AI for Building Code & Zoning Checks: What Works in 2026

There is no AI building code compliance check you can rely on in 2026 - published disclaimers from CodeComply.Ai and UpCodes leave final determinations with qualified humans, and our 20 October 2026 search found no independent accuracy benchmark for any of these tools. What does work, paired with a verification workflow: chat assistants for structuring research on code text you control, code-database assistants (UpCodes Copilot, ICC AI Navigator) for jurisdiction-aware section lookup, and plan-review platforms (CodeComply.Ai, Archistar) for pre-submittal screening.

This guide compares those three tool classes, names the four failure modes that survive a casual check, reports our own October 2026 prompt tests, and ends with a 5-step verification workflow built on the prompts we actually ran. We do not cover BIM-integrated model checkers (Solibri and its kin) or energy-code engines in any depth; both deserve their own article.

Prices and features were verified on 20 October 2026. They change often, so check the vendor page before you quote a figure to a partner.

Can AI check building code compliance?

No, if you mean an AI building code compliance determination you could sign off on. Yes, if you mean code research, section lookup and pre-submittal screening, each followed by a verification step a licensed person performs. Even CodeComply.Ai, which sells “AI plan review” to building departments, leaves the final call to a qualified human and writes that into its public materials.

The disclaimer language is worth reading in full, because it comes from a company whose product is marketed as automated plan review. The CivicPlus-CodeComply.Ai partnership press release of 19 March 2026 states: “Automated compliance checks are informational and do not replace professional review or constitute legal advice. Final compliance determinations remain the responsibility of qualified reviewers.” That sentence is the contract you are working under with every tool in this article, whether it is printed on the page or implied by the licence terms.

Adoption is running well ahead of evidence. According to the RIBA AI Report 2025 (survey January-April 2025, published June 2025), 59% of UK practices use AI at least occasionally, up from 41% in 2024, and 54% have adopted AI for standards and regulatory compliance checking at least rarely. RIBA itself notes the sample is self-selecting, so practices already interested in AI are over-represented.

The US picture is more cautious. The AIA’s adoption study (March 2025, with Deltek and ConstructConnect) found only 6% of US architects regularly use AI, 8% of firms have implemented it, and about 90% of professionals are concerned about inaccurate outputs. Those two surveys describe the same profession a few months apart, and the gap between them says a lot about how loosely “using AI” is defined.

As of our search on 20 October 2026, no published independent accuracy benchmark exists for building-code citation by chat assistants or by the commercial platforms - the part the marketing pages leave out. Every accuracy figure published for these commercial tools was produced by the vendor, on a test set the vendor chose, with a scoring method the vendor did not publish. We say this plainly because the liability sits with you, and a vendor-run number is no defence in a claim.

The three tool classes - and which part of the job each covers

Three classes of AI building code compliance tools exist today, and they do different jobs. Chat assistants (Claude, ChatGPT, Gemini) research and structure code text you paste in yourself. Code-database assistants (UpCodes Copilot, ICC AI Navigator) look up sections inside a maintained database that knows which edition and amendments a jurisdiction has adopted. Plan-review platforms (CodeComply.Ai, Archistar, Symbium) screen uploaded plan sets against digitised rules, and are mostly sold to governments.

Criterion	Chat assistants (Claude, ChatGPT, Gemini)	Code-database assistants (UpCodes Copilot, ICC AI Navigator)	Plan-review platforms (CodeComply.Ai, Archistar, Symbium)
What it works from	Model knowledge plus whatever you paste	Maintained code database with adoptions and amendments	Your uploaded plan set checked against digitised rules
Jurisdiction awareness	None by default	UpCodes: per adoption, including amendments; ICC: current I-Codes, can be scoped to selected US states	Per configured jurisdiction
Typical failure you must catch	Invented or superseded section numbers	Answers limited to database scope	False confidence in a screening report
Price	Free tiers exist; paid plans from $20/month	UpCodes Professional $59 per user per month billed annually, $68 monthly; ICC AI Navigator requires a Digital Codes Premium subscription	CodeComply.Ai from $250/month (single user); Archistar sold to municipalities
Who it’s for	Any architect, with the verification workflow below	Firms doing regular multi-jurisdiction code research	Submitters and building departments

Chat-assistant column tested by us: 2026-10-20, via API access to current Claude and GPT model families. Code-database and plan-review columns: vendor documentation, October 2026 - we have not independently tested those products.

UpCodes Copilot is an AI research assistant layered on the UpCodes database of adopted codes and amendments. Per UpCodes’ own feature page (read 20 October 2026), generic AIs “work from out-of-date code” while Copilot pulls from jurisdiction-specific adoptions. The same page reports vendor-run practice-exam scores: 77% on an ARE 5.0 PDD practice exam, 94% on a plans-examiner practice exam and 96% on a commercial-inspector practice exam, dated April 2024. Those are practice-exam runs by the vendor, never independently audited, and an exam score tells you little about performance on a real project with a city amendment in play.

Pricing as of 20 October 2026: Essentials at $39 per user per month billed annually ($45 monthly) includes 3 Copilot questions; Professional at $59 ($68 monthly) includes unlimited Copilot; a 2-week trial is available. If you already live in the UpCodes database, the question is whether unlimited lookups justify the step up from Essentials.

ICC AI Navigator is the code publisher’s own LLM assistant inside Digital Codes Premium, per ICC’s support documentation (read 20 October 2026). An LLM, or large language model, is the text-prediction engine behind tools like ChatGPT; here it sits on top of the I-Code text ICC itself publishes. It answers from the most recent I-Code editions by default and can be scoped to selected US states, and ICC positions it for basic code questions. It requires an active Premium subscription, so the cost is the subscription you may already carry, and it still will not catch a city amendment that lives outside the I-Codes text ICC publishes.

CodeComply.Ai takes a PDF plan set, converts the drawings to structured data, and flags issues with code citations pinned to specific sheets. The vendor lists IBC, IPC, IMC, NFPA 101, ADA, FHA and state-amended codes, from $250/month for a single-user licence (vendor FAQ, 20 October 2026). The vendor also claims “98% accuracy”. No public methodology backs that number, so treat it as marketing until an independent audit exists.

Archistar AI PreCheck is a government-side plan-review tool, which means architects meet it as applicants rather than as customers. The City of Austin reports “roughly 50% time-savings from the zoning perspective” (Leah Hilton, Zoning Plans Examiner Supervisor, Archistar case study, read 20 October 2026). Surrey, BC launched it publicly on 10 March 2026.

Jerome Thibaudeau, Surrey’s business transformation manager for planning and development, said in Archistar’s Surrey announcement that “roughly 80% of residential permit applications contained significant zoning deficiencies, requiring an average of 1.6 resubmissions” and called the tool “like ‘spell-check’ for zoning”. That 80% figure is a comment on the submissions, and it explains why municipalities are buying.

Symbium does computational-law checks for focused project types such as solar, electrification and accessory dwelling units, and is strongest in California (symbium.com, 20 October 2026). Computational law means the ordinance has been rewritten by hand as machine-checkable rules, which is why coverage is jurisdiction-limited by design. Outside its mapped cities and project types it has nothing to say.

The four failure modes that survive a casual check

The dangerous AI code errors are the ones that look right. Correct values attached to superseded section numbers, stale editions, silently assumed jurisdictions, and local amendments the model has never seen - each one reads as a plausible answer and passes a quick skim by someone who knows the code. The obviously wrong answers are easy to catch; these four are the ones that reach a drawing set anyway.

Failure 1 is the right value with the wrong citation. In our 12-task Claude vs ChatGPT test (18 October 2026), Claude returned all three correct egress values for a 2021 IBC office question and cited the corridor-width table as 1020.2. That is the 2018 edition’s number; 2021 moved it to 1020.3.

The same answer carried a warning about renumbering between editions, which makes the error more insidious: the model knew the risk and still walked into it. A right answer carrying a dead section number survives exactly the casual check you would normally run, because the number you remember matches and the citation looks like a citation.

Failure 2 is the stale edition. The IBC business-area occupant load factor changed from 100 gross to 150 gross in the 2018 edition, per the Georgia DCA summary of 2018 IBC key changes (2018 adoption cycle, opened 20 October 2026). A model answering from pre-2018 training text gives you 100 and sounds completely sure, and 100 was the right answer for long enough that many of us still carry it in our heads. Edition drift is invisible unless you ask for the edition with every number and then check that the edition matches your jurisdiction’s adoption.

Failure 3 is the wrong jurisdiction default. A model answering a code question with no jurisdiction stated has to pick one, and nothing forces it to pick yours; we did not run an unnamed-jurisdiction control, so we treat this as a risk to manage rather than a measured failure. The AEC consultancy Resolved tested Claude on six real tasks (17 March 2026); their verdict on code work was “research accelerator, not a sign-off” (Reid McConkey and Chris Briggs, Resolved). We reached the same conclusion from a different set of tasks.

Failure 4 is the invisible local amendment. Even a model answering from the correct edition misses state and city amendments, and that is exactly the gap UpCodes and ICC market their tools to address. The same Resolved test flagged this directly: Claude’s knowledge “may not reflect the most recent code amendments or local municipal interpretations”.

Zoning is worse. Ordinances are municipal, often PDF-only, frequently amended by council motion, and often missing from, or out of date in, a general model’s training text.

The academic record points the same way. The ARCHER paper (2026) measured naive single-pass LLM prompting on building-compliance rule checking at 0.4681 mean union accuracy versus 0.8498 for a deterministic multi-agent harness, which is a structured pipeline where the model’s output is checked by fixed rules rather than trusted directly. The P4IR paper (2026) states LLM approaches are “prone to generating incorrect and hallucinated computer-processable rules”.

An October 2025 case study found agent pipelines calling a deterministic engine (the DOE COMcheck API) more reliable than retrieval-augmented (RAG) reasoning over code text, on energy-code checks. In ARCHER’s comparison, raw single-pass prompting - what most of us do when we open a chat window and type a question - scored lowest.

What our own tests showed in October 2026

On 20 October 2026 we ran seven code and zoning prompts, ten runs in total: three prompts through both the Claude and GPT model families, four through Claude only, via API access - the same models behind the chat apps, called programmatically. Both families passed an edition trap and a jurisdiction trap when the prompt named the edition and the jurisdiction. These were single runs with no statistical weight.

The Claude runs used the Sonnet tier; the GPT runs used the Astra reasoning tier with web search available. The lesson sits in the conditions above.

Test 1, the edition trap. We asked for the IBC 2021 occupant load factor for general business areas, with exact section numbers. Both models answered 150 gross per Table 1004.5 and both added the Section 1004.8 concentrated-business rule, which is correct. Claude flagged the 100-to-150 history on its own, without being asked.

Two days of tests, two different outcomes. In our 18 October 12-task test the same Claude tier attached a 2018 section number to a correct 2021 value. You cannot predict which answer carries the stale citation. That unpredictability is the whole argument for logging and verifying every citation on every run, including the runs that look clean.

Test 2, the jurisdiction trap. We asked an escape-stair question for a four-storey office in Manchester, UK. Both models reached for Approved Document B Volume 2, Section 3, Tables 3.1 and 3.2, with no IBC default creeping in. The GPT run cited the live gov.uk PDFs. It also refused to pick a stair width from storey count alone, and that refusal is the right answer, because the tables depend on occupant numbers and the number of stairs available.

Both passed with the city named in the prompt; we did not run a control without it. The risk we did not test: without a named jurisdiction, a model may default to another country’s code - so never leave the jurisdiction out.

Test 3, the restricted-text zoning screen. We pasted a fictional ordinance section with planted values and a proposal that violated one requirement. One further requirement could not be determined from the text at all. Both models produced a correct five-row screen: they flagged the height violation, marked floor area ratio as “unclear from provided text” instead of guessing a value, and neither claimed that anything “complies”. Restricting the model to pasted text worked in both families.

Full transcripts are archived with dates and model identifiers. These were single runs of two model families, under conditions we controlled, on a date we can name. Read them as a field note.

A 5-step verification workflow for AI code research

Treat every AI code answer as an unverified draft with a paper trail. The workflow: fix the jurisdiction and edition first, use AI to build the research map, restrict it to text you paste, log every citation into a verification table, then check the log against the adopted code and its local amendments before anything reaches a drawing. This is what AI building code compliance work looks like when the liability stays with you.

Step 1: Pin the jurisdiction and edition before asking anything

Write both down in the project file and state both in every prompt. Our Manchester test named the city, and both models used the right code. Left to guess, a model has to pick some jurisdiction, and nothing forces it to pick yours - so take the guess out of the question.

Step 2: Use AI to build the research map

The answers come later, from the adopted text. At this stage you want a list of chapters to open and questions to answer, with the numbers left blank. Prompt (tested 20 October 2026, Claude family only):

I am an architect starting code research for a project. Act as a research assistant, not a compliance checker. Project: [one-paragraph description - building use, storeys, construction type if known, jurisdiction and adopted code edition]. Produce: (1) a list of the code chapters and section families I most likely need to review for this project type, each with one line on why; (2) for each, the project-specific question I need to answer; (3) a verification list - what I must confirm in the adopted local code and amendments before relying on anything, including the edition year and renumbering risk between editions. Do not state final requirements as settled numbers; frame everything as sections to open and verify.

Step 3: Restrict the model to text you control

Paste the ordinance or code excerpts and forbid outside knowledge. This closes off the stale-edition and wrong-jurisdiction failures at the same time, because the only text in play is the text you chose. Two tested prompts (the first tested 20 October 2026 in both model families, the second in the Claude family only):

Here is the text of a zoning ordinance section for my site, and my proposed building parameters. Using ONLY the ordinance text I provide, build a table: each requirement with its clause number, my proposed value, and a flag "appears to meet", "appears short" or "unclear from provided text". Do not state that anything complies; this is a first-pass screen. If a requirement cannot be evaluated from the data given, say so.

Before I analyze a zoning ordinance, extract every defined or ambiguous term from the ordinance text I paste, and for each term tell me: where it is used, whether the pasted text defines it, and what I need to look up in the full ordinance's definitions chapter before relying on any analysis. Use ONLY the pasted text; do not supply definitions from general knowledge.

A note on the second prompt. In our run it caught that “35 feet maximum height” is meaningless until the ordinance says where height is measured from - the transcript’s words: without the starting point, “‘35 feet’ is meaningless”. That is the kind of gap that sinks a first-pass screen and only shows up at the planning counter.

Step 4: Log every citation before you trust it

The log is a table you fill in by hand at the source. The model builds the empty table; you supply the status column. Prompt (tested 20 October 2026, Claude family only):

I asked an AI assistant some building-code questions and got answers with section numbers. Turn the answers I paste below into a verification log table with columns: claim, code edition cited, section cited, what to check at the source, risk if wrong, status (blank for me to fill). Add a final row for jurisdiction amendments. Do not tell me whether the claims are correct - the log is for my own source check.

Step 5: Check the log against the adopted code and its amendments

Open the adopted edition, open the state and city amendment documents, and close out each row. When the adopted edition is unclear, a phone call to the building department settles it faster than any search. Prompt (tested 20 October 2026, Claude family only):

I am an architect preparing to rely on building-code research. Build me a local-amendment checklist for the following code sections: [code edition and section numbers], in a US jurisdiction whose state adopts the IBC with amendments. List, as a checklist: (1) where state and city amendments are typically published; (2) the specific questions to ask about each of these sections; (3) the renumbering risk - which of these section numbers changed between editions; (4) who to contact when the adopted edition is unclear. Frame every item as something I verify, not something you confirm.

Liability. Building codes, zoning ordinances and their local amendments are jurisdiction- and date-specific, and the responsibility for compliance sits with the licensed professional who stamps the drawings. Every AI output in this workflow - section numbers, values, screening tables - is a research draft to verify against the currently adopted edition in your jurisdiction. The failure modes to catch are concrete: hallucinated or superseded section numbers, stale editions, unstated jurisdiction assumptions and missed local amendments. Nothing an AI tool returns reduces the duty of care, and no tool on this page determines compliance.

The same verification discipline applies to specification text, where a wrong ASTM number carries the same kind of risk as a wrong section number - see our AI spec-writing workflow.

What you can do for free

Free tiers cover research and structuring on text you paste; jurisdiction-aware lookup needs a paid plan. Chat assistants have usable free tiers, UpCodes offers a 2-week trial (its Essentials plan includes only 3 Copilot questions), ICC AI Navigator sits behind a paid Digital Codes Premium subscription, and neither CodeComply.Ai nor Archistar lists a free tier on its vendor pages. Prices below were checked on 20 October 2026.

Tool	Free tier	Real limit	Enough for
Claude / ChatGPT free tiers	Yes	Usage caps (vary by vendor and load)	The restricted-text prompts above, run on one ordinance section at a time
UpCodes	2-week trial	Essentials at $39/mo (annual) includes 3 Copilot questions; daily use needs a higher plan	Evaluating whether Copilot fits your jurisdictions
ICC Digital Codes	Free basic code viewing	AI Navigator requires Premium	Reading the adopted model code text at the source
CodeComply.Ai / Archistar	None listed on vendor pages (20 Oct 2026)	CodeComply from $250/month single user (vendor-stated); Archistar is sold to municipalities	Archistar reaches you as an applicant when your city runs it

Most of the AI building code compliance workflow above runs on free tiers. The free path that works this week: run the step-3 restricted-text prompts in a free chat tier on your own ordinance text, one section per conversation. Our Claude code-check workflow is the single-tool version of the same routine, and the prompt library covers neighboring tasks such as area schedules and accessibility checklists.

FAQ

Can I trust ChatGPT for building code questions?

Do not rely on it for final answers. Invented and superseded section numbers are documented, and in our own tests a Claude tier gave a stale citation on 18 October and a clean answer on 20 October - single runs, and nothing in our runs rules out either family producing one. Treat it as a research accelerator: log every citation, check it at the source, and name the edition and jurisdiction in every prompt.

What is the difference between UpCodes Copilot and asking ChatGPT?

The database underneath. Copilot retrieves from UpCodes’ maintained library of jurisdiction-specific adoptions and amendments, so the text it quotes is meant to be the text your city adopted. A generic chat model works from training text with a cutoff date and no amendment layer. Vendor-run exam scores exist for Copilot; our 20 October 2026 search found no independent audit of either product, and the price difference is just as real.

Can AI check zoning compliance?

This is the weakest area for generic models. Ordinances are municipal, often published only as PDFs, and full of terms the pasted section does not define. Restricted-text screening of pasted ordinance text worked in our 20 October 2026 tests, with both model families flagging what they could not determine. Government-side platforms such as Archistar screen applications in the cities that run them - Surrey calls it “spell-check” for zoning, a pre-screen rather than a determination.

Will AI replace plan reviewers?

In practice, these tools assist plan reviewers. Surrey’s own framing is “spell-check” for zoning - a pre-screen ahead of human examiners - and CodeComply.Ai’s partnership disclaimer leaves final determinations with qualified reviewers. Surrey’s figure of roughly 80% of residential applications arriving with significant zoning deficiencies explains why municipalities buy pre-screens. The legal authority to approve or reject a set of drawings remains with a person, and we see no jurisdiction moving to change that.

Which AI is most accurate for building codes?

The honest answer is that nobody knows. Our 20 October 2026 search found no independent accuracy benchmark for AI building code compliance tools, and every published accuracy figure for a commercial tool comes from the vendor whose product it measures. Academic work such as ARCHER and P4IR shows naive prompting of a language model scoring well below structured, rule-checked pipelines. Pick by workflow fit and verify everything.

Sources

• CivicPlus - CodeComply.Ai partnership announcement (19 March 2026): https://www.civicplus.com/news/nn/civicplus-brings-ai-building-plan-review-codecomply-ai/
• RIBA AI Report 2025: https://www.riba.org/work/insights-and-resources/ai-report/riba-ai-report-2025/
• AIA - Artificial Intelligence Adoption in Architecture Firms (March 2025): https://www.aia.org/about-aia/press/new-research-explores-perceptions-and-opportunities-artificial-intelligence
• UpCodes - Copilot features (vendor benchmark notes): https://cms.up.codes/features/ai
• UpCodes - pricing: https://up.codes/pricing
• ICC - AI Navigator support documentation: https://support.iccsafe.org/ht_kb/digital-codes/icc-ai-navigator
• CodeComply.Ai - product and FAQ: https://codecomply.ai/
• Archistar - AI PreCheck (Austin case study): https://archistar.ai/aiprecheck
• Archistar - City of Surrey launch (10 March 2026): https://www.archistar.ai/blog/city-of-surrey/
• Symbium - computational-law compliance platform: https://symbium.com
• Resolved - We tested Claude on AEC tasks (17 March 2026): https://resolved.ca/blog/we-tested-claude-on-aec-tasks-heres-what-wed-trust-ai-with/
• ARCHER - multi-agent building compliance checking (2026): https://arxiv.org/abs/2607.25566
• P4IR - LLM rule generation for automated compliance (2026): https://arxiv.org/abs/2606.22402
• Automated building code review case study (October 2025): https://arxiv.org/abs/2510.02634

Next time an AI answer hands you a section number, run it through the step-4 verification log before it reaches a drawing. And the newsletter brings a tested workflow like this one every week - one workflow, one tool, no hype. Subscribe.