AI will not replace your design judgment, your stamp, or your responsibility for a building that has to stand up and let people out safely. It will absolutely replace the blank first hour of concept exploration, the afternoon spent hunting through a code chapter, and the tedium of a first-draft project narrative or spec section. This handbook is the practical middle ground: where AI genuinely helps architectural work, the three rules that keep it from becoming a liability, and five workflows you can use this week — no technical background required.
~18 min readno coding5 workflow recipes1 judgment exercise
This is a guide to working with AI tools, written by an engineer — it is not architectural, engineering, or legal advice, and it is not a substitute for the adopted codes, your professional standards, your jurisdiction's requirements, or the qualified professionals of record. When they conflict with anything here, they win.
01
Where AI actually helps — and where it doesn't
The single most useful mental model: treat AI like a fast, tireless junior designer and researcher with no license, no knowledge of your jurisdiction's adopted code, no engineering, and occasional confident lying. You'd happily hand that assistant a concept study or a first-draft narrative. You would never stamp their code interpretation, and you'd never let their rendering stand in for an engineered, buildable design.
Task
Verdict
Why
Concept ideation, mood, form-finding studies
Excellent
Fast divergent exploration; you curate and develop the ideas worth keeping
Models state wrong/outdated/wrong-jurisdiction requirements — see Rule 2
Anything stamped, or a rendering shown as buildable — unverified
Never
Your stamp, your liability — see Rule 1
02
The three non-negotiable rules
Rule 1 — You stamp it, you own it.
Professional responsibility does not delegate to software. Whatever the tool produced — a spec section, a code interpretation, a detail, a narrative — the moment it goes into a stamped set or a client deliverable it is your professional work, carrying your liability for life-safety, code compliance, and buildability. Practically, AI output gets the same scrutiny you'd give a green intern's work: read every line, verify every requirement, redraw what's wrong. The time you save comes from skipping the blank page and the legwork, never from skipping the professional review your seal represents.
Rule 2 — Every code, zoning, and structural claim gets verified. Every one.
Language models generate plausible text, not the adopted code. They will state an egress width, a fire rating, a setback, or a live load — confidently and specifically — that is wrong, from an outdated edition, or from a different jurisdiction than yours. This is life-safety with legal liability attached to your seal. A code requirement from an AI is a pointer to look it up, never a citable fact: verify it in the actual adopted code and local amendments for the specific jurisdiction, and route structural and life-safety questions to the qualified engineer of record.
Rule 3 — Confidential client and site information stays protected.
Free and consumer AI products may retain what you type and use it to train future models. Unbuilt design concepts, client business plans, site security details, unannounced projects, and NDA-covered information can leak once pasted. The discipline: use enterprise tools with contractual no-training and retention commitments, and abstract identifying details even then ("a mixed-use project on a tight urban infill site" — not the address and client). A leaked concept or a breached NDA is a real professional and business harm, not a technicality.
03
Five workflows you can use this week
Each recipe: what to give the AI, what to ask, and what you must verify by hand. The prompts are starting points — adjust the specifics and keep your confidentiality discipline.
1 · Concept ideation and design directions
Give program, site character, constraints, and the feeling you're after:
I'm designing [a small community library] on [a sloping wooded corner site]. Program: [bulleted]. Constraints: [budget/height/context]. Generate 5 distinct conceptual approaches — for each: the organizing idea, how it engages the slope and trees, the primary material palette, and one risk or tension it introduces. Push for range, not safe variations.
You verify: nothing to "verify" here — this is divergence, not fact. But treat every idea as raw material: buildability, code, and site reality get applied by you as you develop the ones worth keeping.
2 · First-draft narrative or design statement
Give the design's actual ideas, audience, and length:
Draft a project narrative for [a client presentation]. The design's core ideas: [bulleted, in your words]. Audience: a non-architect client board. Tone: confident and warm, concrete not flowery. ~300 words. Explain the "why" behind each major move in plain language, ending with how it serves their goals.
You verify: that every claim matches the actual design (the model will invent appealing rationale you didn't intend), and rewrite it in your voice — a narrative that oversells or misdescribes the design is worse than none.
3 · Research starting point for code/materials — never endpoint
Use AI for the map, then confirm in the adopted code and product data:
For [a 3-story mixed-use building] under a code based on the [IBC], what topics and chapters typically govern [means of egress and occupancy separation]? For each: what questions I need to answer and what commonly gets missed. Do not give me specific numbers, widths, ratings, or citations — just the landscape of what to check and where.
You verify: everything — this is a checklist of what to look up, not the answers. The "no specific numbers or citations" instruction steers the model away from its most dangerous failure mode (Rule 2) toward what it's genuinely useful for.
4 · Early rendering / visualization for conversation
Describe the space, light, materials, and mood for an image tool:
Generate a concept rendering: [interior of the library reading room], [double-height space with clerestory light], warm timber and pale plaster, late-afternoon sun, a few people reading, calm and generous mood. Architectural photography style, wide angle. Concept exploration, not a construction document.
You verify: label it clearly as concept imagery to the client — AI renderings show plausible pictures, not buildable geometry (impossible cantilevers, materials that don't exist at scale). Never let it read as an engineered or permittable design.
5 · Draft-set QA prompts and coordination checklists
Ask AI to generate the checklist; you (and the team) run it against the real set:
Generate a coordination QA checklist for a [design development] set on [a mid-size commercial project]: common cross-discipline conflicts, drawings-vs-specs mismatches, and things that frequently get missed between architectural, structural, and MEP. Organize by discipline interface. This is a checklist to run manually, not a review of any drawings.
You verify: the checklist is a memory aid, not a review — the actual coordination and clash-checking is done by the team against the real documents. AI never "reviews" your set for compliance; it just helps you remember what to look for.
04
The judgment exercise: spot the danger
Three scenarios from real practice patterns. Pick what you'd do — the point is calibrating when the rules bite.
1. Sizing an exit corridor, the AI states: "Per IBC, the minimum corridor width for this occupancy is 44 inches." It's specific and confident, and 44" sounds right. Use it?
Rule 2. "44 inches" may be right for some conditions and wrong for yours — occupancy, occupant load, edition, and local amendments all change it, and the model doesn't know your adopted code. Designing egress to an unverified AI number is a life-safety error carrying liability to your stamp. The AI points you to the topic; the adopted code gives the value.
2. A client loves an AI concept rendering and asks "so we can build exactly this?" The image shows a dramatic unsupported cantilever. What do you say?
Rule 1 and honesty. AI renderings are plausible pictures, not engineered geometry — the cantilever may be structurally impossible or unaffordable. Letting a client believe a concept image is buildable sets up a broken promise and a liability. Frame renderings as exploration, always, and manage the gap between image and buildable design explicitly.
3. To get fast AI feedback on a competition concept, it's easiest to paste the full unannounced design and client brief into a free chatbot. Do you?
Rule 3. A consumer tool may retain and train on the concept and brief; deleting the chat doesn't undo retention on the provider's side. Unannounced competition designs and client business plans are exactly the confidential material you must protect. Enterprise tools with contractual no-training terms, and abstract the specifics.
05
Choosing tools: the questions that matter
You don't need to understand the technology to procure it well. You need answers, in writing, to five questions:
Question
Answer you want
Is our data used to train your models?
No, contractually — not "you can opt out somewhere in settings"
How long are prompts and uploads retained?
Defined, short, and deletable on request
Where is data processed and stored?
A stated jurisdiction you can live with
Security posture?
SOC 2 Type II or equivalent, encryption in transit and at rest
Admin controls and audit logs?
Yes — you'll want to know who used it for what
General-purpose enterprise AI (Claude, ChatGPT's business tiers) covers the writing, research-starting, and ideation recipes here. Architecture- and AEC-specific tools add generative design in BIM, code-checking against structured databases, and image-to-model workflows — evaluate those on their own accuracy claims and your firm's data policy. And if your firm has an AI policy, it precedes everything here; if it doesn't, these three rules are a reasonable seed for one.
The picture
The picture: one routing decision for every AI output
The three rules as one routing step. The question is never “is this AI output good?” but “what kind of claim is this?”. Divergent ideas need only your taste. A requirement needs the adopted code. A prompt that names the client needs abstracting before it is sent, not after.
By the numbers
Worked example: auditing one AI-drafted narrative, sentence by sentence
You give the model the library project from Recipe 1 (sloping wooded corner site, double-height reading room under a clerestory) and ask for the ~300-word client narrative from Recipe 2. Here are five sentences from the draft, sorted into the routes in the picture above.
#
Sentence in the draft
Kind of claim
What you do
1
“The reading room rises double height under a band of clerestory light.”
Matches the design
Keep. Tighten into your voice.
2
“Rainwater from the roof feeds a garden along the slope.”
Rationale you never designed
Delete. The client will remember it as a promise.
3
“The main stair meets the required 44-inch width.”
Code claim
Strike the number. A narrative is no place for an unverified requirement (Rule 2).
4
“The timber structure stores carbon for the life of the building.”
Performance claim
Cut, or keep only once your structural and materials choices actually support it.
5
“Set on Elm Street for the Friends of the Library…”
Identifying detail
Fine in the final file. It should never have been in a consumer-tool prompt (Rule 3).
Tally: of five sentences, one survives as written, one needs a check before it can stay, and three go or change. That ratio is normal. The model writes appealing, specific sentences, and specificity is exactly what makes a wrong claim dangerous in front of a client. The draft still saved the blank hour. The audit is the part your stamp pays for, so budget time for it every time.
Trade-offs
Trade-offs: which kind of tool for which work
Read Time as how fast it gets you a usable draft, and Space as the risk or cost it carries.
Case
Time
Space
Why
Free consumer chatbot
Instant, nothing to set up
May retain and train on prompts
Fine for public, non-project questions. Never for unbuilt concepts, briefs or site data.
Enterprise general-purpose AI
Fast once procured
A licence, plus the written answers in the table above
Covers every recipe on this page. Still needs abstraction for the most sensitive work.
Narrower, and their accuracy claims are theirs to prove
Ask which outputs are deterministic and which are generated. Check the generated parts like a junior’s.
No AI
Slowest first draft
No new data or accuracy risk
Still the right choice for anything you cannot abstract or cannot verify.
The verification cost never moves. A requirement from any of these tools still has to be looked up in the adopted code. What the choice of tool changes is the confidentiality risk and how fast the draft arrives.
Watch for
Common mistakes architects make with AI
Verifying only the numbers that look wrong
An odd number gets checked. A familiar one, like the 44 inches in the exercise, slides through. Familiar is not the same as correct for this occupancy, edition and jurisdiction. Verify every requirement, starting with the ones you think you already know.
Checking the model’s edition instead of yours
Confirming that a requirement exists in “the IBC” is not enough. Check it in the edition your jurisdiction adopted, with its local amendments. A real requirement from the wrong edition is still wrong on your sheets.
Abstracting the address but not the project
Removing the client’s name while pasting the full competition brief, a site photo or a unique program still identifies the work. Abstract until a stranger could not tell which project it is.
Labelling the rendering in the meeting, not on the image
Clients forward images. If “concept imagery, not a buildable design” is said aloud but not printed on the picture, the label does not travel with it. Put it on the file.
Treating the QA checklist as a review
Recipe 5 produces a list of things to look for. A finished checklist means nobody has checked the set yet. The coordination review still has to be run by the team against the real drawings and specs.
In the wild
Where this shows up in practice and in interviews
Plan review
Reviewers check your set against the adopted code. An AI-originated wrong requirement that reaches the set comes back as a comment and costs a resubmittal cycle at best, or a field correction at worst. That is why it has to be caught at your desk.
Client presentations
Recipe 2 and Recipe 4 are where AI saves the most time and where the gap between image and building is most visible. Show renderings and narratives as exploration, and keep a record of what was promised.
Firm AI policy and contracts
NDAs, competition rules and client contracts decide what may go into which tool. When a firm writes its first AI policy, the three rules on this page are a usable outline: the stamp, verification and confidentiality.
Job interviews
If you are asked how you use AI, give the routing step, not a tool list: what you hand it, what you verify and where, and what never goes into a prompt. That answer shows judgment, which is what the question is testing.
Frequently asked
Quick answers
Can AI do my code analysis?
No — treat AI as a way to find which topics and chapters to check, never as the source of the requirement. Every value gets verified in the adopted code and local amendments, and structural/life-safety confirmed by the professionals of record. Your stamp carries the liability.
Can I use AI renderings in a client presentation?
Yes, as clearly-labeled concept imagery for exploring feeling and direction. Never present them as buildable or engineered — they're pictures, not geometry.
Does AI-assisted design threaten authorship or IP?
Keep confidential work out of consumer tools (Rule 3), and know your tools' terms on output rights. For client work, your firm's contracts and professional judgment govern authorship, not the tool.
Where should a skeptical architect start?
Recipe 2 — a first-draft project narrative from your own design ideas. Zero code risk, zero site-data risk if abstracted, and it shows the genuine strength (fast, decent writing from your inputs) without touching the failure modes.