Treadstone Associates
Article · 9 min read

Can AI answer building code questions?

It can help you find and read the right provision. It cannot tell you what the code in force says, because there is no single Canadian building code — there is a national model code and a set of provincial codes that adopt and amend it, and a general-purpose model has no reliable idea which one applies to you.

Treadstone Associates · Updated 2026

Key takeaways

  • • Canada has model codes published by the National Research Council, and separate provincial codes that are the ones actually enforced.
  • • Ontario's Building Code Compendium is a specific dated document. Ask an AI a code question without telling it which edition applies and you are asking it to guess.
  • • Retrieval over a code document you have licensed is the usable pattern: the model quotes and locates, you read the provision.
  • • Enforcement is by municipal building officials, and technical disputes go to the Building Code Commission — not to whichever tool produced the answer.

Yes, with a large qualification that changes how you should use it. AI is genuinely good at finding the provision, restating it in plain language, and telling you what else in the document is related. It is unreliable at knowing which document governs your project, and in Canada that is the entire question.

The failure is structural rather than a matter of the model being weak. Ask "what is the minimum stair width" and there is no Canadian answer to give — only an Ontario answer, a British Columbia answer, and a national model-code answer that binds nobody by itself.

Why a Canadian code question is harder than it looks

The Codes Canada group at the National Research Council publishes the National Model Codes on behalf of the Canadian Board for Harmonized Construction Codes. Its publications page lists the 2025 editions of the National Building Code, National Fire Code, National Plumbing Code and National Energy Code for Buildings, with the 2020 editions still listed alongside them — and it tells you directly to contact the appropriate provincial or territorial official to find out which codes and regulations apply in your area.

That instruction is the whole problem in one sentence. Provinces adopt the model code on their own timetable and amend it. Ontario publishes its own Building Code Compendium — the province notes the 2024 Compendium is available through Publications Ontario and tells users to check the filing date on the title page against the most recent version, because it has been reissued. British Columbia maintains its own codes separately again.

So a model trained on a broad sweep of internet text has seen: US code text, multiple editions of the national model code, several provincial codes, and a great deal of forum discussion that mixes all of them. Asked a bare question, it produces a confident average. An average of several codes is not any code.

The tell-tale wrong answers

If an answer cites an "IBC" section, it has given you the International Building Code, which is a United States document and has no force in Canada.

If it quotes a number without naming an edition and a jurisdiction, treat the number as unverified. "The Building Code says" is not a citation; "Division B of the 2024 Ontario Compendium, at the page you are about to open" is.

If it cites a code section that resolves to nothing when you look it up in the actual document, you have found a fabricated reference. This is the single most common failure mode on legislative questions, and it looks exactly like a correct answer.

The pattern that works: retrieval, not recall

The distinction that makes AI safe here is between asking a model what it remembers and asking it to work over a document you have supplied. The second is reliable in a way the first is not, because every sentence it produces can be traced to a page you can open.

In practice that means licensing the code document your province enforces, and pointing the tool at that file. Then the useful questions become: where in this document is fire separation between suites dealt with; show me every clause in this part that mentions guards; summarise the difference between these two sections in plain language; what defined terms does this clause depend on. Each of those is a search-and-restate task, and each answer arrives with a location you can verify in under a minute.

What you should not do is ask it to conclude. "Does my design comply" is not a retrieval question. It is a determination, and the determination is made by someone with statutory authority.

Who actually decides

Ontario is explicit that enforcement of the Building Code, including issuing permits, is carried out by municipal building departments. The province adds that before you hire anyone to prepare design plans for a permit application you should confirm they are registered on the QuARTS public registry — registered practitioners carry a Building Code Identification Number, and the province warns not to assume an unregistered person can support your application.

If the disagreement is technical, there is a defined route. Where a permit is refused over the technical requirements of the Building Code, the appeal is to the Building Code Commission, an independent adjudicative tribunal; where the refusal concerns other applicable law, such as the interpretation of a zoning by-law, the route is to a judge of the Superior Court of Justice. Neither forum has any interest in what a chatbot said.

The Building Code Commission route carries a hard stop worth knowing before relying on any AI-assisted reading of the code: under the Building Code Act, 1992, s.24(4), “the decision of the Building Code Commission is final” once it rules on a dispute over the technical requirements of the code — there is no further appeal on the merits. An AI tool that helped locate the wrong provision at the reading stage cannot be unwound at the hearing stage.

A worked example

A renovation contractor in Ontario is converting the second storey of a mixed-use building and wants to know what the code requires for the exit from the residential unit above a commercial ground floor.

A bad workflow: type the question into a general assistant, receive a specific-sounding answer with a section number, price the work on it. The number may be from the model code, from a superseded edition, or from nowhere at all.

A workable one: open the edition of the Compendium your municipality enforces. Ask the tool to locate every provision dealing with exiting from a dwelling unit in a building of that occupancy and construction type, and to list them with their locations. Read those provisions. Draft your question for the plans examiner from what you have read, and put it to them before the application goes in — the province recommends talking to the building department before applying, precisely so that you learn what drawings and approvals will be required. The AI cut the reading down; the building official still decides.

Where the code question becomes a legal question

There is a point past which this is not a code question at all. If work is already built and the argument is about who pays to fix it, you have moved from interpretation into remedies, and the two run on different clocks and different forums — our sister firm sets out the practical difference in building code complaint versus suing your contractor in Ontario.

Common questions

Is there one Canadian building code?

No. There are national model codes published by the NRC and provincial codes that adopt them with amendments. The provincial one is the one enforced against you.

Can I upload the code PDF and ask questions about it?

Technically yes, and that is the pattern worth using. Check your licence terms for the document first — code publications are copyright works, and what you may do with a copy is set by the terms you bought it under.

Will an AI answer stand up with a building official?

No, and it is worth not being the person who tries. What stands up is the provision itself, the stamped design, and the reviewer’s determination. Ontario’s Building Code pages are the starting point for what the province currently publishes.

Put AI where it helps and keep it away from where it does not.

A 30-minute call is enough to tell you whether AI pays for itself here.