Not all code questions are equal, and the single most useful judgment a builder can make about AI is knowing which kind of code question they are holding. Some questions are safe to let AI draft an answer to, some are safe to let AI summarize but never decide, and some must never receive an authoritative AI answer at all because interpreting them is reserved for a licensed professional who stakes their seal on it. Confuse these three categories and you either waste AI's genuine usefulness or, far worse, let a machine make an interpretation that belongs to a stamping engineer. This lesson gives you the three-tier system for codes, OSHA, and sustainability standards, so that every time AI touches a regulatory question you instantly know whether it may draft, may summarize, or must keep its hands off.

Why Codes Need Three Tiers, Not One Rule

The earlier lessons gave you one discipline, verify AI output, and that is correct but too coarse for codes, because regulatory questions span a huge range of stakes and a huge range of who is allowed to answer them. "Draft me a toolbox talk about ladder safety" and "is this connection adequate under the seismic provisions" are both code-adjacent questions, but treating them with the same posture is absurd: the first is a communication task AI does well, the second is an engineering interpretation that requires a licensed structural engineer's stamp and judgment. A single rule cannot capture that range. You need a triage system that sorts a regulatory question into the right posture in seconds, and that system has three tiers.

The tiers are defined by what the question asks the AI to do, not by the topic. The same body of code, IBC, OSHA 1926, IECC, contains questions that fall into all three tiers depending on what is being asked. The skill is not memorizing which standards are safe, because none are uniformly safe; it is recognizing, for the specific question in front of you, whether you are asking AI to draft communication, to summarize for orientation, or to make an authoritative interpretation, because that determines everything about how you can use the answer. So we define the three tiers precisely, then practice sorting real questions into them.

Tier One: AI May Draft

The first tier is questions where AI produces a communication or document and a human verifies the facts, the same language-work posture from the strengths lesson. "Draft a toolbox talk on fall protection for tomorrow's steel erection." "Write a pre-task plan for this confined-space entry." "Turn these OSHA requirements into a plain-language safety briefing in English and Spanish at a sixth-grade reading level." In each, AI is doing what it does best, turning regulatory substance into clear communication, and the human's job is to verify that the substance is correct, the right standard, the real requirement, before it reaches a crew.

The key feature that puts a question in tier one is that the authoritative content already exists and AI is only rendering it; nobody is asking the AI to decide what the rule is, only to communicate a rule a human has confirmed. The verification is the standard language-work gate, check every cited requirement against the actual OSHA standard or code section, with the safety-specific intensity that a pre-task plan or toolbox talk reaches a crew and a wrong requirement is a safety failure. But the posture is fundamentally permissive: this is work AI should do, because it shrinks the time to produce safety and compliance communications dramatically, and the verification is bounded and clear. Tier one is where AI earns its keep on the regulatory side, and recognizing a question as tier one means you can confidently delegate the drafting.

Tier one questions ask AI to communicate a rule a human has confirmed, not to decide the rule. That is the language-work sweet spot: delegate the drafting, verify the substance, and recover real hours on safety and compliance communication.

Tier Two: AI May Summarize for Orientation

The second tier is questions where AI summarizes a body of regulation to orient a human, who then reads the source and makes any actual determination. "Summarize the egress requirements in IBC Chapter 10 so I know what to focus on." "What are the main IECC envelope requirements I should check on this project." "Give me an overview of the OSHA fall-protection requirements that apply to this scope." Here AI is producing a map, not a verdict, pointing a human at what to read, and the human does the reading and the determining.

The crucial discipline in tier two is that the summary is a starting point and never the answer, because this is exactly where the hallucination and conflation failures from earlier lessons live. An AI summary of IBC Chapter 10 may cite sections that do not exist; an AI summary of OSHA may blend 1910 and 1926; an AI overview of IECC may state a requirement from the wrong edition. So the summary's job is purely to orient, to tell you roughly where in the code to look and what topics matter, and then you go to the published code and read the actual provisions yourself, treating nothing in the summary as authoritative. Tier two is truly useful, because orienting yourself in a thousand pages of code is real work AI accelerates, but it carries a sharp boundary: the summary directs your reading and your reading determines the answer, never the reverse. The danger of tier two is the temptation to let the summary become the determination, which collapses it into the forbidden tier three.

Tier Three: AI Must Never Authoritatively Interpret

The third tier is the bright line: questions whose answer is a code interpretation reserved for a licensed professional, where no AI answer carries any authority and acting on one is professionally and legally indefensible. "Is this connection adequate under ASCE 7-22." "Does this egress design comply, sign off on it." "Is this beam sized correctly." "Interpret whether this condition meets the code for purposes of my stamp." These are not communication or orientation; they are professional determinations that the licensing system specifically assigns to a human who takes responsibility, applies judgment to the specific conditions, and stamps the result, accepting liability for it.

The reason tier three is absolute is that the interpretation is the professional act itself, not a fact to be verified. When a structural engineer determines a connection is adequate, they are not retrieving a fact AI could check; they are exercising licensed judgment about a specific situation, considering things the code requires a professional to weigh, and putting their seal and their liability behind the conclusion. An AI cannot hold a license, cannot exercise that judgment with accountability, and cannot stamp, so an AI interpretation in tier three is not a draft to verify, it is a usurpation of an act the law reserves for a person. The danger here is not that the AI's answer might be wrong, though it might; the danger is that the answer has no legitimate authority no matter how right it sounds, and a professional who leans on it has substituted a machine's output for the judgment their license exists to provide. AI can help the professional orient and communicate around a tier-three question, summarize the relevant provisions, draft the narrative once the determination is made, but the determination itself stays entirely with the licensed human, full stop.

The Tier Is Set by Use, Not by the Question's Words

The most important and most counterintuitive part of the system is that the same question can live in different tiers depending on what you are going to do with the answer, which means you, not the question, set the tier. "What is the required egress width here" is innocent-sounding, but it is tier two when you are orienting yourself before reading the code, and tier three the moment that answer becomes the basis of a stamped compliance determination on a permit set. The words did not change; your use of the answer did, and your use is what determines whether an AI answer is a harmless orientation or an illegitimate substitute for a professional's judgment.

This is why the triage question has to be asked about your intent, not just the topic. Before you accept any AI answer to a regulatory question, you ask what you are going to do with it: communicate it (tier one), use it to direct your own reading (tier two), or rely on it as the determination (tier three, forbidden). The discipline lives in being honest about that intent, because the dangerous move is the quiet slide where you start in tier two, "let me just get oriented," and then, under time pressure, let the orientation become the determination without ever consciously deciding to. That slide is how an AI summary ends up being the de facto basis of a compliance call that should have required a licensed interpretation, and nobody decided to cross the line; they drifted across it. Naming your intent up front, and re-checking it if the stakes of the answer rise, is what keeps a tier-two convenience from silently becoming a tier-three violation.

The OSHA Trap That Sits Across All Three Tiers

OSHA deserves a special note because it is where the tier system gets tested hardest, and where a specific AI failure recurs that you have to watch for at every tier. The trap is the conflation of 29 CFR 1910, which governs general industry, with 29 CFR 1926, which governs construction. These are different bodies of regulation with real differences in their requirements, and a general AI model, having read far more about general industry than about construction, will routinely blend them, citing a 1910 provision as if it governed your jobsite or describing a requirement that is actually different under 1926.

This matters across all three tiers in different ways. In tier one, if you ask AI to draft a pre-task plan and it builds the plan on a 1910 requirement instead of the governing 1926 one, the crew gets the wrong rule, so the substance verification must specifically confirm the construction standard, not just that a standard was cited. In tier two, an AI summary of "the OSHA fall-protection requirements" may silently mix the two, so your orientation is pointed at the wrong provisions, and you have to know to read 1926 specifically. And tier three is unaffected only because you were never going to let AI interpret anyway. The lesson is that the tier system tells you what posture to take, and the 1910-versus-1926 trap reminds you that even within the permitted tiers, the substance verification has to be done by someone who knows that construction lives in 1926, because the model does not reliably know it and will hand you general-industry rules with total confidence. Naming this trap is half of catching it.

Sustainability: The Moving Target Within the Tiers

Sustainability standards add their own wrinkle to the tier system because they change, and AI's training cutoff makes it especially unreliable on what is current. LEED v5, released in 2025 and becoming the sole registration option in mid-2026, carries different requirements than prior versions, and a model whose training leaned on older LEED material will describe the wrong version's requirements with confidence. The same applies to evolving energy codes, WELL, and state-specific requirements like California's Title 24, all of which version and update on cycles the model's frozen knowledge cannot track.

Within the tiers, this means sustainability questions carry an extra currency check on top of the normal posture. A tier-one task, drafting a plain-language summary of the project's sustainability commitments for the team, must verify against the actual mandated version, because the AI may render last version's requirements. A tier-two task, orienting yourself in the rating system, must confirm you are reading the current version's provisions, not the one the model defaulted to. And the determination of whether a design actually achieves a credit, which is often a specialized professional judgment tied to the specific rating system and sometimes to a stamp on the energy compliance, belongs in tier three. The pattern is that sustainability does not change the tiers; it adds a version-currency verification to whichever tier the question falls in, because the moving target is exactly the kind of present-state-of-the-world fact the model's frozen training handles worst. When sustainability is involved, the question after "which tier" is always "which version actually applies, confirmed against the current standard."

The Applied Problem: Sort Ten Code Questions Into the Three Tiers

Here is the exercise that builds the triage reflex. Take ten real regulatory questions from your actual work, a mix of safety, building code, and sustainability, and sort each into one of the three tiers, then defend the call. Do not pick easy examples; pick the messy ones from your real projects, because the messy ones are where the judgment lives.

Work them deliberately. "Draft the toolbox talk for tomorrow's crane pick" is tier one, AI may draft, because the requirements exist and AI is rendering them, verify against OSHA 1926 Subpart before the crew hears it. "Summarize the LEED v5 requirements the owner mandated so I know what to track" is tier two, AI may summarize to orient, then you confirm each requirement against the actual rating system. "Does this guardrail detail meet the code, sign off for the permit set" is tier three, no AI authority, because that is a determination for the licensed professional who stamps it. For each of your ten, write which tier and why, and pay special attention to the questions that feel like they could be in two tiers, because the discipline of deciding the boundary is the entire skill. The question "what does the code require for this egress width" is tier two if you are orienting yourself before reading the code, and tier three the instant the answer is going to be the basis of a stamped compliance determination; the same words land in different tiers depending on what you will do with the answer.

The deliverable is your ten questions sorted and defended, and the lasting product is the reflex: the instant a regulatory question arises, you know to ask "am I asking AI to draft, to orient me, or to interpret," and that question routes you to the right posture before you ever read the AI's answer. This triage is what keeps AI productive on the enormous volume of tier-one and tier-two regulatory work while keeping it rigorously out of the tier-three determinations that belong to a stamp, and it is the foundation for every code-checking and safety-planning lesson in the next level, all of which depend on you knowing, for each question, which tier you are in.

Key Takeaways

• Regulatory questions need three tiers, not one rule, because they span a huge range of stakes and of who is allowed to answer. The tiers are defined by what the question asks AI to do, not by the topic; the same standard contains questions in all three tiers.
• Tier one, AI may draft: turning confirmed regulatory substance into communication (toolbox talks, pre-task plans, plain-language briefings). The authoritative content exists and AI only renders it; verify the substance against the actual standard, with safety intensity. Delegate the drafting confidently.
• Tier two, AI may summarize to orient: mapping a body of regulation so you know where to read (IBC Chapter 10 overview, IECC envelope topics, OSHA scope). The summary directs your reading and your reading determines the answer, never the reverse, because this is where hallucination and 1910-versus-1926 conflation live.
• Tier three, AI must never interpret: code interpretations reserved for a licensed professional (is this connection adequate, does this egress comply for the stamp). The interpretation is the professional act itself, not a fact to verify; an AI answer has no legitimate authority no matter how right it sounds.
• The same words can land in different tiers depending on what you will do with the answer. "What does the code require for this egress width" is tier two when orienting and tier three the instant it becomes the basis of a stamped determination.
• The artifact: sort ten real regulatory questions into the three tiers and defend each call, paying special attention to the boundary cases. The reflex, asking "draft, orient, or interpret" before reading any AI answer, is the lasting product.
• Triage keeps AI productive on the large volume of tier-one and tier-two regulatory work while keeping it rigorously out of the tier-three determinations that belong to a stamp and a licensed human's accountability.