ChatGPT

ChatGPT is the AI tool most construction professionals try first, and for a lot of the writing that fills a QS or PM's week it is genuinely good. Drafting correspondence, restructuring rough notes, summarising a document you do not have time to read properly, producing a first pass at a report. It is not a construction tool. It is a general one that happens to be good at the kind of structured writing our industry runs on. That distinction matters, because it explains both what it is brilliant at and where it falls over. WHAT IT IS ACTUALLY GOOD AT ON A CONSTRUCTION JOB Turning mess into structure. Rough meeting notes into minutes. A voice memo from a site walk into a snagging list. Six weeks of emails into a chronology. This is the highest-value thing it does and it is the least glamorous. First drafts of correspondence. A letter you have written four hundred times and still resent starting. It gets you to a draft you can correct, and correcting is much faster than composing. Interrogating documents. Upload a specification and ask what it says about a specific obligation. It will find things. It will also miss things, so this is a starting point for your reading, not a replacement. Explaining something unfamiliar. A junior QS asking what a pay less notice is will get a decent answer. A junior QS asking whether their specific pay less notice is valid will get a confident answer that may be wrong. WORKED EXAMPLE: INTERROGATING A CLADDING SPECIFICATION A real shape of task. You are pricing a rainscreen package on a Newport office refurbishment. The spec is 62 pages, the tender is due Friday, and you need to know what testing and third-party certification obligations sit with you before you price them. What you do: upload the PDF and ask a specific question. "You are helping a UK subcontract estimator price a rainscreen cladding package. I have uploaded the specification, document ref NPT-SPEC-A2-Rev-C. List every obligation in this specification that falls on the cladding subcontractor for testing, third-party certification, or sample approval. For each one give me: the clause or section number, a one-line summary of the obligation, and who it says pays for it. Only list obligations that are actually in the document. If the specification is silent or ambiguous on who bears a cost, say so explicitly rather than assuming. Quote the wording where the obligation is unclear. Do not include general workmanship clauses." What comes back is a table of a dozen or so items with clause references. On a spec like this it will typically surface fire performance certification for the system, a mock-up panel for architect approval, weather-tightness testing on that mock-up, third-party installer accreditation, and a schedule of material samples. Say two of those you had already priced and one you had not. The specification puts the cost of the weather-tightness test on the mock-up with the subcontractor, and at roughly £4,000 that is a real number on a package this size. Finding it on Wednesday rather than after award is the entire value of the exercise. Now the part that matters, and this is what you should expect rather than hope to avoid. Some of those clause references will be wrong. Not wildly wrong: off by a section, or a summary that reads slightly more favourably to you than the actual wording does. Both kinds of error are invisible until someone opens the document at that page, and both are exactly the kind you would price against. So the workflow is: AI finds the candidates, you open the spec and verify every single one before any of it touches a price. It turns an afternoon of reading into forty minutes of targeted checking. It does not remove the checking. THE LIMITS, PLAINLY It cannot do arithmetic reliably. This catches people constantly because the output looks so convincing. Ask for a cost breakdown that sums to £840,000 and you will get a beautiful table that sums to £847,300. It predicts what a sum looks like rather than calculating it. Anything numeric gets checked in a spreadsheet, every time. It invents clause numbers and standards. Confidently, and in exactly the right format. A BS number that is close but wrong. An NEC clause that does not say what it claims. If a reference is going into a document with your name on it, open the actual source. It does not know your project. It does not know the ground conditions, the history with this client, or what was agreed in the meeting last Thursday. Everything it produces is generic until you make it specific. Its knowledge has a cutoff and it is not a live source. It is not a substitute for checking current Building Regulations, current standards, or anything that changed recently. On regulatory questions this matters a great deal. DATA AND CONFIDENTIALITY Before you paste anything sensitive into it, three checks. What your contract or appointment says about confidentiality. What your firm's policy is, because many now require the business tier only. And whether your tier trains on your inputs, which the consumer tiers often do by default and the business and enterprise tiers generally do not. Practical rule: use a business tier with training off for anything client-related, and redact names and figures on anything genuinely sensitive. If a matter is in dispute, assume everything you type could one day be disclosable. CHATGPT VERSUS THE OTHERS, HONESTLY Claude tends to be better on long documents and holds a careful tone well, which suits contractual writing. Microsoft Copilot's advantage is that it sits inside the Office tools your documents already live in, and on a lot of firms' IT policies that is the deciding factor rather than the model. Gemini is strong if your firm is on Google Workspace. For most people in construction the honest answer is that the differences between the current frontier models matter far less than whether you write a specific prompt and check the output. Pick the one your IT department will let you use, and spend the effort on the prompt. FREQUENTLY ASKED QUESTIONS Is the free version good enough for construction work? For occasional drafting, it is fine. Two things push you to a paid tier: file uploads for interrogating specifications and drawings, which is where a lot of the real value sits, and the data handling, since the consumer free tier's default training settings are not what you want for client documents. If you are using it for work, use a work tier. Can ChatGPT read construction drawings? It can read a drawing and describe what is on it, and it will miss things a competent design manager would catch, particularly relationships across sheets and revisions. Treat it as a second pair of eyes with no experience, not as a checker. For anything dimensional, do not rely on it at all. Will it get UK construction terminology right? Mostly, if you tell it to. Left alone it drifts American: dollars, OSHA, "change order" instead of variation, feet and inches. Say UK construction, say the contract form, say £, and it holds the register fine. The drift is worst on anything regulatory, where an American answer can be confidently and dangerously wrong. Is it safe to use for RAMS or health and safety documents? For a first draft, with a competent person reviewing and approving before it is used, yes. Never for the approval itself. The legal duty for a suitable and sufficient risk assessment sits with a competent person and cannot be delegated to a tool. The usual failure is that the AI leaves residual risk scores identical to the initial ones, which says in writing that your controls achieve nothing. How much does it cost in the UK? The free tier covers casual use. Paid individual plans are around £20 a month, with business tiers priced per seat above that and enterprise pricing on application. Check the current pricing page, since the tiers move regularly. For most construction professionals the paid individual tier is the point where file uploads make it genuinely useful, and the business tier is where the data handling becomes appropriate for client work.