1. Which live decisions currently depend on an AI output? Not which tools are licensed. Which decisions.
2. For each of those, which role is accountable, who currently holds it, and do they know?
3. Where does the AI get its information: the actual contract and current revisions, or general model knowledge?
4. If a client asked tomorrow how you govern AI, what document would you send?

Most firms find the first question harder than expected, because the honest answer includes things nobody procured. That is not a failure of discipline. part 2, the AI in your business that nobody procured is about that gap, and it is the widest one between what a construction business believes about its AI use and what is actually happening inside it.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. Your firm has no AI policy and no procurement record for any AI tool. Does that mean AI is not being used?
2. Why is enterprise grade data security an incomplete answer to the question "is our AI use safe"?
3. A project manager asks a model to make a delay email "more professional and contractual". What has changed about that email that matters?
4. What is the difference between an AI system making a decision and an AI system influencing one, and which is more common?
5. Why is the first governance exercise a discovery exercise rather than writing a policy?

Which sources is this part built on?

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

• ONS, Artificial intelligence in UK businesses: 2023 to 2026
• RICS, Artificial Intelligence in Construction (2025)
  Caution: Three separate traps. First, the adoption depth figures and the sentiment figures come from two different surveys; do not present them as one dataset, and do not round 67 and 69 into a single "about 70 per cent of the profession". Second, the skills survey is global and covers all surveying disciplines, so its figures are not UK construction figures. Third, these are cross-sectional categories from one survey, not a cohort, so nothing in them shows a pilot "becoming" production.
• RICS, Construction Productivity Report 2026
  Caution: The 0.4 per cent figure is McKinsey 2024 quoted by RICS, not a RICS finding. Attribute it to McKinsey via RICS or cite McKinsey directly. The 17 per cent attaches to automation specifically, not to automation and digitalisation jointly.
• Insolvency Service, Company Insolvency Statistics, June 2026
  Caution: This series is published monthly and the figure is superseded roughly every four weeks. Always state the twelve month window in the sentence so the claim ages into a dated fact rather than a wrong one.
• Get It Right Initiative, Financial and economic impact of error (2016 research)
  Caution: The 5 per cent and 21 per cent figures measure different things. Never present 21 per cent as recorded rework.
• CITB, Construction Workforce Outlook 2026 to 2030

A firm can have no AI strategy, no AI procurement and no AI budget, and still have AI shaping its project records every week. The question is not whether to allow it. That decision has already been taken, quietly, by people trying to do their jobs.

What were the answers to the previous five questions?

1. Your firm has no AI policy and no procurement record for any AI tool. Does that mean AI is not being used?

No. It means AI use is unrecorded. Copilot ships inside Microsoft 365, AI features arrive inside estimating, design and document management software by update, and consumer accounts need no purchase order. Absence of procurement evidence is evidence of absent governance, not absent use.

2. Why is enterprise grade data security an incomplete answer to the question "is our AI use safe"?

Because it answers only where the data went. An enterprise secure system can still summarise forty emails and miss the one that changes the commercial position, work from a superseded revision, or produce fluent wording that concedes something. Security and correctness are separate problems with separate controls.

3. A project manager asks a model to make a delay email "more professional and contractual". What has changed about that email that matters?

The author knew what they meant. The redraft contains wording chosen by a language model: which programme, whether it was agreed, whether causation has been established. Making language better and making a project record correct are different operations, and only one of them was requested.

4. What is the difference between an AI system making a decision and an AI system influencing one, and which is more common?

Making means the output is actioned without meaningful human involvement. Influencing means a human decides, but some of the reasoning in front of them originated with a model. Influence is far more common, far harder to see, and currently far less governed.

5. Why is the first governance exercise a discovery exercise rather than writing a policy?

Because a policy written without knowing what is happening governs an imaginary organisation. You cannot classify by consequence, assign owners or set proportionate controls over use cases you have not found.

What does shadow AI actually look like on a live project?

It looks like ordinary competence. A quantity surveyor pastes a subcontract clause in to understand it faster. An estimator asks a model to interrogate a 180 page specification because reading it properly costs half a day. A bid manager improves a methodology. A design coordinator summarises a technical submission. None of it feels like implementing anything.

The top path is the one every governance framework anticipates: business case, due diligence, deployment, register entry, review cycle. The bottom path has no gate anywhere along it. By the time an output has become an email, a report or a commercial assessment, it is a project record, and the organisation carries it.

Why is the rewritten email the sharpest example?

Because it is the smallest possible intervention with the largest possible consequence, and because everybody has done it.

None of that is a hallucination. The model did precisely what it was asked. The employee asked for better prose and received, alongside it, a contractual position. That is why "check the AI output" is too vague an instruction to be a control: checked against what, by whom, for which risk?

Is this now a regulatory matter rather than a matter of taste?

For surveying firms, yes. The RICS professional standard on the responsible use of artificial intelligence took effect on 9 March 2026 and is mandatory for members and regulated firms where AI has a material impact on the delivery of surveying services. It expressly recognises that AI becomes involved in producing professional work both knowingly and unknowingly, and requires firms to keep a written record of AI use and a risk register (RICS).

The wider industry is not formally bound by that standard, but the direction is unambiguous, and the same logic reaches anyone carrying a duty. A principal designer or principal contractor under CDM 2015 holds their duties personally and organisationally. Nothing about a model participating in the reasoning redistributes them.

What should you do about it this week?

Nothing punitive. The single most effective change is to the question you ask, and the discovery exercise in part 4 is built on it.

Do not ask whether anyone is using unauthorised AI. You will get denial, or one minor confession, and neither gives you the picture. Ask instead where AI is currently helping people do their work, and be precise about what you are promising. You are promising that nobody is in trouble for telling you. You are not promising that everything continues unchanged, because some of what you find will need controls around it, and an assurance you break in week four is worse than one you never gave.

Say what the amnesty covers and what it cannot. It covers the fact of use: nobody is disciplined for telling you they have been using a tool the firm never gave them. It cannot cover an actual data breach or a breach of client confidentiality, because those carry notification duties the firm does not control, and a personal data breach starts a seventy two hour clock whether or not you would rather handle it gently. Brief interviewers on two escalation triggers: personal data or client confidential information in an unmanaged account, and anything touching a live dispute or a statutory submission. On either, stop, say plainly that this one goes to the data protection lead today and that the escalation is about the information rather than the person, and route it the same day.

It is also worth being clear internally that this is not a character issue. People reach for these tools because the alternative is reading 180 pages on a Thursday evening. The pull is real and it is not going to weaken. See also shadow AI in the workplace and controlled AI adoption rather than blanket bans.

That leaves an obvious objection, and it is the one directors raise first. If AI can now read the contract, challenge the valuation and interrogate every record on the project, what exactly is left for the professional? part 3, AI can read the contract. It cannot become the dutyholder answers it, and the answer is not sentimental.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. Under CDM 2015 and, in England, Part 2A of the Building Regulations, who can hold a dutyholder role, and could that ever be a software system?
2. What is the difference between work that AI can perform and responsibility that AI can carry?
3. Which parts of a project manager's week are genuinely at risk from automation, and which are not?
4. If a firm automates every routine judgement a graduate used to make by hand, what happens to their ability to catch the machine when it is wrong, and when does that bill arrive?
5. Why does an approval button on a screen not necessarily constitute human oversight?

Which sources is this part built on?

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

• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.
• RICS, RICS first-ever standard on responsible AI use now in effect
• HSE, Construction (Design and Management) Regulations 2015
• ICO, consultation on draft guidance about automated decision-making, including profiling
  Caution: Published 31 March 2026. The consultation closed on 29 May 2026 and the guidance remains in draft pending the final version, so cite it as draft guidance whose consultation has closed, never as settled law. Re-check the status before quoting it. Note also that section 80 does not define meaningful human involvement: Article 22D reserves that to regulations the Secretary of State has not yet made.

The popular version of this argument is that machines cannot replace human intuition. That is weak, and it is losing. The strong version is legal: UK construction requires a named person to carry the duty, and a model cannot be named.

What were the answers to the previous five questions?

1. Under CDM 2015 and, in England, Part 2A of the Building Regulations, who can hold a dutyholder role, and could that ever be a software system?

Dutyholders are commercial and domestic clients, designers, principal designers, contractors, principal contractors, and workers: identifiable people or organisations, with skills, knowledge, experience and organisational capability requirements attached under CDM, and express competence requirements under Part 2A in England. A software system is none of those things. It cannot be appointed, cannot demonstrate experience, and cannot be held to account.

2. What is the difference between work that AI can perform and responsibility that AI can carry?

Work is the assembling, comparing, extracting and drafting. Responsibility is the acceptance of consequences for a judgement. The first is fully transferable to a machine. The second is not transferable at all, which is why capability and accountability have to be governed separately.

3. Which parts of a project manager's week are genuinely at risk from automation, and which are not?

At risk: the finding, comparing, transcribing and assembling. Not at risk: deciding what a programme movement means, whether an instruction is justified, what to escalate, and carrying the consequences of that. The dividing line is not difficulty. It is whether being wrong creates a liability somebody has to answer for.

4. If a firm automates every routine judgement a graduate used to make by hand, what happens to their ability to catch the machine when it is wrong, and when does that bill arrive?

They lose it, and the bill arrives at the abnormal case, which is the only case they were retained for. Lisanne Bainbridge named this the ironies of automation in 1983: automating routine work removes the human from ordinary practice while still expecting competent intervention when the situation is not routine.

5. Why does an approval button on a screen not necessarily constitute human oversight?

Because oversight requires the reviewer to see the reasoning, the sources, the assumptions and the conflicting information, and to have the authority and competence to disagree. A button records assent. It does not evidence involvement, and where the decision is a significant one about an individual, that distinction now has consequences in law.

What does the statutory framework actually require?

It requires named parties. Under CDM 2015, which applies across Great Britain, commercial and domestic clients, designers, principal designers, contractors, principal contractors and workers each carry defined duties. Regulation 8 requires a designer or contractor appointed to a project to have the skills, knowledge and experience, and if they are an organisation the organisational capability, necessary to fulfil the role, with a matching duty on whoever appoints them to satisfy themselves of it. For building work in England, Part 2A of the Building Regulations 2010, inserted on 1 October 2023, goes further and places an explicit duty on a principal contractor to plan, manage and monitor the building work and coordinate matters relating to it so that it complies with all relevant requirements. Wales, Scotland and Northern Ireland have their own arrangements.

Note the reach, because it matters outside buildings. Part 2A bites on building work as the Building Regulations define it, which on a rail, highway or water scheme means the station or the control building rather than the viaduct, the tunnel or the treatment stream. The duty that reaches all of it is CDM, and its sharpest provision is regulation 8: nobody may accept an appointment as a designer or contractor unless they hold the skills, knowledge, experience and, as an organisation, the capability the work requires, and whoever appoints them must take reasonable steps to satisfy themselves of it. That test can only be applied to a person or a company.

The same architecture recurs wherever consequences are physical. A temporary works coordinator is appointed by name under BS 5975 by a designated individual who is themselves named. A panel engineer under the Reservoirs Act 1975 must be on a specified panel and nobody else may certify. A chartered engineer signing a design answers personally to an institution that can remove their title. None of those appointments can be held by a system.

The direction since Grenfell has been towards more clearly identified accountability, not less. The Building Safety Act 2022 added a gateway regime and a golden thread of information, both of which turn on somebody being answerable for what was decided and why.

Set that beside a technology that can now read every document on a project in an afternoon, and the tension resolves cleanly rather than dramatically. The machine takes the work. The duty stays put.

Does that mean nothing changes for the professions?

No. It means the composition of the roles changes substantially while the roles persist.

This is consistent with the broader evidence. The ILO and NASK index of occupational exposure to generative AI finds around one in four workers in occupations with some exposure and about 3.3 per cent in the highest band, and is explicit that exposure means task transformation rather than job replacement (ILO, 2025). The ONS finds most UK businesses reporting no change to headcount from AI at all (ONS, 2026).

What is the actual risk, if not job losses?

That the person left holding the duty is no longer practised enough to discharge it.

Lisanne Bainbridge set this out in 1983, long before anything resembling a language model existed (Ironies of Automation90046-8)). Automate the routine, and you remove the human from the ordinary case while still expecting expert intervention in the abnormal one. The more reliable the automation, the less recent the human's experience, and the moment they are needed is precisely the moment they are least ready.

Applied to a graduate quantity surveyor: every difficult clause goes into a model, every awkward email gets redrafted, every long document gets summarised rather than read. Productivity rises immediately. Five years later, the question is whether that person is an exceptional QS equipped with AI, or someone highly capable at operating AI who never built the judgement required to challenge it. Those are very different professionals and only one of them can sign anything.

RICS has taken the same position in regulatory language: members using AI in material professional work must understand its limitations and failure modes, assess reliability, apply professional judgement and maintain oversight. The surveyor's judgement, not the system output, sits at the centre.

How should this change what a firm builds?

It gives you a design rule that is more useful than any policy sentence: before automating a workflow, decide where the human remains and why.

• Some activities can be fully automated, because being wrong is cheap and the error is visible immediately.
• Some should be assisted, with the professional inside the reasoning rather than receiving its conclusion.
• Some should produce a recommendation and stop, requiring an explicit decision.
• Some should escalate automatically, because the value or the risk crosses a threshold.
• Some should remain untouched, specifically so that the organisation retains people who can still do them.

That last one is counterintuitive and it is deliberate. Preserving a small amount of human effort where a machine could do the job is not inefficiency, it is competence maintenance, and it is the direct answer to Bainbridge. Related: why AI will not replace construction jobs and what an AI admin system actually includes.

A slide widely circulated as IBM internal training material from 1979 put it more briskly: a computer can never be held accountable, therefore a computer must never make a management decision. The original document has never been located in IBM's archive, so treat it as a well travelled aphorism rather than a citation. The reasoning holds regardless of who wrote it.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. What is the first thing a construction business should build: an AI policy, or something else?
2. Why should a register record a system and a use case together rather than just naming the tool?
3. What single question, asked of an employee, reveals the consequence of a given AI use faster than any other?
4. What should you do with the workarounds people describe during a discovery exercise?
5. Why does a register that grows over time indicate success rather than deterioration?

Which sources is this part built on?

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

• HSE, Construction (Design and Management) Regulations 2015
• The Building Regulations 2010, Part 2A: dutyholders and competence
  Caution: England only. CDM 2015 applies across Great Britain; Part 2A does not. Never present the two as co-extensive. Competence is the Part 2A term; CDM 2015 deliberately uses skills, knowledge, experience and organisational capability instead.
• Building Safety Act 2022
  Caution: The Act is the enabling statute. The operative procedure, including change control, is in the Building (Higher-Risk Buildings Procedures) (England) Regulations 2023. The regime is England only: Wales operates separately and Scotland has building warrants and no gateways.
• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.
• Bainbridge, L. (1983), Ironies of Automation, Automatica 19(6), 775 to 779
• ILO and NASK, Generative AI and Jobs: a refined global index of occupational exposure
• ONS, Artificial intelligence in UK businesses: 2023 to 2026
• Simon Willison, A computer can never be held accountable
  Caution: Cited for the provenance, not for the quotation. IBM’s own commentary pages state the 1979 attribution without qualification, so they cannot evidence a claim that the attribution is unconfirmed. Attribute the line as a widely circulated slide, never as a verified corporate publication. The idea stands on its own merits.

Discovery costs four weeks and one uncomfortable decision: that nobody gets into trouble for what they report. Get that right and the register is accurate. Get it wrong and every control built on top of it sits on a false picture.

What were the answers to the previous five questions?

1. What is the first thing a construction business should build: an AI policy, or something else?

A register, built from a discovery exercise. Policy is the cheap part and can be written in an afternoon once you know what you are governing. Discovery is the part that takes four weeks and cannot be skipped.

2. Why should a register record a system and a use case together rather than just naming the tool?

Because the risk lives in the pairing. Copilot summarising an internal meeting and Copilot analysing contractual correspondence share a product name and nothing else: different data, different consequence, different control. A single row saying "Microsoft Copilot" governs nothing.

3. What single question, asked of an employee, reveals the consequence of a given AI use faster than any other?

"What happens if it is wrong?" It cuts past the description of the tool straight to the exposure, and people answer it accurately because it is a question about their work rather than about their compliance.

4. What should you do with the workarounds people describe during a discovery exercise?

Record them prominently. A workaround marks a place where the formal process is failing badly enough that a busy person built their own route around it. That is both a risk and the best automation candidate list you will get.

5. Why does a register that grows over time indicate success rather than deterioration?

Because growth means people are reporting new use rather than concealing it. A register frozen at its first count is not a stable organisation, it is one where nobody trusts the process enough to add anything.

What does a discovery exercise involve?

Four weeks, a handful of interviewers, and a decision at the outset to treat this as a survey rather than an audit. The distinction is not presentational. If people believe that admitting to using a consumer tool will get them into trouble, they will not tell you, and you will build your entire control framework on a false picture.

Nominate interviewers people already trust inside each department. Do not send compliance to ask a design team about AI. Book twenty minutes, ideally at a point in the week when the person has recently done the work rather than at four o'clock on a Friday, and ask six questions.

1. What AI system do you currently use at work?
2. What task does it save you from doing?
3. What information do you give it?
4. What do you use the answer for?
5. How do you check it has given you the right answer?
6. What would happen if it got it wrong?

What goes into a register row?

One row per meaningful pairing of a system with a use case. This is the AI register, and it is the artefact every later part refers back to. Start it in a spreadsheet: visibility first, database later, and if you find twenty three use cases then you have twenty three rows and no rounding.

For RICS regulated firms this is not optional housekeeping, and it is two artefacts rather than one. The professional standard requires a written register of AI systems that have a material impact on service delivery, recording the purpose, the date of first use and the date of next review; and, separately, a risk register carrying a red, amber, green rating that must be reviewed and updated at least quarterly by the staff responsible for decisions about the firm use of AI (RICS). This register is the first of those. A firm already running a corporate risk register should cross reference into it rather than duplicate. Firms outside RICS regulation should expect their clients and their insurers to arrive at the same requirement by a different route.

What do people actually report when you ask properly?

A predictable spread, and one or two things that stop the room.

• Meeting transcription and summarisation, which spreads fastest because it needs no decision from anyone and shows a visible saving on day one, and which almost never has an owner.
• Email drafting and refinement, including on correspondence that is a project record.
• Document interrogation: specifications, subcontracts, technical submissions.
• AI features that arrived inside existing software by update, which people do not describe as AI at all.
• Personal subscriptions being used for work, on work information, on a personal account.

That last category is the one worth handling carefully. It carries the highest data risk on the list, because work information sits on an account the firm cannot see, cannot search and cannot delete, and it exists because no approved route does. Treating it as misconduct produces a smaller register and the same behaviour. Treating it as unmet demand produces an approved route and a register you can believe.

How do you keep it from becoming a dead document?

By attaching a rhythm and a use to it on the day it is created.

And pick three quick decisions in week four, once the classification in part 5 is in hand: one Green use case to approve and publicise, one Amber to add guidance to, one Red to route through a named approver. That demonstrates within a month that governance widens what people are allowed to do rather than narrowing it, which is the only argument that will keep them reporting honestly. See what a proper AI audit looks like and what an automation consultant actually does for the adjacent work.

You now have a list of real uses with real consequences. part 5, classify by consequence, not by which logo is on the screen turns that list into a control framework, and the two variables that set it are what it costs to be wrong and how late you would find out.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. Two people use the same AI product. One summarises internal minutes, the other drafts contractual notices. Should they carry the same controls, and why?
2. What two variables should determine whether a use case is Green, Amber or Red?
3. Why is "is the tool enterprise grade" the wrong basis for classification?
4. Which kind of use case feels low risk but should be classified Red?
5. What is the minimum control that should attach to any Red classified use?

Which sources is this part built on?

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

• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.
• ICO, Employment practices and data protection: monitoring workers
  Caution: The 70 per cent figure is not in the guidance and is not a finding about workers. It is a general public survey about a hypothetical. Do not write that seven in ten workers regard monitoring as intrusive.
• CIOB, Artificial Intelligence (AI) Playbook 2024
• UK Government, AI Playbook for the UK Government

The instinct is to treat consumer tools as dangerous and enterprise tools as safe. That sorts by data security and ignores correctness entirely, which is the half that produces the expensive failures.

What were the answers to the previous five questions?

1. Two people use the same AI product. One summarises internal minutes, the other drafts contractual notices. Should they carry the same controls, and why?

No. The product is identical and the exposure is not. A poor internal summary costs a clarifying email. A defective notice can lose an entitlement or start time running. Controls follow consequence, so the two sit in different bands despite sharing a licence.

2. What two variables should determine whether a use case is Green, Amber or Red?

The consequence if the output is wrong, and how late that error would surface. High consequence caught immediately in normal review is manageable. Moderate consequence discovered on site or in an adjudication is not.

3. Why is "is the tool enterprise grade" the wrong basis for classification?

Because it answers where the data sits, not whether the answer is right. An enterprise secure system will still work from a superseded revision, miss a project specific amendment or state responsibility it cannot evidence. Security is a precondition, not a classification.

4. Which kind of use case feels low risk but should be classified Red?

Summarising a design team meeting where a fire strategy assumption changed. It looks like note taking. The summary becomes the record of a decision affecting statutory compliance, and the error would not surface until building control or later. Red, on both variables at once.

5. What is the minimum control that should attach to any Red classified use?

Four are non-negotiable: a named decision owner, explicit approval by an authorised role working from the underlying evidence rather than the AI summary, a defined escalation route, and a retained record of what was reviewed and decided. The band table adds three more that apply once the use case is running: defined delegated authority, formal testing before deployment, and periodic assurance afterwards.

Why does the two variable model work better than a list of tools?

Because a list of tools goes out of date at the next software update, and because it sorts on the wrong attribute. Consequence and lateness are stable properties of the work itself.

Lateness deserves more weight than it usually gets. Construction has spent decades building gates precisely because the cost of an error rises steeply with the distance between the mistake and its discovery. An AI error caught by the person who prompted it costs nothing. The same error discovered when the cladding is up costs whatever it costs.

What sits in each band?

Green is assistance where being wrong costs a correction. Amber is anything that materially informs professional work. Red is anything where being wrong reaches safety, statute, certification or a material sum, and the control steps up at each boundary.

Note what is absent from all three: any mention of a product. The same assistant can legitimately appear as Green for drafting an agenda, Amber for interrogating a specification and Red for anything touching a gateway submission under the Building Safety Act. One register, three rows, three sets of controls.

Where do firms get the classification wrong?

In four recurring places, all of them in the same direction.

• Meeting minutes. Filed as Green because it is note taking. Minutes are frequently the only record that a decision was made, by whom, and on what basis. If a summary can become the evidence of a commercial decision, it is Amber. If it can become the evidence of a decision bearing on statutory compliance or safety, a design team meeting where a fire strategy assumption moved being the standard case, it is Red. One activity, one tool, three bands, and the two variables decide which.
• Email refinement. Filed as Green because it is writing help. Outbound project correspondence is a project record with contractual weight, as part two set out. Internal chat is Green, correspondence to the other side is Amber.
• Anything involving a graduate. Firms classify by the seniority of the user rather than by the consequence of the output. An assistant quantity surveyor producing a commercial assessment is not exercising authority at all, because the authority sits with whoever signs. What has changed is that the signatory is now approving work that reads as though it were senior, so the review requirement rises and the reviewer has to work from the evidence rather than from the document in front of them.
• Embedded features. AI that arrived inside estimating or design software by update often gets no classification at all because nobody procured it. It is doing exactly the work you would classify Amber if you had noticed.

How do you make the classification survive contact with a busy week?

By making the control proportionate enough that nobody has an incentive to route around it. A three day approval queue for a meeting summary teaches people to stop using the approved route, and you are back to a register you cannot believe.

Write the gate into the register row itself rather than into a separate policy document. A control that lives next to the use case it governs gets followed. A control that lives in a policy on SharePoint gets acknowledged once at induction. Related: controlled AI adoption not blanket bans and private AI versus public chatbots.

Classification tells you how much human control a use case needs. It does not tell you what that control consists of, and "a human checks it" is not an answer. part 6, what meaningful human involvement now has to mean makes it one.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. Since 5 February 2026, what does UK law say makes a decision "solely automated"?
2. What four attributes does a human reviewer need before their involvement counts as meaningful?
3. What is automation bias, and what organisational conditions make it worse?
4. "A competent person reviews it" is not a control. What would a written control have to specify before a reviewer at five o clock on a Friday knows exactly what to look at?
5. Why should a firm log overrides as carefully as it logs approvals?

Which sources is this part built on?

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

• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.
• Building Safety Act 2022
  Caution: The Act is the enabling statute. The operative procedure, including change control, is in the Building (Higher-Risk Buildings Procedures) (England) Regulations 2023. The regime is England only: Wales operates separately and Scotland has building warrants and no gateways.
• ICO, consultation on draft guidance about automated decision-making, including profiling
  Caution: Published 31 March 2026. The consultation closed on 29 May 2026 and the guidance remains in draft pending the final version, so cite it as draft guidance whose consultation has closed, never as settled law. Re-check the status before quoting it. Note also that section 80 does not define meaningful human involvement: Article 22D reserves that to regulations the Secretary of State has not yet made.
• Data (Use and Access) Act 2025, section 80

An approval button records assent. It does not evidence involvement. Where a decision is about a person, section 80 of the Data (Use and Access) Act has made that difference a legal one. Everywhere else it is now the standard you will be measured against anyway.

What were the answers to the previous five questions?

1. Since 5 February 2026, what does UK law say makes a decision "solely automated"?

Section 80 of the Data (Use and Access) Act 2025 provides that a decision is based solely on automated processing where there is no meaningful human involvement in the taking of it. It replaced Article 22 of the UK GDPR and moved the framework from prohibition towards a risk based model with mandatory safeguards.

2. What four attributes does a human reviewer need before their involvement counts as meaningful?

ICO draft guidance names three: competence to understand the subject and the system, authority to alter or reject the output, and discretion that is genuinely exercised rather than nominal. Construction work makes a fourth essential: access to the reasoning and the sources. The test is whether the person can exercise real influence before the decision is applied.

3. What is automation bias, and what organisational conditions make it worse?

The tendency to give an output excessive credibility because a system produced it. It worsens when leadership visibly trusts the tool, when overrides require justification but acceptance does not, when review time is squeezed, and when the reviewer cannot see how the conclusion was reached.

4. "A competent person reviews it" is not a control. What would a written control have to specify before a reviewer at five o clock on a Friday knows exactly what to look at?

Seven things, and their point is specificity rather than completeness. What the AI does with no human step, what it may only recommend, the exact evidence a human must check, the judgement a human must make, the role authorised to make it, the triggers that force escalation, and what is kept. "Check it is right" satisfies none of those, which is why it is not a control.

5. Why should a firm log overrides as carefully as it logs approvals?

Because an override is the only signal that tells you where the system is wrong. A high override rate on one flag type means the rule is wrong, the knowledge is stale, or the process differs from the written policy. All three are worth knowing and none of them appear in an approval log.

What changed in law, and does it reach construction?

It reaches construction, though not by the route the statute itself takes. Section 80 bites on significant decisions about individuals involving personal data, so a great deal of construction work sits outside its direct scope. Reading it narrowly would still be a mistake.

The reason is that Parliament has now put a statutory test around the thing every AI governance framework leans on. A decision is solely automated where there is no meaningful human involvement in taking it. Note what was not done: the Act does not define meaningful human involvement, and Article 22D reserves that meaning to regulations the Secretary of State has power to make and has not yet made. What exists in the meantime is the draft ICO guidance, which is the clearest available statement of what a UK regulator considers oversight to be. It has no legal force over a commercial or technical decision, and it is still a sound working test to borrow for one: a firm that cannot meet it on a Red use case should assume the use case would not survive scrutiny.

The ICO position is unusually direct: involvement must be active and not a token gesture, and the test is whether a person can exercise real influence over the decision before it is applied, with the authority, discretion and competence to alter it. Reviewers should be trained to understand the system's logic, outputs, limitations and risks (ICO draft guidance, 2026).

What does the matrix actually contain?

Seven fields, completed for every Amber and Red entry in the register. This is the mechanism; everything else supports it.

The value is in the third and fourth rows. "Human must verify: the output is accurate" is not a control, because it does not tell a tired reviewer at five o'clock what to look at. "Confirm the cited clause matches the signed and amended contract" does.

How does the gate fit into the flow of work?

It sits between the flag and the action, and it is chosen by the consequence of the decision rather than by whoever happens to be at the screen.

Two design decisions in that diagram do most of the work. The first is that the authority engine is deterministic. Whether someone is permitted to approve a given decision is a rules question with a correct answer, and answering it probabilistically is an unforced error.

The second is that the checker is not the drafter. If the same model both produces the analysis and confirms it is sound, the second step adds confidence without adding assurance. AI cannot mark its own homework, and where the consequence is high, the challenge function should sit on a different model family and a different prompt lineage.

Why do overrides matter more than approvals?

Because approvals tell you the system was used and overrides tell you where it is wrong.

ICO guidance on AI and data protection is explicit that human reviewers must have the authority to override the output and be confident they will not be penalised for doing so, and that policy and training alone cannot create that confidence: a supportive culture is also required. Its draft guidance on automated decisions adds that you should keep a record of how each review was carried out, and monitor the outcomes. In a construction business the culture half is the hard part. If every override triggers a conversation about why the expensive platform was ignored, the path of least resistance becomes acceptance, and automation bias becomes company policy.

Handled well, the override record becomes the most useful dataset the firm owns. A planner who overrides a criticality flag because finishing on Friday determines Monday mobilisation, which the programme does not model, has told you something true about either the programme or the rule. Either way the system improves, and the person who spotted it should be the one who gets the credit.

Practical rule: make the reason field mandatory, keep it free text, and read it monthly. See operational AI versus chatbots and instruction to be agreed for why recorded reasoning matters commercially as well as legally.

You now have classification and control. What you do not yet have is a clear picture of what you are controlling against, because most firms are watching for the wrong failure. part 7, construction AI does not usually fail by inventing things sets out the eight that actually happen.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. Why is "check for hallucination" the wrong instruction to give a construction team?
2. In what three ways can an AI output be entirely factual and still wrong on a live project?
3. Who in a construction business is best placed to catch a context failure, and why is it not the IT department?
4. How would you test an AI system for the superseded document failure mode before deploying it?
5. What should "human must verify" say for a use case involving contractual terminology?

Which sources is this part built on?

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

• Data (Use and Access) Act 2025, section 80
• ICO, consultation on draft guidance about automated decision-making, including profiling
  Caution: Published 31 March 2026. The consultation closed on 29 May 2026 and the guidance remains in draft pending the final version, so cite it as draft guidance whose consultation has closed, never as settled law. Re-check the status before quoting it. Note also that section 80 does not define meaningful human involvement: Article 22D reserves that to regulations the Secretary of State has not yet made.
• ICO, Guidance on AI and data protection: ensuring individual rights in AI systems
• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.
• UK Government, AI Playbook for the UK Government

A model that invents a clause is easy to catch, because the clause does not exist. A model that quotes the right clause from the wrong revision produces something that survives every check a firm has, and is wrong.

What were the answers to the previous five questions?

1. Why is "check for hallucination" the wrong instruction to give a construction team?

Because it directs attention at fabrication, which is rare and easy to spot, and away from context failure, which is common and hard to spot. A team told to watch for invented facts will pass a confident, correctly cited answer drawn from a superseded revision.

2. In what three ways can an AI output be entirely factual and still wrong on a live project?

It quotes the standard form where your contract has been amended. It answers from the drawing revision that was current until last Tuesday. It states a settlement position that is reasonable on these facts and inconsistent with the position the firm is running on three other live disputes.

3. Who in a construction business is best placed to catch a context failure, and why is it not the IT department?

The QS who knows the contract was amended, the design lead who knows which revision is current, the planner who knows the logic was changed without agreement. Catching these requires knowing how construction information behaves, which is a professional competence and not a technical one.

4. How would you test an AI system for the superseded document failure mode before deploying it?

Deliberately give it a superseded drawing alongside the current one and ask a question the two answer differently. A system that answers confidently without flagging the conflict has failed. Build the same trick into the golden test set for every category.

5. What should "human must verify" say for a use case involving contractual terminology?

Confirm the term used matches the defined term in this specific contract rather than general usage, and that the mechanism named is the one the contract actually provides. Not "confirm the output is accurate".

What are the eight categories?

Wrong contractual terminology, a missed project amendment, a superseded document, a conflict with company policy, the wrong specification revision, an unrecognised statutory requirement, a misread approval status, and an unsupported statement of responsibility.

Worked briefly, in the form they take on site:

What do all eight have in common?

The model reasoned correctly from information that was incomplete, and produced fluent output that concealed exactly how incomplete it was. There is no moment of obvious error to catch.

This has a direct consequence for who does the checking. Fabrication can be caught by anyone careful. A superseded revision can only be caught by someone who knows which revision is current, and a missed amendment only by someone who knows the contract was amended. Human in the loop only works if the human retains and applies precisely the expertise these failures require, which is the argument from part three arriving from a different direction.

It also explains why hallucination dominates the conversation. It has a clean story: the system said X, X is false, therefore check facts. The eight above each require understanding something about how construction information behaves, and that is harder to put on a slide. It is also why they are the ones that actually happen.

How do you build the check into the way work is done?

Three points in the cycle, none of which is a reminder to be careful.

1. Before deployment. Build a golden test set from a completed project: a hundred to two hundred real events with known correct outcomes, including deliberately difficult cases in each of the eight categories. Measure precision and recall before anyone relies on the system, and record which matters more here: a notice trigger agent that is highly precise and catches a third of the triggers is worse than useless, because it has taught you to trust it. This needs no technology to produce and it is the most valuable asset in the whole programme.
2. During operation. Sample outputs against the eight categories rather than checking generically whether they look right. Ten outputs a month, checked against a specific list, surfaces real weakness faster than ad hoc spot checking ever will.
3. After an incident. Classify it against the eight. Over a year that tells you which category your particular deployment is weak in, which is the only basis on which to strengthen anything.

The single behavioural change that catches the most is to move the team from asking for answers to asking for evidence. Not "is this a variation" but "identify the contractual provisions, correspondence and records relevant to whether this is a change, and show the sources". Not "are we delayed" but "compare the programmes, identify movement to critical activities and show which records support the apparent causes".

Related reading while the module settles: hallucination risk in professional documents and which AI model for construction teams.

That completes the control module. The next four parts put it to work, starting with the discipline that touches every other one. part 8, the project manager who compressed the work, not the thinking follows a project manager through a single Wednesday morning.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. What is the difference between using AI to make a decision and using it to compress the work that precedes a decision?
2. A revised programme lands. What should you ask for, and what should you not ask for?
3. Why does a specific, constrained request produce better evidence than an open question?
4. What is the risk of a structured, authoritative looking output, and what habit counters it?
5. Which decisions in a project manager's week should be explicitly marked as human only?

Which sources is this part built on?

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

• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.
• BSI, BS EN ISO 19650 information management
• Building Safety Act 2022
  Caution: The Act is the enabling statute. The operative procedure, including change control, is in the Building (Higher-Risk Buildings Procedures) (England) Regulations 2023. The regime is England only: Wales operates separately and Scotland has building warrants and no gateways.
• The Building (Higher-Risk Buildings Procedures) (England) Regulations 2023
  Caution: England only. There is no such thing as a "gateway re-submission": the mechanism is change control.
• Building Safety Regulator
  Caution: Application volumes and median decision periods are published periodically and move month to month. Do not quote a specific in-flight figure in evergreen copy.

The wrong way to use AI is to outsource the judgement. The right way is to compress the work that prevents judgement from happening, which on most projects is most of the week.

What were the answers to the previous five questions?

1. What is the difference between using AI to make a decision and using it to compress the work that precedes a decision?

Making the decision means the output is acted on. Compressing the work means the professional receives structured, sourced evidence and then decides. The first changes who is accountable. The second changes only how long it took to be ready to decide.

2. A revised programme lands. What should you ask for, and what should you not ask for?

Ask for a structured comparison of the two programmes, naming both and stating the data date and contractual status of each. Activities added and removed; changes to original duration, remaining duration and percent complete; logic changes including relationship type and every lag; calendar and constraint changes; and the scheduling settings, specifically retained logic against progress override, because that one setting can move planned completion by weeks without a single duration changing. Then movement in total float and free float, and the difference between the longest path and the total float critical set, which in P6 are not the same thing. Do not ask whether the project is delayed or what to do about it. The first is retrieval, the second is judgement.

3. Why does a specific, constrained request produce better evidence than an open question?

Because scope is what makes an answer checkable. "Is this a variation" invites a plausible narrative. "Does this scenario meet the definition in clause 4.2.1, and show the clause" produces something with a reference you can open.

4. What is the risk of a structured, authoritative looking output, and what habit counters it?

Automation bias: the more organised the output looks, the less it gets questioned. The counter is routine spot checking. If it says a clause requires something, open the clause. Not every time, but often enough that the habit does not decay.

5. Which decisions in a project manager's week should be explicitly marked as human only?

Giving an instruction. Certifying or notifying a payment. Accepting a programme, or notifying reasons for not accepting it. Notifying, or deciding not to notify, a change event. Certifying completion. Write the list in the vocabulary of the contract you are actually on, because a list written in the wrong regime's words is the first sign nobody has read the conditions, and write it down before somebody needs it.

What is actually in the morning?

A revised programme from the contractor. An email thread from the designer about an ambiguous specification. Yesterday's site meeting minutes, taken by someone without the project context. Three messages from the site manager about progress. A query from the QS about a possible variation. An instruction draft that may or may not be justified. A delay notification. A payment application with supporting documents. A compliance query from the client.

The output required from all of that is a short list of decisions: what the programme movement means for the critical path, whether the instruction is compliant and within authority, what the delay notification does to the forecast, what to approve. None of that can be automated and none of it should be.

What can move is everything that has to happen before any of it can be thought about. That is the distinction the whole module turns on.

What does a good request look like?

The same five part structure every time, which is worth teaching once rather than distributing a library of prompts nobody will type.

What comes back is a list with references rather than a narrative. The follow up is then the question that actually matters, and it is a question only the PM can frame: does this change affect the payment application date or the handover milestone, and show me the logic.

The same structure applies to the specification thread. Retrieve the specification and the correspondence, ask what each email states or modifies with dates, ask what the most recent definitive position is, and ask what contradictions exist. The output is not the answer. The output is a precise question to put to the designer: your email on Tuesday said mild steel would be acceptable if stainless was delayed, and the specification requires stainless. Before precedence, confirm why stainless was specified, whether that is design life, exposure, or bimetallic action with adjacent components, and whether carbon steel meets it here. If it does not, this is a rejection rather than a clarification. If it does, an email does not change the scope: that needs an instruction, and an instruction is a change event.

Where does this go wrong?

In four places, and they are worth naming to the team before they happen rather than after.

• Confinement is not immunity. Retrieval from a controlled source lowers fabrication risk substantially. It does not eliminate it. The clause the system identified still has to be opened occasionally.
• Missing context. The system will not know that a specification clause was overtaken by an email the site team treats as controlling. Only someone who knows the project spots that gap.
• Automation bias. Structured output invites acceptance. This is the failure mode most likely to affect a good project manager, precisely because the output is usually right.
• Scope creep. A tool that compares programmes well invites "just approve this payment". Different question, different verification, different authority. The human only list exists for this moment.
• Documentary answers to technical questions. Which document takes precedence is the easy half and the half a retrieval system is good at. Whether a substitution is acceptable at all is an engineering judgement about design life, exposure and compatibility, and no hierarchy of documents answers it. A system that reconciles a conflict has often concealed the question.

What does the junior get out of it?

Potentially a great deal, and potentially nothing, depending entirely on how the firm sets it up.

The failure case is the one from part three: a graduate who receives answers so quickly they never work through the reasoning, and who five years later cannot challenge the system they depend on. The better case uses the same tool as a challenger rather than an oracle. Ask it not to answer yet, but to list the questions worth considering first. Ask it to identify weaknesses in your interpretation. Ask for three plausible readings and what information would distinguish between them.

That is closer to having a patient senior colleague available at nine on a Tuesday than to having the work done for you, and it is the one use that genuinely accelerates professional development rather than deferring it. More on the training design in training sceptical construction teams and on the tooling in three AI solutions every contractor needs.

Project management is mostly sequential: something happens, you respond. part 9, design management is a dependency problem, not a document problem takes on the discipline where everything happens at once and the expensive failures are the connections nobody spotted.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. Why is design management a harder problem for AI than project management?
2. What is the difference between a clash and a dependency, and which is more dangerous?
3. A client moves a wall by one metre. What are the second order effects, and why do they matter more than the first?
4. What information does a system need before it can say anything useful about buildability?
5. What can AI not tell you about Building Regulations compliance, however good the retrieval is?

Which sources is this part built on?

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

• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.
• CIOB, Artificial Intelligence (AI) Playbook 2024
• ICO, consultation on draft guidance about automated decision-making, including profiling
  Caution: Published 31 March 2026. The consultation closed on 29 May 2026 and the guidance remains in draft pending the final version, so cite it as draft guidance whose consultation has closed, never as settled law. Re-check the status before quoting it. Note also that section 80 does not define meaningful human involvement: Article 22D reserves that to regulations the Secretary of State has not yet made.

Any single discipline getting something wrong is recoverable. The costly case is structural assuming one thing, MEP assuming another, and nobody finding out until the contractor does.

What were the answers to the previous five questions?

1. Why is design management a harder problem for AI than project management?

Because project management is largely sequential and design management is parallel. Dozens of disciplines generate interdependent information at once, and the failures live in the relationships between documents rather than in any one of them.

2. What is the difference between a clash and a dependency, and which is more dangerous?

A clash is two elements occupying the same space, which is geometric and detectable. A dependency is one discipline relying on an assumption another has quietly invalidated. Dependencies are more dangerous because nothing intersects, so nothing is flagged.

3. A client moves a wall by one metre. What are the second order effects, and why do they matter more than the first?

First order: structural resize, ceiling zones, duct rerouting. Second order: fire strategy assumptions, construction sequence, procurement lead ins, and possibly a gateway re-submission. The second order effects consume the programme, and they are the ones nobody costs at the point of agreeing the change.

4. What information does a system need before it can say anything useful about buildability?

The contractor's actual planned construction sequence, real site constraints including access and existing structures, temporary works requirements, and logistics such as crane capacity and delivery routes. Without those, buildability analysis is guesswork dressed as analysis.

5. What can AI not tell you about Building Regulations compliance, however good the retrieval is?

How the body reviewing it will interpret your approach. Regulations are applied through judgement and interpretations differ, and for a higher-risk building there is no choice of body and no informal conversation at all. What AI can do is align the disciplines' assumptions and expose where they diverge, so the question you put is precise rather than open.

What does the design manager actually have to hold in their head?

On a medium project: structural, architectural, mechanical, electrical and public health, civils, specialist packages such as lifts and fire safety, client requirements, regulatory compliance and contractor buildability input. Forty to fifty documents produced or revised in a month, each carrying a status code and a revision under BS EN ISO 19650, still widely called suitability on site, all interdependent. Note that status and revision answer two different questions: the status code says what the information may be used for, and only the revision says whether it is current.

The traditional response is coordination meetings, design check meetings, an RFI system, specification reviews, mark ups, and the stage sign offs the RIBA Plan of Work 2020 requires at each information exchange. All necessary. All producing more information, which is the recursive part of the problem. The distinction that matters is between Stage 3, where geometry gets coordinated, and Stage 4, where design intent meets contractor's design and specialist design and separate parties' assumptions first have to be true simultaneously. A project that signs off Stage 3 because nothing in the federated model intersected, and calls that coordinated, has the failure below waiting for it.

What fails is rarely a discipline being incompetent. Structural allows half a kilonewton per square metre to the soffit for services and finishes. MEP hangs a chilled water run and a fan coil unit. Architecture adds a ceiling with an acoustic layer. Fire adds protection to the duct. Every discipline sits inside its own allowance, the aggregate does not, and nobody owns the aggregate. Nothing clashes, because nothing intersects.

That is the general shape of it. Clash detection tests nominal geometry. The dependencies that cost money live in tolerance, deflection, aggregate load, construction sequence and third party test evidence, none of which a model represents, and all of which are assumptions one discipline holds about another.

Where does AI actually help?

In three places, ordered by how much of the value they carry.

Cascade is the one that changes the conversation with a client, because it converts a request into a costed consequence before anyone agrees to it.

The output is not a duration. It is a structured consequence the design manager can put to a client with the uncertainty left visible: moving that wall reaches structural, architectural and MEP, and the design work has to happen in that order because MEP cannot resolve until the other two are fixed. It revises the fire strategy, which changes the core sequence and brings the shaft procurement decision forward. If this is a higher-risk building past Gateway 2, a change to compartmentation is likely to be a major change under the change control regulations, which means an application to the Building Safety Regulator and a decision before the work can start. I can give you the design effort. I cannot give you the regulatory date, and nobody who offers you one should be believed.

That is a materially better conversation than structural says it is fine but MEP might have issues, and it is better precisely because it does not end in a number. The value of the cascade is that it makes the shape of the consequence visible before anyone commits, not that it produces a duration the system is in no position to promise.

What is different about the risk here?

The consequences of a missed connection are higher, and the failures are quieter.

• Implicit dependencies. Two details that never clash but where one assumption invalidates another. The system sees what is drawn, not what is implied.
• Missing site context. An existing tree not in the model, an unrecorded utility, a neighbouring structure. The design works and cannot be built.
• Interpretation divergence, and which regime you are in. For conventional work a local authority or a registered building control approver may reach different views on the same clause, and no retrieval predicts which. For a higher-risk building that risk changes shape rather than disappearing: building control is the Building Safety Regulator, there is no alternative body and no informal steer, and every question is a formal submission answered on the regulator timescale.
• Specialist package constraints. Lift loadings, plant clearances and equipment access are often contractually or proprietarily defined. A geometric check will not see a contractual constraint.
• Statutory triggers. On a higher-risk building in England, a material change of use to any part of the building is expressly a major change under the change control regulations. The client must apply to the Building Safety Regulator for approval of that change, and the work must not start until it is granted. A review scoped to technical compliance will not raise it, and it will not know that it did not.

What changes about the coordination meeting?

It stops being a search and becomes a resolution.

The traditional meeting opens with a question nobody can answer well: are there any clashes? The honest answer is usually yes, but I am not sure which ones matter. The assisted version opens with six critical clashes, four tight conditions and three sequence risks, each routed to the discipline that has to move, each with the clearance required and the clearance provided. The meeting spends its time deciding rather than discovering.

The design manager coordinates. They do not hold the design. Design intent sits with the appointed designers under their appointments, contractor's designed portions sit with the contractor and flow down to the specialists, and specialist packages sit with the specialists. A system that says MEP should move has expressed a preference, not a determination: which party moves is answered by the design responsibility matrix and the appointments, not by the model.

Surfacing an issue by machine transfers no design responsibility either. What it changes is the record. Before, an unnoticed dependency failure turned on whether reasonable skill and care would have found it. After, there is a dated entry showing it was raised, and whether anyone acted. A flagged item left open is a materially worse position than one nobody ever saw, and that register is disclosable. Every flag needs a disposition with a name and a reason against it. See also what a good AI pilot looks like and MCP and the connected construction office.

Design decides what gets built. part 10, the most valuable AI on a project may save nobody any time, published Wednesday, 26 August 2026 deals with the discipline that has to prove what happened, months or years after it did.

What five questions should you be able to answer now?

Attempt these before tomorrow. Each has a defensible answer, and each is answered at the top of the next part.

1. Why is commercial management described as an evidence game rather than a numbers game?
2. What is a challenge function, and why might it save no time at all?
3. A CVR says a package is 74 per cent complete. What would you compare that against?
4. Why does a contemporaneous commercial record beat a reconstructed one, and what does that mean for final accounts?
5. What should happen when a commercial manager overrides an assessment, and what makes the difference between a strong and a weak position later?

Which sources is this part built on?

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

• BSI, BS EN ISO 19650 information management
• RIBA Plan of Work 2020
• Building Safety Act 2022
  Caution: The Act is the enabling statute. The operative procedure, including change control, is in the Building (Higher-Risk Buildings Procedures) (England) Regulations 2023. The regime is England only: Wales operates separately and Scotland has building warrants and no gateways.
• The Building (Higher-Risk Buildings Procedures) (England) Regulations 2023
  Caution: England only. There is no such thing as a "gateway re-submission": the mechanism is change control.
• Building Safety Regulator
  Caution: Application volumes and median decision periods are published periodically and move month to month. Do not quote a specific in-flight figure in evergreen copy.
• The Building Regulations 2010, Part 2A: dutyholders and competence
  Caution: England only. CDM 2015 applies across Great Britain; Part 2A does not. Never present the two as co-extensive. Competence is the Part 2A term; CDM 2015 deliberately uses skills, knowledge, experience and organisational capability instead.
• RICS, Responsible use of artificial intelligence in surveying practice, 1st edition
  Caution: The quarterly review obligation attaches to the risk register, not to the register of AI use. The client disclosure duty is mandatory and unconditional wherever the material impact test is met: it is not a graduated judgement call. Most of the firm level duties bind RICS-regulated firms specifically rather than all members.