How Can Architecture Firms Use AI for Code Review Without Liability Risk?

Most architecture firm partners I talk to spend an unreasonable share of their week on code research. The egress calculations on a tenant improvement nobody is paying full design fees for. The historic preservation review that requires reading three different overlay districts. The energy code path comparison for a project where the client wants the cheaper option but the AHJ has not yet adopted the latest IECC edition. The hours add up, and they fall on the firm's most experienced architects because junior staff cannot be trusted to spot the edge cases that get the firm in trouble at plan review.

At a 15-person firm, this works out to 8 to 15 hours per week of senior-architect time spent on code research that, in theory, any qualified architect with three hours of focused reading could handle. In practice, the senior architects do it because they have the AHJ knowledge to know when the published code does not match what the local building official will accept.

AI changes the math on this if (and only if) the workflow is set up correctly. AI handles the published-code research layer well. It cuts a four-hour code analysis assignment to 30 minutes. It runs comparison checks across code editions in seconds. It produces first-pass compliance memos that read like a senior associate did the work. What AI does not handle is AHJ-specific interpretation, relationship-driven flexibility with a specific building official, or edge cases where the published code is silent and local interpretation is what matters. That part stays with the licensed architect.

This guide walks through the architect-in-the-loop workflow that captures the speed without putting state licensure or professional liability coverage at risk. It covers the categories of code work where AI is genuinely helpful, the categories where AI must not produce the final answer, and the insurance question most firms have not yet asked their broker.

Why this matters for architecture firms specifically

Architecture firms are caught in a quiet margin squeeze. Fees have not kept up with the cost of running a firm. Code complexity has increased every cycle for the last 20 years. Energy code, accessibility code, building code, and zoning code now intersect in ways that make routine projects require more research time than 10 years ago. Junior staff cannot reliably do the research without senior review, which means senior architects do more code work and less of the design and AHJ-relationship work that distinguishes the firm.

The firms that figure out AI for code research reclaim 6 to 12 hours per week of senior-architect time. Some reinvest that into more projects. Some into deeper design exploration. Some into business development. All three beat reading IBC sections aloud at 9pm.

The firms that do not keep losing the speed competition. A client comparing two firms on project turnaround picks the one that completes schematic design and code analysis in less calendar time. AI lets a 15-person firm move at the speed of a 50-person firm.

What an AI-augmented code review system actually does

An AI-augmented code review system is a workflow that uses a Business or Enterprise tier AI assistant (Claude or ChatGPT) on top of the firm's design and project management stack (Revit, AutoCAD, BIM 360, Bluebeam, Deltek Vision, ProjectWise) to handle three tasks: drafting first-pass code analysis memos, comparing code provisions across editions and jurisdictions, and producing compliance summaries the licensed architect reviews and finalizes.

Three things make this different from generic AI use:

• It runs inside a clear regulatory boundary. The workflow defines what AI can produce (first-pass research, comparison summaries, structural framing of compliance arguments) and what it cannot (the final interpretation, the AHJ-specific judgment, the stamped deliverable). The licensed architect always owns the final answer.
• It uses the firm's accumulated AHJ knowledge as context. The firm's prior AHJ correspondence, plan review comments, and successful permit submissions become the training context for AI on similar future projects.
• It produces a defensible audit trail. The firm can show, on demand, which code research AI drafted and which the licensed architect verified or wrote from scratch.

Think of it as a junior architect who has read every code section the firm has ever cited, never sleeps, and produces first-pass code analysis in 15 minutes that used to take three hours.

Before you start

You need:

• A Business or Enterprise tier AI account with a Data Processing Addendum signed by the firm. Mandatory before any project information enters the workflow.
• Read access to the firm's existing project archives (BIM 360, ProjectWise, or wherever past project documentation lives). The historical AHJ correspondence and prior code analyses are the most valuable training context for new code research.
• Updated client agreements with appropriate AI use disclosure language. If your engagement letters and AIA contract amendments do not address AI use yet, do that first. Most state architecture associations have model language available.
• A real project to test the workflow on. Pick something with moderate code complexity (a tenant improvement in a multi-occupancy building, a small commercial project in a jurisdiction with overlay zoning, a renovation that triggers existing-building code provisions). Avoid first-time use on anything high-stakes.

One thing to settle before any project document enters AI: the licensure and liability question. We have a dedicated section on this below. It is non-negotiable. The firms that skip these steps are one bad code citation away from a stop-work order, a liability dispute, or a state board complaint.

Task 1: Build the firm's code-research scaffolding and AHJ knowledge base

The failure pattern: every architecture firm has institutional knowledge about the AHJs it works in, but almost no firms have it written down. The senior architects know that one reviewer in the local building department always asks for a separate egress diagram showing the path from each space to two exits. Mid-level architects do not know this until they get the plan review comment back. The firm reinvents the wheel on every project.

What to build first, before any project-specific code research:

Here are 8 of the firm's prior code analyses across the AHJs we work in most often. Read them and extract the patterns. Which code editions does each AHJ typically work under, which local amendments matter most for the project types we do, which provisions does each AHJ interpret unusually compared to the published code, what is the firm's standard format for a code analysis memo, and what kinds of questions does each AHJ typically ask in plan review. Output a structured knowledge base organized by AHJ.

This is a one-time three or four hour investment by the senior architect leading QA. The output becomes the firm's anchor knowledge base. New AHJ entries get added as the firm wins work in new jurisdictions. The knowledge base also becomes onboarding material for new project managers. For firms with design specialization (healthcare, K-12, multifamily), run a second pass to capture the specialty-specific code knowledge as a separate anchor.

Task 2: Generate the first-pass code analysis memo

Most projects start with a code analysis memo that establishes the basic envelope: occupancy classification, construction type, allowable height and area, applicable code edition, fire-resistance ratings, egress requirements, accessibility scope. The senior architect typically writes this from scratch, pulling sections from the IBC and local amendments, taking three to five hours per project.

What to ask AI for instead:

Generate a first-pass code analysis memo for this project: a 12,000 square foot multi-tenant retail and restaurant building in [City, State], submitted under the locally adopted IBC 2021 with state amendments. Tenants include three retail spaces (Group M), one full-service restaurant with a 95-occupant dining area (Group A-2), and one quick-service restaurant (Group A-2). Use the firm's standard memo format. Identify occupancy classifications and mixed-use considerations, construction type given the footprint and adjacent properties (Type IIA per our preliminary analysis), allowable height and area calculations under IBC Section 506, mixed-occupancy separation requirements, fire-resistance ratings for the structural frame and exterior walls, egress capacity and travel distance limits, and accessibility scope under the IBC and ADA. Flag any provision where state amendments diverge from base IBC. End with three clarifying questions the firm should resolve with the AHJ before schematic design progresses.

The prompt forces specificity. Generic prompts produce generic outputs. Specific prompts (with actual square footage, occupancy details, jurisdiction, and code edition) produce outputs that read like a senior associate did the first pass. The licensed architect reviews, verifies citations against the actual code text, adds AHJ-specific knowledge, and finalizes. A four-hour memo becomes a 90-minute memo. For renovation projects triggering IEBC provisions, add: "This is a renovation. Identify which compliance method (Prescriptive, Work Area, or Performance) is appropriate and walk through the requirements."

Task 3: Run comparative code research across editions and jurisdictions

Projects often span multiple jurisdictions or sit at the boundary of code editions. A regional client with three buildings across three cities, each with a different adopted code edition. A renovation where the existing building was permitted under an older code and the renovation triggers current-code requirements for some sections but not others. A project type the firm has not done in the AHJ before.

What to ask AI for instead:

The firm has done a similar project type (medical office building, 15,000 square feet, Group B with limited Group I-2) in [Home City, Home State] under IBC 2018 with state amendments. We are now doing a similar project in [Target City, Target State] under IBC 2021 with different state amendments. Compare the two regulatory environments. Identify the provisions that have changed materially between IBC 2018 and 2021 for this occupancy and project type. Identify the state amendment differences. Flag any provision where the firm's standard approach in Home State will not work in Target State. Output a comparison memo.

The prompt does what senior architects do mentally but rarely write down: a structured comparison of regulatory environments. AI is good at this when given the actual jurisdictions and editions. The licensed architect reviews, verifies, and adds AHJ-specific knowledge AI cannot have. For multi-jurisdiction portfolio clients (national retailers, restaurant chains, healthcare systems), this workflow becomes a recurring engine. The firm builds one comparison memo per jurisdiction the client operates in. New projects pull from existing memos, AI updates them for code edition changes, the licensed architect reviews.

Task 4: Draft the energy code compliance pathway analysis

Energy code analysis is one of the most time-consuming research tasks in modern architectural practice. The IECC, ASHRAE 90.1, state-specific energy codes, and voluntary above-code programs (LEED, ENERGY STAR, local stretch codes) intersect in ways that take a senior architect or specialist consultant several hours to untangle.

What to ask AI for instead, before the senior architect or energy consultant does the deep dive:

Generate a first-pass energy code compliance pathway analysis for this project: 25,000 square feet of office space, Climate Zone 4A, jurisdiction adopts IECC 2021 with state amendments, targeting LEED Silver. Compare the three compliance pathways available under IECC 2021 (Prescriptive, Total Building Performance, and ASHRAE 90.1 Performance). For each, summarize the scope of analysis required, typical advantages and trade-offs, and the design flexibility for the project team. Identify the pathway most likely to be efficient for this type and climate zone. End with the data the design team needs to gather before the energy consultant can run a final compliance model.

The constraint that matters: "first-pass." AI produces the framing memo that the energy consultant uses to scope the actual compliance work. AI does not produce the compliance model itself. Firms that skip specialist verification are firms that fail final code review on energy compliance. For projects pursuing voluntary certification, the same prompt structure compares IECC requirements against LEED or WELL thresholds.

Task 5: Build the accessibility code compliance summary

Accessibility code review is the second most time-consuming research category and one of the highest-liability areas in architectural practice. ADA, ANSI A117.1, IBC Chapter 11, state accessibility codes, and project-specific federal program requirements (Fair Housing Act for multifamily, Section 504 for federally funded buildings) intersect in ways that produce real liability exposure when missed.

What to ask AI for instead:

Generate an accessibility compliance summary for this project: a 6-story multifamily residential building, 110 dwelling units, in a jurisdiction adopting IBC 2021 with state amendments and ADA 2010 standards. Identify the applicable accessibility standards for residential occupancies, the scope of accessible, Type A, and Type B unit requirements under IBC Chapter 11 and the Fair Housing Act, the common-use area scope, and any state-specific provisions that exceed the federal floor. Flag the provisions most often cited in plan review comments for similar projects. End with three or four questions the design team should resolve before schematic design progresses.

The constraint that matters: accessibility code is unusually fact-specific to the project type. AI is good at framing and standard provisions. AI is bad at project-specific edge cases. The licensed architect or accessibility specialist owns the final review. For commercial projects with mixed-occupancy accessibility scope, the prompt extends: "Identify the accessibility requirements for each occupancy in the building, where they differ, and where the design must accommodate both."

Task 6: Generate the code analysis cover memo for AHJ pre-application meetings

Most AHJs offer pre-application meetings. The architecture firm walks the AHJ through design intent, proposed code analysis, and open questions before formal permit submission. These meetings are high-value when prepared well and waste of everyone's time when prepared badly.

What to ask AI for instead, in the days leading up to a pre-app meeting:

Generate a pre-application meeting cover memo for this project: [project description]. Summarize the project scope in two paragraphs, list the firm's preliminary code analysis findings (occupancy, construction type, allowable height and area, key code questions), name the three to five specific questions the firm wants to resolve with the AHJ, and propose a schedule for design progression based on the answers we expect. Voice: professional, prepared, respectful of the AHJ's time. Length: under 800 words. Include a one-page summary at the top for the building official's reference.

The pre-app cover memo written this way reads as a firm that has done its homework. The AHJ reviewer walks into the meeting with the right context. The questions are sharp. The meeting takes 45 minutes instead of 90, and the firm leaves with the answers it needed. That difference, multiplied across a year of pre-app meetings, is meaningful billable-hour recovery.

The architecture-specific prompts that actually work

After watching architecture firms use AI on code work, the difference between AI output that is genuinely useful and AI output that puts the firm at risk comes down to four prompt moves.

Specify the AHJ and the code edition. "Code analysis for a retail building" produces generic output. "Code analysis for a 12,000 square foot retail building in [City, State] under the locally adopted IBC 2021 with state amendments" produces output the firm can actually use. The jurisdiction and edition are non-negotiable inputs.

Specify the project facts. Square footage, occupancy classifications, construction type, climate zone, target voluntary programs. The more specific the facts, the more useful the output. Vague project descriptions produce vague code analyses.

Specify the constraint that matters. Most code research has one constraint that makes or breaks the analysis. "Verify all citations against the actual code text" is one. "Flag every provision where state amendments diverge from base IBC" is another. "Identify the cheapest compliant pathway" is a third. Pick the constraint that, if AI got it wrong, the firm would have to redo the work.

Specify what AI does and what the architect does. Explicit role separation in every prompt. "AI produces first-pass research and structured framing. The licensed architect reviews, verifies citations, adds AHJ-specific knowledge, and finalizes." That single instruction keeps the workflow inside the licensure boundary.

The state licensure and liability non-negotiables

This section is short because the rule is simple, but it is the most important section in this guide.

Do not put any of the following into the consumer tier of any AI tool:

• Project drawings, BIM models, or any CAD file content
• Client confidential information covered by the engagement agreement
• AHJ correspondence, plan review comments, or permit submission documents tied to a named project
• Internal firm working papers on a specific project's code strategy
• Personally identifiable information about clients, contractors, or building occupants
• Anything subject to a confidentiality or NDA the firm has signed

The practical workflow: build the firm's AHJ knowledge base and code-research scaffolding on the consumer tier with anonymized examples, then move to the Business or Enterprise tier the moment any project-specific information enters the workflow. The Business tier with a Data Processing Addendum is the only defensible foundation.

State architectural licensure requires the licensed professional to take responsibility for the work, regardless of how the underlying research was conducted. AI does not stamp drawings. AI does not sign code analyses. The licensed architect does. Document the firm's AI use in the QA manual. Show the QA process in client agreements and AIA contract amendments. Be ready if a state board reviews the firm's practices.

Professional liability and E&O insurance is the part most firms have not addressed. Most policies written before 2024 are silent on AI. Carriers issuing new policies are adding riders or exclusions specifically for AI involvement in code interpretation. Call the firm's broker. Ask: does the policy cover work where AI assisted in code research, and does it cover errors where AI hallucinated a code citation. If uncertain, get the rider added. The cost is rounding error compared to a single missed code requirement that becomes a construction-defect claim.

AHJ-specific interpretation cannot be delegated to AI. The senior architect's local knowledge is the most valuable part of the firm. AI augments it. AI does not substitute for it. If your firm has signed an Anthropic Business or OpenAI Enterprise agreement with a Data Processing Addendum, the rules on data flow are different. Ask your firm's general counsel or operations director what is covered. Do not assume.

When NOT to use AI for code work

AI is a generalist tool. It is not the right answer for every code research task.

Skip it for:

• Anything the licensed architect will stamp without a real review. AI produces first-pass research. The licensed architect reviews every citation and conclusion before the deliverable goes anywhere. Firms that stamp AI output without verification are one bad citation away from a state board complaint.
• AHJ-specific interpretations that depend on personal knowledge. AI does not know that a specific reviewer always asks for a particular detail. The senior architect knows.
• Final energy code compliance modeling and accessibility plan reviews. AI is good for the framing. Specialists own the final analysis.
• Complex zoning interpretations and historic preservation cases. Zoning code is highly local and often interpreted by the planning department in ways that diverge from the published text. Historic preservation has a layer of judgment AI cannot replicate.

A simple rule: AI is an unfair advantage on the 60% of code research that is published, structured, and verifiable. Trust the licensed architect for the 40% that depends on AHJ relationships, local interpretation, and professional judgment.

The quick-start template

Here is the prompt scaffold that works across most architecture firm code research use cases. Copy it, fill in the brackets, paste into Claude or ChatGPT at the Business tier.

Generate a [type of code analysis: building code, energy code, accessibility, zoning, comparative jurisdiction] for this project: [project descriptor with square footage, occupancy, construction type as known].

Jurisdiction: [City, State, AHJ name].

Adopted code: [IBC edition, IECC edition, ASHRAE edition, ADA standards, with state amendments named].

Project type and use: [tenant improvement, ground-up new construction, renovation, change of occupancy].

Specific code questions: [the named questions the firm needs to answer].

Constraint that matters most: [verify all citations, flag state amendment divergences, identify the cheapest compliant pathway].

Format: [first-pass memo, comparison memo, pre-application cover memo].

Length: [under 1,000 words, under 1,500 words].

Role separation: AI produces first-pass research and structured framing. The licensed architect reviews, verifies citations against the actual code text, adds AHJ-specific knowledge, and finalizes.

That is the whole pattern. For most code research, this scaffold is enough.

For recurring use, store the scaffold in the firm's project management tool (Deltek Vision, ProjectWise, or Procore) as a code analysis template. Each new project gets a new instance, with the firm's standard QA review steps built in by default.

Bigger wins beyond the single project

Once the firm has the basic code research workflow running, the next layer of value shows up in places that are not single projects.

A firm-wide AHJ knowledge base that compounds. Every project adds to the AI training context for the next project in the same AHJ. Set up a structured database in Notion or ProjectWise with one entry per AHJ: adopted codes, common plan review comments, named amendments, reviewer-specific quirks. Within a year, the firm has institutional knowledge that survives senior architects retiring.

A faster project intake and fee-proposal cycle. AI-assisted code research means the firm produces credible fee proposals faster than competitors. AI gives the firm a quick read on the code load, the proposal goes out faster, and the firm wins more projects on responsiveness.

A QA process that surfaces missed citations earlier. Use AI as a second-pass reviewer on stamped deliverables: "Identify any citations that may not match the actual code text, any provisions that may have been missed, and any inconsistencies between sections." AI catches small errors humans miss after staring at the document too long. The licensed architect verifies the AI's flags.

A specialist consultant relationship that gets sharper. Energy consultants and accessibility specialists are expensive. Firms that come to those specialists with AI-augmented first-pass analyses get more efficient consultant hours. The specialist verifies and refines instead of starting from zero.

The professional services AI consulting connection

This is one tool in one category. The bigger AI question for architecture firms is structural: the work that used to require a four-architect team for two weeks now requires one senior architect and AI for two days. Firms that adapt their staffing models, pricing structures, and senior-team focus to that reality will compound margin and project capacity. Firms that do not will keep losing senior architects to firms that figured it out.

If your firm is wrestling with the bigger question (which workflows AI reshapes, what the new staffing model looks like, how senior architects and project managers should spend their time), the AI Consulting in Professional Services page covers the full scope: where AI fits in agency, consulting, accounting, and architecture work, the common adoption failure modes, and what an engagement looks like when it works.

For the individual firm reading this: pick the next code analysis on your desk. Run the workflow on it this week. The hours saved are the case for everything else.

Closing

The goal is not to ship more code analyses. It is for senior architects to spend their time on design judgment, AHJ relationships, and client conversations that distinguish the firm. AI for code research is the cleanest entry point: the workflow is contained, value shows up on the next project, and the compliance frame is manageable if you address licensure and professional liability upfront.

Pick the next code analysis. Run the workflow on it this week. Compare the senior-architect time before and after. The case for rolling AI into the rest of the firm's research and analysis work writes itself.

If you want to talk about how AI fits into your firm at the program level, the AI Consulting in Professional Services page lays out the full picture and how an engagement works.