Can AI Answer Building Code Questions? A Builder's Safe-Use Routine

A builder can use ChatGPT or Claude to find which code section probably applies, but should never rely on the answer. Chatbots misquote clauses, mix editions and miss local amendments. Use the chatbot as a fast index, read the clause yourself, then confirm the adopted edition with the building department or building control.

The danger is not that the answers are always wrong. They are usually plausible, written in confident language, and right often enough to lull you. The failures cluster in exactly the places that cost money: a stair rise limit from a different edition, a fire-separation exception that was removed, a local amendment the chatbot has never seen. A wrong answer that looks right is worse than no answer.

Follow me on Instagram@sagnikteaches

What a chatbot gets right and where it slips

Think of a chatbot as a well-read apprentice who has skimmed many versions of the code and cannot remember which is which. That apprentice is genuinely useful for orientation and genuinely dangerous as a final word. The split is fairly predictable.

Connect on LinkedInSagnik Bhattacharya

Notice what the useful rows share: in each, you stay the person who reads and decides. The rows where it fails share something too, because they all depend on a fact outside the chatbot's reach, either which edition governs your job or what the authority has said about it.

Subscribe on YouTube@sagnikteaches

The five-step routine

This is the routine I would put on a laminated card in the site office. It takes ten to fifteen minutes for a typical question and it keeps the chatbot in its lane.

1. Write the situation, not the question. Give the building type, the element, the dimensions and what makes it unusual. "Can I do this?" invites a confident guess. A described situation lets the chatbot list which topics apply.
2. Ask for sections to read, not for the answer. Tell it to name the likely chapters and clause titles, and to say plainly where it is unsure.
3. Open the real document. Find the clause in the edition the authority has adopted. Read the whole clause, the exceptions, the footnotes and the tables.
4. Confirm the edition and any amendments. Ask the building department or building control which edition and amendments apply to your application date.
5. Record it. Save the clause reference, the edition, who confirmed it and when, in the job file next to the drawings.

Steps three and four are the ones people skip, and they are the whole point. A chatbot answer without them is a rumour with good grammar.

A loft conversion stair, checked step by step

Take a small residential builder converting a loft in a two-storey house. The client wants a steeper, more compact staircase to save floor area on the first floor. The builder has a hunch the proposed rise and going are too tight but wants a quick view before phoning anyone.

Here is the prompt, written as a situation.

I am a builder working on a loft conversion in a two-storey house.
The new stair will serve one habitable room (a bedroom with an
en-suite) and rises about 2.6 m floor to floor. I want to use a
rise of about 220 mm and a going of about 220 mm, with a winder
at the turn.

Do not tell me whether this complies. Instead:
1. List the building code topics that would normally govern a stair
   like this (pitch, rise and going, headroom, handrails, winders,
   guarding, fire escape from a loft).
2. For each topic, name the section title you would look for.
3. Mark anything where requirements differ between code editions or
   between loft stairs and ordinary stairs.
4. List three questions I should put to the building department.
5. Say clearly which parts you are unsure about.

A reasonable illustrative response lists six topics: rise and going limits, pitch, headroom over the stair, handrail height and both-side requirements, winder going measured at a set distance from the narrow edge, and the extra fire protection that can apply when a loft adds a storey above the existing escape route. It then suggests questions such as "which edition applies to this application", "does your office permit a reduced going on a loft stair, and under what conditions", and "does the new storey change the escape route requirements for the existing stairs".

What would you fix in that output? Two things. First, the chatbot says a pitch limit "is usually around 42 degrees" for private stairs. That figure may or may not match the adopted edition, so you do not repeat it to the client or write it on a drawing. You note "pitch: check clause" and move on. Second, it mentioned loft stairs as possibly having relaxed rules, but never said under which conditions. That gap is where the inspector's answer matters most. So the builder opens the adopted edition, finds the stair chapter, reads the loft stair exception in full, and notices the exception only applies when the stair serves a single habitable room. Because the bedroom has an en-suite, the builder pauses and puts that exact question to the department.

Time spent: five minutes with the chatbot, ten reading the document, five on an email. What the chatbot saved was the ten minutes of working out which chapters to open and the half hour of drafting a clear question. What it did not do was decide anything.

Uploads, boundaries, basements, joists and inspections

Uploading your own copy of the code

Many builders have a licensed PDF of the code. Uploading it improves the quality of the answers because the chatbot can quote your copy. Claude can read PDFs up to 1,000 pages, with visual layout analysed only up to 100 pages, so a full code book is at the edge of what it will take in. That limit matters, because tables and diagrams are exactly the pages where the visual reading is needed. Ask for the page number and clause label of every quote, then open that page. If the quoted text is not on that page, discard the answer.

Using only the attached document, find the clause on handrail height
for stairs in a dwelling. Quote it word for word, give the page
number and clause label, and list any exception or table the clause
refers to. If you cannot find it, say so rather than answering from
general knowledge.

The last sentence matters. Without it, a chatbot that cannot find the clause tends to fill the gap from memory, and you cannot tell the difference from the outside.

A window in a bathroom wall near a boundary

Say a builder is adding a bathroom window in a wall close to a neighbouring property. Two code ideas collide: ventilation and light on one side, fire spread across a boundary on the other. A chatbot is good here at saying the two topics exist and that the second often limits glazing near a boundary. It is poor at telling you the distance and the permitted unprotected area, because those depend on the edition and the building's construction. The routine says to take the topic names, read both sections, then ask the authority for the permitted openings. If the chatbot had simply said "fine, 1 m is enough", the builder would have walked into a failed inspection.

Rewriting a vague basement question

Before: "What is the minimum ceiling height in a basement?" The answer will be a single figure, probably blended from several jurisdictions. After: "I am finishing a basement in a detached house to create a bedroom and a living area. List the headroom, egress window, smoke alarm and damp-proofing topics I need to look up, and tell me which differ for a habitable room versus a utility room." The second prompt yields a reading list, and the answer to the original question comes from the document, with the edition recorded.

The joist table that matched no edition

This one is an illustration, but it is an easy trap to fall into. A builder asks a chatbot for the maximum span of a timber joist for a new floor and gets back a tidy table of figures, then uses it to order timber. At inspection, the span tables in the adopted edition assume a different strength grade and a heavier dead load for the planned floor finish, so the joists are one size too small. The chatbot has not lied; it has produced a plausible table that matches no single edition. Why models produce confident tables like this is covered in AI hallucinations explained for business owners. The routine would have caught it at step three, because structural tables are the clearest example of numbers you never copy from a chat window. Joist spans should come from the published span tables the authority accepts, or from an engineer's calculation.

Preparing for an inspector visit

This is the safest, most valuable use. Ask the chatbot for a pre-inspection checklist built from your description of the stage.

Here is a description of the stage I am about to call for inspection:
foundations for a single-storey rear extension, strip footings
600 mm wide, 1.0 m deep, poured, rebar in place, drainage run
crossing under the footing in a sleeve.

List what an inspector is likely to check at this stage, as a
checklist with a tick box each. For each item, say what document or
measurement I should have ready. Flag any item where you are not
sure it applies to every authority.

The output is a checklist you then compare with the authority's own inspection list. You will find items the chatbot missed and items it invented, and that comparison is itself a useful way to learn what the authority cares about. If you want the corrected list on a phone at the inspection, a site visit checklist built in Google Forms shows a simple way to set it up.

How to ask the building department so you get a useful reply

Departments are busy, and a vague email gets a vague reply or none. A chatbot can help you draft a clear one, which is a point where it clearly earns its keep. Keep it to one question per email, attach the relevant drawing, and name the clause you have read. A good message follows this shape.

Subject: Application [reference] - confirming stair requirement

Hello,

We are converting a loft at [address, or job reference] under
application [reference], submitted on [date]. The new stair serves
one bedroom with an en-suite. I have read [clause reference] on loft
stairs and want to confirm two points:

1. Which edition and amendments apply to this application?
2. Does the single habitable room condition in [clause reference]
   still apply when the room has an en-suite?

Drawing [number] is attached. A short reply by email is fine, and I
will keep it on the job file.

Thank you,
[your name]

Notice the message shows you have done the reading. Departments tend to answer a reader's precise question faster than they answer "what are the rules for loft stairs?". Keep their reply, because it is the record that protects you if a different officer reviews the job later.

Prompt habits that lower the error rate

• Give the edition if you know it. "Answer for the 2021 edition" at least stops it blending, though it cannot stop it misremembering.
• Ask for confidence markers. "Mark each statement as quoted from the attached text, or from general knowledge." This separates grounded answers from guesses.
• Ask it to argue the other side. After an answer, ask "What would an inspector challenge here?" The reply often exposes the assumption you both made.
• Keep one topic per chat. Long mixed conversations drift, and early assumptions leak into later answers.
• Start fresh when you doubt it. Ask the same question in a new chat, or in a second chatbot. Disagreement tells you exactly which point to read in the document.

Disagreement between two chatbots is a useful signal, but agreement is not proof. Both can share the same wrong blend of editions.

Where purpose-built code software fits the routine

Search for this topic and most of what comes back is software built for code questions rather than a general chatbot. UpCodes, for example, sells UpCodes Copilot, which it describes as an AI-powered compliance research assistant. CodeComply sells AI code checks on drawings to design professionals, lists awareness of local amendments as a feature, and offers separate precheck and plan-review products to the authorities who review submissions. Tools like these answer from the code text they hold, so they suffer less from the edition-blending that trips up a general chatbot.

They speed up steps one to three of the routine; they do not replace steps four and five. The tool still has to hold the edition your authority adopted, and you still need to confirm that it does. Its answer is still a reading of the code rather than the authority's interpretation of it, and the person who signs the drawings still carries the decision. Before paying, test any such tool on five questions you have already settled with your own building department. Score each answer on three points: did it name the right section, did it quote the right figure for your adopted edition, and did it show where the figure came from? A tool that scores well and shows its sources is a fast index. One that gives right-looking answers without sources is still only a pointer, because you cannot see where the figure came from. Pricing is on the vendors' own pages or comes through a sales demo, so weigh it against the fifteen minutes the routine already takes per question.

What stays with you, however good the tool gets

Responsibility for compliance sits with the builder and the designer, and the authority's approval is the one that counts. No chatbot answer changes that. Nor does it change what your insurer or a client's solicitor will ask if something goes wrong: which edition, which clause, who confirmed it. If your answer to that is "the chatbot said so", you are in a weak position. If it is "clause reference, edition, email from the department dated such-and-such", you are in a strong one.

It is also worth separating code questions from design questions. Where the answer affects structure, fire safety, means of escape or energy performance, the right person is the designer or engineer who signs the drawings. A chatbot may help you understand what they told you; it should not replace asking them. The same goes for products: a manufacturer's certificate or technical data sheet beats a chatbot's summary of it every time.

Reading a clause properly once the chatbot has pointed you to it

Pointing is the chatbot's job; reading is yours, and reading a code clause is a skill with a few habits worth stating. Start with the scope sentence at the top of the chapter, because it says which building types the chapter covers, and a stair rule for dwellings may not be the one for a shop. Next read the definitions, since words such as "habitable room", "storey" and "private stair" are defined terms with narrower meanings than everyday speech. Then read the clause itself, the table it points to, and the exception that usually follows it. Many mistakes come from stopping after the first sentence.

Keep a pencil note beside each requirement: the figure, whether it is a minimum or a maximum, and where it is measured from. A going measured at the centre line of a winder is not the same as one measured at the narrow edge, and the chatbot will happily blur the two. When the clause refers to another standard, write down its name and edition and ask the authority whether that referenced standard is accepted. If you ever find that the chatbot's summary disagrees with the clause you are holding, the clause wins, and it is worth telling the chatbot so in the same chat, because it stops it repeating the error to you in that conversation.

Fifteen minutes of checking against a replaced window

Suppose, as an illustration, that a builder relies on a chatbot figure for the clear opening of an escape window in a basement bedroom. The window is ordered, fitted and the plasterboard is up before the inspector measures the opening and finds it short of the required clear area. Replacing the window and making good might take two days of two people plus materials, which at an assumed combined labour rate of $90 an hour for 16 hours is $1,440 before the window itself and the delay to the next trade. Against that, the routine costs perhaps fifteen minutes of reading and a five-minute email. Even if the routine only prevents one such error a year, it pays for itself many times over. The point of the sum is not the exact figures, which are assumptions, but the shape: checking is cheap, and rework discovered at inspection is expensive.

Keeping client details out of code questions

Describing a situation does not require pasting a client's name, address or contract. Strip those out. Use the job reference or "the client" and describe the building in general terms. On consumer plans, chats may be used to improve the models unless you switch that setting off in the privacy controls, while business plans generally do not train on business content by default. Check the setting on whichever plan you use before pasting drawings or reports.

Setting up a simple routine for a small crew

Write the five steps into a one-page procedure and give it to anyone who asks a chatbot about the code. Add a short log: date, question, sections found, edition confirmed, who confirmed. After a month you will have a record of the questions that keep recurring, which is the best way to decide what deserves a proper checklist. The same list shows which departments reply quickly and which need a phone call. If your jobs often involve unusual conditions, consider a shared Project in ChatGPT or Claude, loaded with your licensed code text and a standing instruction to quote with page numbers and refuse to answer from memory. That helps the quality of the pointers; it does not alter the rule that the authority decides.

If you want the routine written into how your crew works, that is the kind of job my AI implementation consultation covers: get in touch here. For the wider picture of where AI fits a building business, see the AI guide, and for the contract side of the same find-then-verify habit, read what builders should test in contract review software. When a code answer changes the scope, pricing the change accurately is the next problem.

Questions builders ask about AI and the code

Can I upload the code book PDF and trust the answers?

It improves the odds, because the chatbot can quote your own copy, but it does not remove the risk. Long documents get skimmed, tables and exceptions get missed, and the PDF may be the wrong edition. Treat the quoted section number as a pointer, open the page yourself and read the whole clause, including its exceptions.

Which edition of the code applies to my job?

The edition adopted by your local authority on the date your application was made, plus any local amendments. A chatbot cannot know that. Ask the building department or building control, and write the edition and the date of the reply on the job file.

Is it fine to use AI to prepare questions for the building inspector?

Yes, and it is one of the best uses. Give it the situation and ask for the clauses that probably apply and the three questions an inspector would want answered. Then send those questions to the department and let them answer.

Should I tell the client I used AI to research a code point?

You do not need to mention a search aid any more than a search engine, but never present a chatbot answer as a ruling. What you tell the client is the confirmed position from the authority, with the clause reference and the date you confirmed it.

Further reads

• Construction Contract Review With AI: Which Clauses It Flags and Misses — Same find-then-verify habit applied to contract clauses.
• ChatGPT for Building Design: Briefs, Options and Code Checks — Using ChatGPT earlier, at brief and option stage.
• AI in Home Building: Buyer Enquiries, Warranty Calls and Handovers — Where code questions surface in enquiries and handover.
• Best AI App for a Construction Company: Logging Near Misses — Log the near misses a wrong assumption can cause.
• Simpro Form Builder: Draft Job Sheets and Safety Checklists With ChatGPT — Turn your code checks into a reusable job sheet.
• How Small Builders Use AI to Price Variations and Change Orders — Price the change when a code answer alters the scope.
• Best AI Tool for Construction Design: What It Drafts and What Needs a Stamp — Which AI design outputs a small builder can issue (sketches, option comparisons, specification wording) and which need an engineer's or architect's sign-off.
• How to Write a Method Statement for a Small Build, With AI Help — Draft a small-build method statement with AI: sequence, plant, hazards, controls, a competent person's sign-off and the errors to catch before work starts.
• Loft Conversion Quote Template: Stages and Provisional Sums With AI — How to split a loft conversion quote into fixed price, allowances and provisional sums, with a filled-in stage table and two AI prompts.
• AI for Electrical Code: Can ChatGPT Answer Code Questions Safely? — How an electrical firm can test an AI on ten known code questions, catch edition and amendment errors, and confirm each answer in the adopted code book.
• AI for Home Construction: Where It Helps a Builder, Plot to Handover — Seven build stages from enquiry to handover, with the job AI can take and the job that stays yours at each, plus which stage to try first.
• Can AI Design a Fire Alarm System? What an Installer Must Still Do — AI can draft fire alarm paperwork and checklists, but design, commissioning and certification stay with a competent person working to the applicable standard.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general building-regulation practice (adopted editions, local amendments and the building department's role); Claude upload limits and ChatGPT and Claude privacy settings (vendor help pages); UpCodes Copilot and CodeComply product pages.