Compliance and Code-Checking Assistance
Deliberate Academy Editorial Team
Reviewed for accuracy and professional relevance
You're 4 lessons in — don't lose your progress.
Sign up free to save where you are and earn a verified certificate when you pass.
- Use AI tools to orient quickly to which codes, standards, and jurisdictional amendments likely apply to a given design condition
- Explain why an AI-generated code citation is a research lead, not a compliance determination, and what verification step closes that gap
- Apply a structured verification protocol before any AI-assisted code research enters a compliance memo, submittal, or plan check response
- Identify the specific failure modes of AI code research: fabricated section numbers, outdated editions, and missing jurisdictional amendments
A mechanical engineer preparing a plan check response for a hospital HVAC upgrade used an AI tool to identify which sections of the mechanical code and ASHRAE 170 likely governed the project's ventilation rate requirements. The tool returned a helpful, well-organized summary in about ninety seconds -- work that used to mean forty minutes searching through code text and cross-references. The summary also cited a specific ASHRAE 170 table number that, on verification, did not exist in the current edition; the actual requirement was in a differently numbered table with different values for the space type in question. That gap between orientation and verification is the entire subject of this lesson.
AI as a Code Research Orientation Tool
Codes and standards research is one of the most time-consuming parts of engineering documentation, and one of the areas where AI's speed advantage is largest. General-purpose tools like ChatGPT and Claude, and specialized platforms such as UpCodes, can quickly identify which code families likely apply to a described condition, summarize the general requirements of a code section in plain language, and flag related sections an engineer might not have thought to check.
This is valuable as a starting point precisely because it compresses the search phase of research -- narrowing from "which of thousands of code sections could apply" to "these five sections are worth reading in full." It is not valuable as an endpoint. AI tools do not reliably distinguish between code editions, do not reliably know which jurisdictional amendments apply to your specific project location, and will produce a specific-sounding section or table number with the same confidence whether that number is correct or fabricated.
Use AI code research the way you would use a knowledgeable but unverified colleague's memory: as a fast way to generate a shortlist of what to check, never as the check itself. A useful prompt pattern is: "What code sections and standards likely govern [condition] for a [building/facility type] in [general jurisdiction type, not a specific verification request]? List the code families and general topic areas, not specific section numbers you are not certain of." This reduces the tool's incentive to fabricate a specific-sounding citation and keeps its output at the orientation level where it is actually reliable.
A Missing Table Number in a Hospital Plan Check Response
Context
A mechanical engineer preparing a plan check response for a hospital HVAC upgrade used an AI tool to research ventilation rate requirements under ASHRAE 170. The tool's summary cited a specific table number and stated the required air changes per hour for the space type in question, formatted confidently alongside correct general context about the standard's scope.
Action
The firm's plan check protocol required every code citation in a submittal response to be checked against the actual, current standard text before submission -- a rule introduced after an earlier project experienced a rejected submittal over a similar issue. The engineer pulled the current edition of ASHRAE 170 and searched for the cited table number. It did not exist in that edition. The correct requirement was in a different table, with a lower minimum air change rate than the AI-cited figure had implied for a similar-sounding space type.
Outcome
The engineer corrected the citation and the ventilation rate before submitting the response, avoiding a plan check rejection that would have cost the project one to two review cycles -- typically four to six weeks on this jurisdiction's review schedule. The firm's technical director used the incident as a training example: the AI tool had not been wrong about which standard applied, only about the specific table number and value within it, which is precisely the kind of confident, partially correct output that a rushed verification step is most likely to miss.
An engineer asks an AI tool which fire-rating requirements apply to a stairwell enclosure in a mixed-use building, and the tool responds citing a specific IBC section number and a two-hour rating. The engineer includes this citation directly in a code compliance memo without checking the actual code text. What is the most significant risk in this workflow?
Select one answer.
A Verification Protocol for AI-Assisted Code Research
Before any AI-assisted code research enters a compliance memo, specification, or submittal response, verify: (1) the code family and edition -- confirm which edition is currently adopted in the project jurisdiction, since AI training data mixes editions from different years; (2) jurisdictional amendments -- many jurisdictions amend base codes (IBC, IMC, NEC) with local requirements that a general-purpose AI tool is unlikely to know unless specifically trained on that jurisdiction's amendments; (3) the specific section, table, or figure number -- locate it in the actual current text, not just in the AI's summary; (4) the actual requirement value -- confirm the AI's stated number (a rating, a rate, a dimension) matches the source document exactly, not an approximation; (5) applicability -- confirm the cited provision actually applies to your specific occupancy, space type, or condition, since codes frequently have exceptions and conditional applicability that a summary can miss.
Treat any AI-generated code citation that you cannot immediately locate in the actual current standard or code text as fabricated until proven otherwise, not as a citation you failed to find. This inverts the natural instinct to assume the tool is right and your search was incomplete -- AI-generated section and table numbers that do not exist in the current edition are a known and common failure mode, not a rare edge case.
A specifying engineer wants to use an AI tool to identify which NFPA standards likely apply to a proposed battery energy storage system installation. Which use of the tool best matches the professionally appropriate pattern from this lesson?
Select one answer.
Exercise
Your Task
Choose a code or standard research question relevant to your current work. Ask an AI tool the orientation-style question from this lesson's tip callout -- asking for the general code families and topic areas that likely apply, not specific section numbers. Then independently locate and read the actual current standard text for the topic areas identified. Write down: which topic areas the AI correctly identified, whether any section or table number it mentioned (if it offered one despite your prompt) matched the actual current text, and one thing the actual code text told you that the AI's summary did not capture accurately.
Success looks like
- You independently located and read the actual current standard text, not just the AI summary
- You can identify at least one detail (an exception, a conditional applicability clause, a specific value) that the AI summary omitted or got wrong
Watch out for
- Treating the AI orientation step as sufficient and skipping the independent verification step because the summary looked complete
- Assuming a section number the AI provided is correct simply because it looks properly formatted
Hint
If your discipline has a maintained standards register or subscription database (ASTM Compass, NFPA LiNK, a firm-maintained code library), use that as your verification source rather than a general web search, which can surface the same mixed-edition confusion an AI tool does.
- AI tools compress the search phase of code and standards research, quickly narrowing which code families and topic areas likely apply to a described condition -- this is genuinely useful as an orientation step.
- AI-generated code citations, including specific section, table, and figure numbers, are research leads, not compliance determinations -- models mix content across standard editions and do not reliably know jurisdiction-specific amendments.
- Apply a five-point verification protocol before any AI-assisted code research enters a compliance document: confirm edition, confirm jurisdictional amendments, locate the actual section or table, confirm the exact value, and confirm applicability to your specific condition.
- Treat an AI-generated citation you cannot locate in the current standard text as fabricated until proven otherwise -- this is a known, common failure mode, not a rare exception.
- Prompt AI code research tools for general orientation rather than specific section numbers where possible -- this reduces the tool's tendency to state a fabricated citation with unwarranted confidence.