Chris M. / 9 June 2026 / Reviewed 12 August 2026

Using ChatGPT, Gemini and Claude in a UK Construction Business: A Practical Guide

A site manager pastes a photo of a snagging list into ChatGPT to get a tidier version to send the client, on the same free account they use at home for recipe ideas. Nothing about that action feels different from using a calculator. It is not the same action. The free consumer tier of ChatGPT is the only one of nine account types this guide checked, across all three major chat tools, that has no data processing agreement available at any price, and by default it trains the underlying model on what gets pasted into it.¹ The paid business version of the identical-looking product sits under a contractual promise that the same input will never be used that way.² Nobody chose that gap on purpose. It exists because the free tab and the enterprise account look the same, and almost nobody reads past the login screen.

This guide covers what actually changes by tier and vendor, what UK data protection law and the surveying profession's own new standard require, where the evidence says these tools are genuinely unreliable, and what a construction business can do that is reliably useful.

The data question, tier by tier

This is the part worth getting exactly right, because getting it wrong is the single most damaging mistake available here.

ChatGPT. The free, Plus, Pro and Go consumer tiers train on conversations by default, and OpenAI's own wording is direct about it: "ChatGPT, for instance, improves by further training on the conversations people have with it, unless you opt out."³ The opt-out is a toggle in Settings, off by default, and turning it off stops future conversations being used but does not undo past training or delete existing history. Team, Enterprise and API accounts sit on the opposite default: "we do not train on any inputs or outputs from our products for business users," and only these business tiers can sign a data processing agreement at all.⁴

Gemini. The consumer app at gemini.google.com trains on conversations by default too, through a setting called Gemini Apps Activity, and Google states plainly that with it on, "a subset of chats are reviewed by human reviewers... to help improve Google services."⁵ Turning Activity off stops future training use but Google still holds the chat for 72 hours to generate the response, and states it may still use conversations "to respond to you and help protect Google... including with help from human reviewers" even with Activity off. The unpaid developer API carries an explicit warning most people never see: "Do not submit sensitive, confidential, or personal information to the Unpaid Services."⁶ The paid API and Google Workspace's business version of Gemini both commit not to train on customer content and both can be covered by a data processing agreement.⁷

Claude. Anthropic's current consumer wording frames training as an opt-in choice made in privacy settings, called Model Improvement, and declining keeps the standard 30-day retention rather than the five years that applies if a user opts in.⁸ Worth flagging directly: contemporaneous reporting on how this setting was rolled out in August 2025 describes existing users being shown a pop-up where the prominent action opted them in, with a smaller toggle needed to opt out, and the same pre-set toggle shown to new signups.⁹ Anthropic's own current page and that reporting do not read identically, so the safe instruction is to check the actual toggle in your own account rather than assume a default either way. Claude for Work, Team, Enterprise and the API sit on a hard contractual line instead: "Anthropic may not train models on Customer Content from Services," with a data processing agreement automatically incorporated into the commercial terms.¹⁰

The pattern across all three is the same. The free version anyone can start using in ninety seconds is the version with the weakest data terms, and the version with the strongest terms is the one requiring a business account and, in most cases, a signed agreement nobody sets up by accident.

What UK law actually requires, and what does not bind you yet

Site photographs, subcontractor names and addresses pasted into any of these tools are personal data, and processing them is governed by UK GDPR regardless of which chat tool is doing the processing. The Information Commissioner's Office maintains standing guidance on AI and data protection, most recently updated as it works through a formal review following changes to UK data law.¹¹ Sending that data to a US-based vendor is an international transfer, and the ICO has published a specific three-step test for identifying when a transfer needs additional safeguards.¹²

It is worth being precise about what is and is not currently binding. The government's approach to AI regulation, the pro-innovation white paper published by the Department for Science, Innovation and Technology, sets out non-statutory principles for existing regulators to apply.¹³ It creates no new AI-specific law and no new AI regulator. What actually binds a UK construction business today is the UK GDPR and the Data Protection Act 2018, exactly as they applied before any AI tool existed, not the white paper's principles.

RICS now has a mandatory standard on this, and it is a real change

If your practice includes RICS members, this is not optional guidance to skim. RICS's professional standard, Responsible use of artificial intelligence in surveying practice, took effect on 9 March 2026, and its own wording is unambiguous about its status: professional standards are the ones members "must comply with," as distinct from the guidance notes members are merely expected to follow.¹⁴

The standard requires, among other things, a written AI risk register reviewed at least quarterly, documented due diligence before adopting any AI system with a material impact on service delivery, and clients told in writing, in advance, when and for what purpose AI will be used on their instruction, including the extent of professional indemnity cover for that use where available. The most consequential single line is this one: any AI output with material impact must be reviewed by "an appropriately qualified and named surveyor who accepts responsibility for its use." An AI-drafted answer is not a finished answer under this standard. It is a draft with your name still required on it.

CIOB has published an AI playbook covering ethics, cyber security and a decision checklist, but as practice guidance rather than a mandatory standard.¹⁵ RIBA's most recent published material is an adoption survey, reporting 59% of architecture practices now using AI, alongside a statement that RIBA is developing its own guidance rather than having already published one.¹⁶ No equivalent document from the Institution of Civil Engineers was found. On insurance, no UK insurer has published a construction or surveying-specific position on how AI-assisted work affects professional indemnity cover; what exists is general market commentary noting that AI-specific exclusions are not yet standard but are considered increasingly likely as the risk develops.¹⁷ That gap, not a settled market position, is the honest answer if a client or insurer asks.

Where the evidence says these tools genuinely fail

This is not a caution to be polite about. It needs a real case, and there is one.

In June 2025 the High Court's King's Bench Division, sitting as a Divisional Court, handed down judgment in a joined referral covering two separate sets of proceedings in which lawyers had used generative AI to produce legal arguments and witness statements containing fabricated case citations, put before the court unchecked.¹⁸ This is not an American story imported for effect. It happened in England, in 2025, to qualified professionals whose job was precisely to check their own sources before relying on them. The lesson for construction is not "never use AI for anything written." It is that fabricated, entirely plausible-sounding citations and case references are a documented, real failure mode, and anything an AI tool produces that claims to cite a standard, a regulation or a precedent needs the citation checked against the actual source before it goes anywhere near a client, a tribunal or a report with your name on it.

Numerical reliability is a genuine, ongoing weakness rather than a solved problem. Recent academic benchmarking finds logical and arithmetic error rates rising by as much as fourteen percentage points as the numerical complexity of a task increases, with wide variation between models.¹⁹ Reading construction drawings is a similarly early and unsettled area: independent benchmarking of vision-capable models against construction drawing extraction tasks has documented models hallucinating or fabricating data outright, with larger models not reliably outperforming smaller ones.²⁰ Treat both of these as real, current limitations rather than solved problems a newer model version has quietly fixed.

The uncomfortable finding on productivity

The single strongest piece of evidence available on whether AI tools actually make people faster comes from a randomised controlled trial, not a vendor survey. METR ran a study of sixteen experienced open-source developers completing 246 real tasks in codebases they already knew well, split between working with AI tools and without.²¹ The developers using AI took nineteen percent longer to complete the same work. Before the study, they had expected AI to speed them up by roughly a quarter. Afterwards, having just been measured taking longer, they still believed AI had sped them up by about twenty percent. They could not correctly perceive that the tool had slowed them down.

This is a narrow, specific result, expert developers, familiar codebases, an early-2025 generation of coding tools, and it should not be generalised into a blanket claim that AI slows down every task in every context. What it should do is puncture any assumption that a feeling of increased speed while using these tools is reliable evidence of actually being faster. That feeling is worth checking against a clock, not trusting on its own.

Where it becomes genuinely useful

Set against those failure modes, there is a real, evidenced set of tasks these tools do reliably help with: summarising a long document, restructuring rough notes into a clean report, drafting a first pass at correspondence for a human to edit, comparing two versions of a document for differences, and turning a messy meeting recording into structured notes. The common thread across all of them is that a competent person still reads the output before it goes anywhere, and the tool is doing compression and reformatting of information that already exists, not generating new facts, new citations or new numbers that need to be independently correct.

UK construction is barely using any of this yet

Two separate official surveys, run differently and not directly comparable to each other, agree on the direction even where their exact figures differ. The Office for National Statistics found that 13% of UK construction businesses report using AI, against 58% in information and communication, describing construction as one of the lowest-adopting sectors it measured.²² The Department for Science, Innovation and Technology's own adoption research, run on a different methodology and timeframe, found 88% of construction businesses neither using nor planning to use any AI technology, against 80% across all sectors.²³ Whichever figure you use, construction is trailing most of the economy on this, which is worth knowing before assuming your competitors have already solved the problem this guide is about.

A short checklist before you paste anything in

Check which account you are actually logged into, not which product name is on the tab, since the free and paid versions of the same tool sit under genuinely different data terms. Never paste in anything containing a person's name, address or image unless you have checked which tier you are using and what UK GDPR requires for that transfer. Treat any citation, standard reference or case name the tool produces as unverified until you have checked it against the real source yourself. And if RICS governs your practice, read the actual standard rather than a summary of it, because "must comply with" is not the same commitment as "should consider."

Related reading

For the specific comparison between Microsoft's own assistant and a bespoke build, see Microsoft Copilot or a custom construction AI assistant. On the two failure modes this guide covers in most depth, hallucination risk in professional documents and which AI model for construction teams go further into named tools and named error types. And if chat tools are already spreading through your business without anyone deciding that on purpose, shadow AI in the workplace covers what that actually looks like.

Where to check this yourself

• "What has the ICO said about using AI with personal data?" The regulator's own AI and data protection guidance is free at ico.org.uk.
• "What does the RICS standard on AI actually require?" The full standard is free to read at rics.org.
• "What is the government's actual policy on AI regulation?" The white paper is free at gov.uk.
• "How many UK construction businesses actually use AI?" The Office for National Statistics publishes the current figures free at ons.gov.uk.
• "What does CIOB say about using AI responsibly?" CIOB's own AI playbook is free at ciob.org.

What could not be established for this guide

Whether OpenAI or Anthropic currently hold live certification under the UK extension to the EU-US Data Privacy Framework could not be confirmed from a primary source and should be checked directly against the official register rather than assumed either way. No named UK insurer has published a position specific to construction or surveying on how AI-assisted work affects professional indemnity cover, and this guide does not invent one. No equivalent standard or guidance from the Institution of Civil Engineers was found. And no settled, peer-reviewed figure exists yet for LLM arithmetic error rates or construction-drawing reading accuracy; the research in both areas is early, methodologically inconsistent between studies, and likely to date quickly, so treat any specific percentage in this space as indicative rather than final.