AI falls flat wherever the right answer depends on something the model cannot see: the amendment your city made to the 2024 IBC (International Building Code), the beam behind the soffit, the corridor width in a plan it drew as a picture, or the fact that it does not know what it does not know. It is good at language: minutes, summaries, first drafts, scripts. It is unreliable at codes, geometry, numbers, and confidentiality, and it carries none of the liability.
That is the reality check in one paragraph. The rest of this post is the specific AI limitations architecture firms run into on real projects, sorted by how much they cost when they bite, and what the National Institute of Standards and Technology (NIST) suggests you do about them.
Where does AI get codes and zoning wrong?
At the edition and at the amendment. The ICC (International Code Council) publishes the 2024 I-Codes (the International Codes), and they are current, but each state and city adopts on its own schedule; many jurisdictions are still enforcing 2021 or 2018 editions with local amendments layered on top. California runs its own 2025 Building Standards Code, effective January 1, 2026. A chat assistant asked about corridor width or allowable height answers from whichever edition dominated its training data, and it does not say which one that was.
Zoning is worse, because it is local by definition. We have seen an assistant quote a setback from the wrong city with the same confidence it quotes the right one, and cite a section number that does not exist in either ordinance. The fix is not to stop using it. The fix is to treat every code or zoning statement as a lead, pull the adopted ordinance, and verify the section before it goes in the code summary. We wrote that up for zoning research and for drawing-set code review.
Why can’t AI reason about a floor plan?
Because a large language model (LLM, the kind of AI behind chat assistants) works on text, an image model works on pixels, and neither works on geometry. Ask an LLM to lay out an accessible restroom and it will describe one fluently and get the 60-inch turning circle wrong in the drawing, because it cannot measure. Ask an image model for a stair and count the risers against the floor-to-floor height; they will not match.
Details are the expensive version of this. An AI rendering of a rainscreen looks like a rainscreen and has no drainage plane, no flashing sequence, and no fastener. The AIA (American Institute of Architects) Trust and Victor Insurance both report water intrusion and building-envelope failures as consistently among the most frequent and costly sources of claims against design professionals, and nothing about a generated picture of a parapet changes the physics behind those claims. A detail is a sequence of decisions in order. Pictures do not have an order.
This is also why text-to-BIM is further off than the demos suggest. A model is thousands of dimensions that must agree with each other, and that is the thing these tools are worst at.
Can you trust the numbers it gives you?
Not without checking every one. LLMs make arithmetic errors, invent quantities, and, most dangerously for a profession that cites benchmarks, fabricate statistics with plausible-sounding sources attached. Ask for the average utilization rate for architecture firms and you may get a precise decimal and a study that does not exist. Deltek Clarity 2026 found 78% of architecture and engineering firms using generative AI, and research, one of its most common jobs, is exactly where fabricated citations do the damage.
Our rule: any number an assistant produces is a hypothesis until someone finds it in the source document. That includes the numbers on this blog, which is why every figure here carries an edition year and a URL.
What happens to confidentiality and liability?
Two separate problems, both yours.
Confidentiality first. A consumer chat tool’s terms may allow it to train on what you type, and a client under a nondisclosure agreement (NDA) for a site acquisition did not agree to that. NIST’s AI Risk Management Framework 1.0 (January 2023) organizes the work into four functions, Govern, Map, Measure, and Manage, and its Generative AI Profile (July 2024) adds the specific risks of tools like these, including data leakage, confabulation, and intellectual property. You do not need to adopt the whole framework. You need a list of which tools are approved for which classes of information, which is the heart of a firm AI policy.
Liability second. The standard of care does not move because the tool changed, and the AIA Code of Ethics and Professional Conduct holds that licensees seal only work they have reviewed and are competent to judge. If an AI-assisted code summary is wrong, the claim names the architect, and the AIA Trust’s risk guidance reads the same way it did before AI. On authorship, the U.S. Copyright Office’s Part 2 report (January 2025) states that purely AI-generated material is not copyrightable and human authorship is required, a position the D.C. Circuit affirmed in Thaler v. Perlmutter (2025). The design that is yours is the one a person made.
Why does AI flop on your firm’s own files?
Because it finds what is there. Point an assistant at the shared drive and ask for the firm’s standard parapet detail, and it returns the 2019 version, the 2022 version, and a version someone modified for a project in Phoenix, with no way to know which is current. Ask it how the firm handles additional services and it finds nothing, because the answer lives in a partner’s email.
AI does not improve content; it amplifies whatever quality the content already has. That makes the unglamorous work the prerequisite: audit the content library once a year, decide who owns it, and move the answers out of email into a knowledge base. The AIA Firm Survey Report 2024 puts about three in four U.S. firms at fewer than 10 people, and at that size nobody owns the library unless someone is told to. A firm with a clean library gets good answers from a mediocre assistant. The reverse is not true.
So what is it still good for?
Plenty, as long as you are honest about the category. Language tasks with a human reviewer: meeting minutes and action items, first drafts of proposals and RFI (request for information) responses, summaries of a 90-page geotechnical report, scripts for Dynamo and Excel, and research that produces leads rather than facts. Deltek Clarity 2026 found generative AI use rising from 64% to 78% of firms in a year, and that is where the adoption is.
The practical test for any new use is whether the answer depends on local facts, geometry, or numbers being right. If it does, the assistant drafts and a person decides. If it does not, let it run. Most of the trouble comes from firms that skipped the test.
AECPulse Project Manager keeps its AI assistants on the tasks they are good at (minutes, RFI drafts, and early warnings on budget drift), with a named reviewer on every output before it leaves the firm.