How to Use AI in Architecture Design: Concept Options for Small Practices

A small practice keeps the client brief intact by turning it into a numbered register of requirements first, then checking every AI-made massing, layout and image against that register before the client sees it. AI proposes options, the register decides which survive, and the architect chooses and develops them.

That sounds slow, but it moves the checking to the cheap end of the job. A three-person practice that checks each step against twelve numbered requirements finds a failed option in minutes instead of at the client meeting. Take the single-storey house extension worked through below, from meeting notes to a three-option comparison: the checks catch an area that looks fine and is not, a planning limit read the wrong way and a rooflight that exists only in a picture. Every plot size, rule and figure is illustrative.

Follow me on Instagram@sagnikteaches

Turn the meeting into a register the whole job is checked against

After the first client meeting, paste your notes into a chat assistant and ask for a register: one row per requirement with an ID, the requirement, who said it, how it will be tested and its priority. Tell the assistant to write NOT STATED wherever the client gave no number. The notes in this example (illustrative) are deliberately messy:

Connect on LinkedInSagnik Bhattacharya

Meeting 1. Family of four plus dog. Wants one big kitchen-dining-living
space, at least the size of what they saw at a friend's, say 40-something
square metres. Needs a utility with an outside door. Small study,
quiet. Downstairs loo, wheelchair friendly for the grandparents. Doors
to the garden, no step. Keep the big tree. Budget 190k, we reckon
about 3,150 a square metre for this kind of work. Want options in
three weeks.

The register (illustrative) comes back as a table, and then you complete it yourself. Where the client gave no figure, the assistant must say so, and you decide whether to propose one:

Subscribe on YouTube@sagnikteaches

Two rows show the point of the exercise. R3's size is the practice's proposal, not the client's, and the register says so; otherwise it would later be remembered as something the client demanded. R7 turns a budget into an area using your own cost per square metre, which gives each option a number to be tested against. The assistant arranged the words, but the 42, the 9 and the 4.0 came from you.

Then confirm the register in a short call or email, ask the client to reply "agreed" and call the result version 1. That reply is the cheapest piece of design insurance in the job. For a longer treatment of room-by-room briefs, see the tutorial on room data sheets from a client brief.

Pull the planning limits out of the documents, with page references

Planning or zoning documents are long, and the useful rules are numbers buried in paragraphs. Paste only the relevant pages, not the whole document, and ask for each rule with its page and its exact sentence. An invented extract in the style of a local policy shows the method:

Extract (illustrative, page 14): Extensions must not raise total
building coverage above 40% of the plot area. A single-storey rear
extension must not exceed 3.8 m in height and must be set back at
least 6.0 m from the rear boundary and 1.0 m from side boundaries,
unless a lower setback is justified by a site-specific design
statement.

Task: list each numerical rule with page and exact sentence, then
restate each as arithmetic for a plot of 380 m2 with an existing
building footprint of 96 m2. List any exception separately.

A first reply from an assistant (illustrative) can read the coverage rule as a limit on the extension itself, giving "up to 152 m²". That is 40% of the plot, but the policy says total building coverage. The corrected arithmetic is 0.40 × 380 = 152 m² for everything on the plot, less the existing 96 m², which leaves 56 m² for the extension. It is a one-line error that changes the whole brief, and it was easy to spot only because the prompt demanded arithmetic and the exact sentence. The reply should also list the height limit (3.8 m), the rear setback (6.0 m), the side setback (1.0 m) and, separately, the exception: "unless a lower setback is justified by a site-specific design statement". That last line is not a rule you can design to. It is a question for the authority.

Check each extracted number against the document yourself before it enters the register. Then use the list to arrive at any pre-application conversation with precise questions. Many authorities offer one, sometimes for a fee, so check what yours provides. The assistant has no knowledge of recent amendments, appeal decisions or how an officer reads a rule. The tutorial on where to draw the line on planning applications sets out which parts of this work stay with the architect.

Massing options: strategies from the assistant, numbers from the model

Ask the assistant for three different organising strategies and, for each, the register items it puts at risk. Give it the register and ask for strategy names and risks only, not dimensions. For this house, the strategies are a linear extension along the rear wall (A), an L-shape wrapping the garden corner (B), and a courtyard scheme with the utility and study on the side (C). The assistant predicts that B will struggle with the tree and the rear boundary, which is a useful hint and nothing more.

The geometry comes from you. Build each strategy as a simple massing in the modelling tool you already use, or in a concept tool such as Snaptrude, which describes itself as built for early-stage concept design and lists its Pro plan at $50 a month on yearly billing or $60 monthly (its page prints prices in more than one place and they do not all agree, so confirm before buying). Then measure footprint, height, distance to the rear boundary and distance to the tree from the model. Those measurements, not anything the assistant says, fill the numeric cells of the comparison:

• Option A: footprint 52 m², height 3.6 m, rear boundary 6.4 m, tree 4.6 m. Coverage is (96 + 52) ÷ 380 = 38.9%.
• Option B: footprint 58 m², height 3.7 m, rear boundary 5.8 m, tree 3.7 m. Coverage is (96 + 58) ÷ 380 = 40.5%.
• Option C: footprint 55 m², height 3.8 m, rear boundary 6.1 m, tree 4.2 m. Coverage is (96 + 55) ÷ 380 = 39.7%.

Option B already fails three requirements before any layout exists: the tree (3.7 m against 4.0 m), the rear boundary (5.8 m against 6.0 m) and coverage (40.5% against 40%). That is the first brief check, and it takes five minutes. A practice that finds this at the client meeting has lost the meeting's credibility. One that finds it here drops B or reshapes it.

Layout options: use a generator for adjacency ideas, not for sign-off

Floor-plan generators produce plausible layouts quickly, which makes them useful for seeing alternative adjacencies and troublesome for anything with a limit. Maket, for example, says most floor plan options are ready within two to three minutes of submitting requirements. Its free plan gives 30 one-time credits, its Homeowner plan costs $20 a month for 300 credits and its Pro plan costs $100 a month for 3,000. A generated or uploaded floor plan costs 20 credits a floor, and its pricing page says the AI-assisted upload on the Homeowner plan isn't guaranteed to be 100% accurate (Pro uploads are described as human-validated). The page says nothing about building codes. Exports sit on paid plans only. So treat what comes out as a sketch from a colleague who has never seen the site.

The brief check for this step is the assistant's job, within strict rules. Describe each layout in a short paragraph, then ask for a status per requirement with evidence. The rule that matters is in the prompt below: if the description does not say, the answer is UNKNOWN.

For each requirement in the register, write MET, NOT MET, PARTLY or
UNKNOWN. MET and NOT MET need a quote from the description and the
numbers. If the description does not say, write UNKNOWN. Do not
assume. Do not suggest design changes.

OPTION C: Single-storey courtyard scheme. Kitchen-dining-living 42.5 m2
facing the garden through sliding doors to a terrace. Utility 6.0 m2
with a side door. Study 8.6 m2 at the front of the extension, away from
the kitchen. WC 3.1 m2. New floor area 55 m2. Ridge height 3.8 m.

The reply (illustrative) is more useful for what it refuses to say than for what it confirms:

An assistant allowed to assume would mark R5, R6 and R8 as MET, because "sliding doors" suggests level access, a courtyard scheme suggests clearance from a tree and a rear extension suggests the boundary is respected. All three are guesses presented as evidence. With the UNKNOWN rule, the unknown and partly known items become your to-do list: you read them off the model and the survey. For C, they come back as a threshold to be detailed, 4.2 m to the tree, 6.1 m to the boundary, 39.7% coverage and a south-east orientation. For a fuller account of where generators fail a real brief, see AI floor plan generation for architects.

Concept images: keep the picture honest

Images carry more weight in a client meeting than any table, which is why they need their own check. Image generators such as Midjourney or Firefly work from a text prompt, or a reference image you upload, and hold no model of your massing. Render add-ins work differently. Veras, now part of Chaos, says it works with Revit, SketchUp, Rhino, Archicad, Vectorworks, Forma and Allplan, and as a web app, and that with precise prompts only the specified geometry and materials change, keeping the massing intact. That is the vendor's claim. You still compare the output with your elevation.

Do the comparison with a short list rather than a glance. Open the image next to the elevation of the option and list every feature that appears in the picture: roof form, number of storeys, window positions, rooflights, chimney, balustrades, planting that hides something. Then tick each against the drawing. A typical mismatch (illustrative): the picture of Option A shows a pitched roof with a rooflight, while the model has a flat roof at 3.6 m and no rooflight. Either the render is re-run with the geometry locked, or the picture is captioned "mood only", or you draw the elevation honestly and use the image for material and light. Never let a picture promise a feature the plan does not have. Add a caption to every image: "Concept image, not to scale, not a design". The Midjourney tutorial for architects covers image workflows for client meetings in more detail.

Rights and client data deserve a minute before you use any image tool on a live job. Commercial-use terms and indemnity differ by vendor and plan, and you should read yours. The tutorial on copyright and client data risks lists the questions.

The comparison matrix, and what the architect decides

With the numbers measured, build one matrix per version of the register. Cells hold numbers, not adjectives:

The matrix does not choose. It tells the architect what each option costs against the brief, and it shows B is out unless the brief or the rules change. A is short on living space by 2 m². It can grow by that much and stay under the 56 m² extension cap, but only if the rear boundary distance of 6.4 m survives the change, and the model must confirm that. C is short on the study by 0.4 m², which a wall move of about 0.2 m along a 2 m side might fix, though that takes 0.4 m² from the living room and leaves 42.1 m², which still passes R1. The architect decides which of those trades to take to the client, and with what reasoning about light, structure and cost.

Then ask the assistant for a plain-English summary of the trade-offs, using only the matrix as input, and edit it. The quick sum below shows why you keep the register live. In week two the client adds a 12 m² home gym:

• A: 52 + 12 = 64 m² breaks the 60 m² budget limit, and coverage becomes (96 + 64) ÷ 380 = 42.1%.
• B: 58 + 12 = 70 m², with coverage of 43.7%.
• C: 55 + 12 = 67 m², with coverage of 42.9%.

Every option now fails R7 and R9. Version 2 of the register makes that visible in a minute, and the client conversation becomes a real choice between the gym, a second storey, a larger budget or a planning discussion about the coverage limit. Without a register, someone redraws all three options and discovers it on the third.

One check per step, in one place

Keep this table on the project file. It is the whole discipline in six lines, and each row is a gate: nothing moves on until its check is done.

The pattern behind the table is the same each time. The assistant or tool produces something fast, a person with the source in front of them confirms or corrects it, and the confirmed version goes into the register or the matrix. If a step has no check, treat its output as a rumour.

What the toolkit costs and where the hours go

Costs below are list prices in USD at the time of writing, and the hours are planning figures for one project of this size:

The $100 is a chat plan at $20, Maket's Homeowner plan at $20 and Snaptrude Pro on monthly billing at $60. Your own baseline for a three-option concept phase is your timesheets, so compare these 12 hours with those rather than with anyone's average. If your practice already owns a modelling tool and a chat plan, the extra spend for the first project can be close to nothing.

Site, structure and quality calls the architect keeps

The workflow above removes clerical effort, not judgement. These stay yours:

• Visiting the site, and reading orientation, levels, neighbours and the tree on the ground.
• Structure and buildability: spans, openings, routes for services, how the thing is built.
• What a planning rule means for this site, and how much risk to take on an exception.
• Whether an option is any good: proportion, light, the way a room feels, the client's unspoken preferences.
• The client relationship, including the conversation when two requirements conflict.
• Professional liability: the drawing is yours whether or not an assistant helped make it.

Write the practice's rules for client data and AI use down once, and keep them with the project file. The architecture AI policy tutorial has clauses to adapt. A short note on how each AI step was checked against the register is also worth keeping, because it is what you would show an insurer or a client who asked how the options were made.

Questions small practices ask about AI concept work

Can AI-generated concepts be shown to a client?

Yes, as concept images with a caption saying they are not to scale and not a design. Show the numbers from your own model beside them. Before using any image tool on client work, check its commercial-use terms and whether the plan you pay for covers them, because the terms differ between plans and between vendors.

Does an AI planning-rule check replace pre-application advice?

No. It turns a policy document into a checklist you can verify, and it flags exceptions worth raising. The authority's officer still decides what a rule means for your site, and policies change. Use the check to arrive at the pre-application conversation with precise questions, where that service is offered, rather than to skip it.

How do I stop the brief drifting as options multiply?

Version the register. Any change, whether from the client, the planning rules or the budget, creates a new version that the client confirms by email, and every option is re-checked against it. Two options that were checked against different versions of the brief should never sit side by side in a comparison.

Which tools should a three-person practice try first?

Start with a chat assistant for the register and the rule extraction, and the modelling tool you already use for massing. Add a floor-plan generator or an image tool only after the register is working, and trial the cheapest plan for one project. Each added tool is another output you must check against the brief.

Further reads

• AI for Architects: Draft Design Schematics From a Client Brief — Go deeper on drafting schematic options from a brief.
• Claude vs ChatGPT vs Gemini for Architects: Which to Pay For — Compare which assistant a practice should pay for.
• Architecture Firms Adopting AI Design: What to Tell Your PI Insurer — What to tell your professional liability insurer about AI design.
• AI Tools for Architects 2026: A Toolkit From Brief to Handover — A toolkit mapped from brief to handover.
• Can AI Draft a Design and Access Statement? What the Architect Checks — Draft the statement that follows concept options, with checks.
• How to Use AI in an Architecture Firm: A First-Quarter Pilot Plan — Plan a first quarter of AI across the whole practice.
• How to Use Claude AI in Architecture: Specs and Planning Reports — Summarise specifications and planning documents with Claude, check every clause against the original, and keep client material out with a never-paste list.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Maket pricing page (credits, uploads, accuracy note, exports), Snaptrude pricing page, the Veras product page on Chaos's site, and OpenAI and Anthropic plan pages; the site, planning extract and project numbers in the example are illustrative.