Yes, for the thinking that comes before the drawing: AI can turn a client brief into a checked list of requirements, a schedule of accommodation, adjacency diagrams and two or three option concepts. It cannot be trusted with dimensions, building-code compliance, structure or services coordination, so the architect still produces the plans and signs off every number.
The split follows one rule: AI is good with words and relationships and poor with geometry. A brief is made of words, so the early stages go well. A plan is geometry, and a generated plan tends to look right while being wrong in ways only a scale rule reveals: a door that swings into a stair, a bathroom too small for a wheelchair to turn in, a bedroom with no wall long enough for a bed.
What goes to AI and what stays on your desk
Schematic design has seven jobs that AI touches differently. Four are language and relationships; three are geometry and judgement. A workable split, which your practice will adjust by a line or two:
| Job | AI can draft | You do by hand |
|---|---|---|
| Reading the brief | Requirements list, contradictions, missing information, questions for the client | Deciding which tensions to raise, and how, with this particular client |
| Schedule of accommodation | Room list with target areas and notes | Area arithmetic in a spreadsheet; sizes driven by accessibility or code |
| Adjacency and zoning | Pair scores, bubble diagram, zoning ideas | Judging noise, privacy, daylight, orientation and structure |
| Concept options | Three organising ideas with strengths and risks | Choosing, developing and testing against the real site |
| Massing and mood images | Image prompts, reference sets, rough visuals | Real massing in your modelling software: scale, storeys, heights |
| Plans, sections, elevations | Nothing you would issue | Every dimensioned drawing, door swing, stair, escape route and coordination point |
| Client presentation | Wording of option summaries and question lists | What to show, the caveat on every image, the conversation itself |
The first four jobs repay the effort. The last three are where practices get hurt, because the output looks finished. A render with a pitched roof and a sunny terrace feels like progress, and it is also the thing a client will remember and hold you to.
One family brief, with its problems left in
Every figure from here on is invented. The practice is small, the enquiry is for a new-build house, and the family numbers five: one parent works from home three days a week, and a grandparent is moving in who uses a wheelchair full time. The plot is 18 m by 35.5 m (about 640 m²), street on the north edge, garden to the south. The construction budget is $640,000 and the client's wish list says "about 210 to 230 m²". The architect's own comparator, from the last three similar jobs, is $2,600 to $3,100 per m² gross.
The wish list, as the client wrote it:
- Four bedrooms, one of them a ground-floor suite with a level-access shower room and a small sitting area.
- Open-plan kitchen, dining and living space "with a table for eight".
- A separate study, quiet enough for video calls.
- Utility and boot room, plus a ground-floor WC.
- Two covered car spaces, and garden access without steps from the suite and the living space.
That text reaches the AI only after it has been cleaned of names and health detail. How, and why, is covered near the end of this page.
Stage one: ask the AI to attack the brief
A common first move is to ask for a summary, which returns the client's own wishes in tidier words. Ask for an attack instead: what contradicts what, what is missing, what the budget will bear. A prompt built from five parts keeps the answer usable: the role, the context you supply, the inputs tagged so they cannot be confused, the task in numbered steps, and the rules, including an instruction to show arithmetic.
You are helping an architect read a client brief before any drawing starts.
CONTEXT: new-build family house. Plot 18 m x 35.5 m, street on the north edge,
garden to the south. Construction budget $640,000. The architect's cost range for
comparable houses is $2,600 to $3,100 per m2 gross.
BRIEF (anonymised): [paste the cleaned brief here]
TASKS
1. List every requirement as one line, tagged BRIEF (stated) or ASSUMED (you inferred it).
2. List contradictions between requirements, quoting the two lines in tension.
3. List missing information that would change the layout, as questions for the client.
4. Compare the wish-list area with what the budget supports at the cost range above.
Show your arithmetic.
RULES: do not propose a layout. Do not invent rooms or areas. If unsure, say so.A realistic result looks like this, with a note on what you would do with each line:
| Type | What the AI returned | What you do with it |
|---|---|---|
| Tension | "Table for eight" open-plan space against a study that must be quiet for calls | Ask where calls happen; propose the study off the hall, not off the living space |
| Tension | Suite described as private, yet "close to the kitchen for meals" | Ask which matters more; the options will trade one against the other |
| Gap | No bin, bike or garden store anywhere in the brief | Add an external store to the schedule, marked ASSUMED |
| Gap | "Level access" undefined: threshold height, ramp, hoist? | Ask whether an occupational therapist has assessed the grandparent's needs |
| Assumed | Bedrooms three and four at 13 m² and 11 m² | Confirm with the client before they harden into the schedule |
| Budget | $640,000 supports about 206 to 246 m² gross at $3,100 to $2,600 per m² | The 210 to 230 m² wish list is affordable only in the middle or lower half of your range: say so at the first meeting |
Two things the AI missed, which you would add: the suite's sitting area competes with the kitchen for the same south frontage, and nobody has asked about parking for the grandparent's transport. The first is a design judgement and the second comes from having done this before. The model reads what is on the page; it has not stood in a client's kitchen.
The budget line is also where I would be careful. The model's division is easy to check (640,000 divided by 3,100 is 206; divided by 2,600 is 246), but the cost range is yours. Never let it supply a cost per m² of its own.
Stage two: a schedule of accommodation, with the sums in a sheet
Ask for one row per space with three columns (area, source marked BRIEF or ASSUMED, and a note) and tell it to leave the totals to you. The totals go in a spreadsheet, for a reason the next table makes plain. Here is a first draft from a run that did not carry that instruction, set against the architect's corrected version:
| Space | AI draft (m²) | Checked (m²) | Why it changed |
|---|---|---|---|
| Hall and coats | 9 | 9 | No change |
| Kitchen, dining, living | 46 | 52 | A table for eight, an island and a sofa group need more floor |
| Utility and boot room | 8 | 8 | No change |
| Study | 10 | 10 | Position changes, not size |
| Ground-floor WC | 3 | 3 | No change |
| Suite bedroom | 16 | 16 | No change |
| Suite shower room | 5 | 7.5 | A typical ensuite size; this client needs level access and a turning space |
| Suite sitting area | 10 | 10 | No change |
| Plant and storage | left out | 4 | Architect adds heating, hot water and electrics space |
| Primary bedroom, ensuite, dressing | 29 | 29 | No change |
| Bedrooms three and four | 24 | 24 | No change |
| Family bathroom | 7 | 7 | No change |
| Media room | 14 | 0 | Not in the brief; the AI invented it |
| Total net | "188" stated; rows add to 181 | 179.5 | Sum in the sheet, always |
Four of those corrections are the usual kind: a living space sized without the brief's table for eight, a dropped line, an invented room and a size taken from a typical house rather than this client. The fifth is the AI's own total. It stated 188 beside rows that add to 181, and neither matches the corrected 179.5. Language models can do arithmetic and still slip on a column of decimals, which is why the sums live in =SUM() cells, the model only fills the rows, and the prompt now says "no totals".
Gross area comes from your own net-to-gross factor, not the model's guess. At an assumed 82%, 179.5 m² net divides to about 219 m² gross, inside the client's range and, at the middle of your cost range, inside the budget too.
One habit pays for itself: back-test the process on a closed project. Feed in the original brief from a house you have already built, produce the schedule, and compare it with what was drawn and built. The gaps tell you what your chosen assistant routinely misses, and you learn it before a live client does. For the room-by-room detail that follows the schedule, see how to produce room data sheets from a client brief.
Stage three: adjacency scores first, diagram second
Ask for adjacency as scores with a reason quoting the brief, on a scale such as +3 (must touch), +2 (prefer), 0 (indifferent) and −2 (keep apart). The scores matter more than the picture, because you can argue with them one pair at a time.
| Pair | Score | Reason given |
|---|---|---|
| Suite bedroom and suite shower room | +3 | Daily use; transfer distance matters |
| Suite and garden | +3 | Brief asks for step-free garden access |
| Suite and kitchen/dining | +2 | Meals shared; "close to the kitchen" |
| Utility and kitchen | +3 | Laundry and food prep share services and routes |
| Study and open-plan living | −2 | Video calls need quiet |
| Study and entrance hall | +1 | Quiet end of the house, visitors never pass the living space |
| Bedrooms three and four and family bathroom | +3 | Children's daily route |
| Plant and utility | +2 | Shared pipework and electrics |
Then ask for the same pairs as diagram code, which free viewers will draw. Graphviz's DOT language works, and so does Mermaid. The output is a few lines of text:
graph G {
"Suite bedroom" -- "Suite shower" [label="+3"];
"Suite bedroom" -- "Garden" [label="+3"];
"Suite sitting" -- "Kitchen/dining" [label="+2"];
"Utility" -- "Kitchen/dining" [label="+3"];
"Study" -- "Living" [label="-2", style=dashed];
}Treat the rendered diagram as a prop for the conversation, then redraw it on tracing paper over the plot outline, where the real constraints (the north street edge, the boundary limits, existing trees) are. A typical failure: the first diagram places the suite beside the primary bedroom upstairs because both are "bedrooms". The scores said otherwise, which is why you keep them.
Stage four: three options that differ in organising idea
Ask for options that differ in how the house is organised, not in cladding or roof shape. Give the model the cleaned brief, the schedule and the scores, and add rules: each option must keep every +3 pair adjacent, state its main weakness in one sentence, and say what it costs relative to the others.
Using the schedule and adjacency scores above, propose three layout options for the house.
Each must differ in organising idea (not style). For each give: a three-line concept,
which +3 pairs it satisfies, the single biggest risk, and the effect on external wall length.
Do not give dimensions. Do not draw. Mark every assumption.What comes back, condensed to what you would keep:
| Option | Organising idea | Gets right | Risk to check | Cost note |
|---|---|---|---|---|
| A: Wing | Single-storey suite wing on the east side, two-storey block to the west, patio in the angle | Step-free garden from suite and living; privacy for the suite | Long external wall; wing must clear the plot's side limits | Compare envelope length with B |
| B: Bar | Compact two-storey rectangle across the plot, suite at the rear of the ground floor | Least wall per m², simplest structure | Suite sits under the primary bedroom; deep plan starves the study of daylight | Baseline |
| C: Pavilions | Suite as a separate pavilion joined by a glazed link | Independence; easy to repurpose later | Link is area the brief did not ask for; two roofs | 8 m² link at $2,850 per m² is $22,800 before the extra wall |
Use your own must-haves as kill criteria. If the client's one non-negotiable is step-free garden access from the suite, any option that needs a ramp is out, however good its section looks. Two surviving options is the right number to sketch. If the model produces six, ask it to rank them against your tests and discard the bottom four without explanation, because explanations for discarded options are where the time goes.
If you want the rendering side of concept work in more depth, the tutorial on concept images for client meetings covers one image tool in detail, and the tutorial on floor plan generation shows where plan generators fail a real brief.
Sketches and mood images: where the picture lies
Three kinds of tool do three different jobs. General chat assistants do stages one to four. Image generators (Midjourney, Adobe Firefly and the image features inside the chat assistants) produce mood and massing pictures. Your own modelling software, SketchUp or Revit or whatever you use, produces the real massing, with the schedule's areas as targets. The mistake is asking the second to do the third.
An image prompt needs the same discipline as a brief. Compare a lazy prompt with a constrained one:
- Lazy: "modern family house with garden, photorealistic". The result will be handsome, anonymous and unrelated to your plot.
- Constrained: "Two-storey timber-clad main block with a single-storey wing to the east forming an L around a south-facing paved patio; flat green roof on the wing, pitched roof on the main block; level threshold from patio into a ground-floor room; eye-level view from the garden; overcast light; no people, no text." Now there is something to check.
Check every generated image against five points before it goes anywhere near a client:
- Count the storeys and compare them with the option.
- Look for steps or a raised threshold at the suite door.
- Match windows to rooms: a bedroom window on a wall that the plan puts against the boundary is a fault.
- Test the shadows against the north-south axis you gave it.
- Check the roof and the walls could plausibly hold each other up. A cantilever with no visible support is a picture, not a structure.
Label every image "concept only, not to scale" on the image itself, not in the covering email. Run a reverse image search on any picture that will be shown to a client, in case it resembles an existing building too closely.
On cost: Adobe lists Firefly Standard at $9.99 a month (2,000 generative credits) and Pro at $19.99 (4,000), and says its own Firefly models' outputs can be used commercially. For partner models inside Firefly, it says you are responsible for deciding whether the content suits your project. If you need one look across several options, Midjourney offers style references and moodboards, and Firefly has a style reference that works only with Adobe's own image models, as last recorded. Prices and plan names move often, so confirm on the vendor page before you subscribe.
Checks that stay with the architect
Whatever the model produced, these are the checks that decide whether an option is fit to show. Use the same table every time, so nothing depends on memory.
| Check | How | Example result |
|---|---|---|
| Area arithmetic | Sum the schedule in a sheet | 179.5 m² net |
| Gross area | Net divided by your efficiency factor | 179.5 ÷ 0.82 = 219 m² |
| Cost at your rates | Gross × your low, middle and high rate | $569,400 / $624,150 / $678,900 against a $640,000 budget |
| Break-even rate | Budget ÷ gross area | $640,000 ÷ 219 = about $2,920 per m² |
| Footprint against site limits | Check the rules you looked up yourself | About 146 m² ground floor on a 640 m² plot |
| Accessibility | Door widths, turning spaces and thresholds against the standard you work to | Suite shower room redrawn at 7.5 m² |
| Stairs and levels | Draw a section through the stair and the suite threshold | Headroom and level change confirmed on paper |
| Daylight and orientation | Sun path on the actual plot | Study daylight confirmed or moved |
| Structure and services | Spans, bearing walls, risers, plant position | Raise with the engineer before the second client meeting |
| Open questions | Count the unanswered client questions | Two outstanding: store size, transport parking |
The cost rows deserve a second look. At $2,600 per m² the 219 m² house costs $569,400, at the $2,850 midpoint it costs $624,150, and at $3,100 it costs $678,900, which is $38,900 over. The budget holds only if the rate comes in at about $2,920 or below, a little above the middle of your range. Saying that on day one is more useful to the client than any rendering.
Client data and copyright: the pre-flight check
A brief contains more than design requirements. It holds names, ages, health conditions, finances and sometimes a neighbour's complaint. Decide what goes into the model before you decide which model to use.
| In the pack | Risk | Do this instead |
|---|---|---|
| Names, addresses, children's ages | Personal data sent to a third-party service | "Client", "Child 1 (9)"; leave addresses out; describe the plot by dimensions |
| Health information | Treated as sensitive personal data in many places | State only the design requirement, not the diagnosis |
| Budget and finance | Commercially sensitive | Give a cost-per-m² band, or keep the budget out and test it in your sheet |
| Surveyor's drawings and site plan | The surveyor's copyright and licence terms | Type in the dimensions you need; do not upload the file |
| Consultants' drawings | Confidentiality owed to others | Keep them out of the tool entirely |
| Client's reference images | Copyright in images you may not hold | Use for your own reference only; do not feed to generators |
Here is the same sentence before and after cleaning. Before: "Mrs [surname]'s mother, 81, recovering from a stroke, uses a wheelchair and must be able to reach the garden alone." After: "One occupant is a full-time wheelchair user; assume level access to the garden and a 1,500 mm turning space in the suite bedroom and shower room." The second version gives the model what it needs and leaves nothing sensitive in the prompt. Swap the 1,500 mm for whatever turning space your own standard requires.
Then there is the plan you use. Claude's pricing page, as of October 2026, lists Free, Pro and Max as "Opt-out" for model training and Team and Enterprise as "None by default"; Team seats are $25 a month billed monthly or $20 billed annually, for teams of 2 to 150. ChatGPT Business is similar: no training on business data by default, $25 a seat monthly or $20 billed annually, two seats at least. On individual plans, such as ChatGPT Plus at $20 a month, Claude Pro at $20 monthly ($17 billed annually) and Google AI Pro (last listed at $19.99), the training switch sits in privacy settings and its label changes, so look at it on the day. A small practice may find the business plan worth its price on this point alone. The wider data-handling question has its own tutorial: copyright and client data risks of AI in architecture practices.
Copyright is the other half. Anthropic's consumer terms assign you whatever rights it holds in outputs, with the words "if any" in the sentence, and OpenAI's terms are worded similarly. That is an assignment of rights that may not exist. One national copyright office's January 2025 report concluded that prompts alone do not give a user enough control over the expressive result to make them its author, while a person's own selection, arrangement and creative modification of AI output can be protected. In practice: treat raw AI sketches as working references, develop the design in your own drawings, and do not promise a client exclusive rights in an image a model produced.
There is a trap in the same area. Vendor indemnities for generated output are narrow. Adobe says "eligible enterprise customers" can receive IP indemnification for select Firefly outputs. OpenAI's output indemnity, as last read, covers its API and Enterprise-type customers rather than ChatGPT Business, and it excludes output you have edited. The changes that make a design yours can therefore take it out from under the vendor's protection. Tell your professional indemnity insurer how you use these tools; the tutorial on what to tell your professional indemnity insurer about AI design sets out what to ask.
What the half day costs, and a way to test it
With the process fixed, a first-option pack for a house like this can be put together in about half a day. The minutes below are assumptions, so time your own run on a closed project before you trust them.
| Stage | By hand (min) | With AI and checks (min) |
|---|---|---|
| Read and stress-test the brief | 90 | 30 |
| Schedule of accommodation | 75 | 40 |
| Adjacency scores and diagram | 90 | 35 |
| Three option concepts | 150 | 60 |
| Mood and massing images | 120 | 45 |
| Architect's checks (table above) | included | 70 |
| Total | 525 (8 h 45) | 280 (4 h 40) |
That is a saving of 245 minutes, a little over four hours. At an internal cost of $85 an hour it is about $350 of time on one project, set against $20 a month for an individual plan. The saving only counts if the checks stay in the right-hand column; drop them and you have bought speed with risk. A saving that matters just as much: the architect arrives at the first meeting with the contradictions already listed and the budget already tested.
If you would rather map this with someone against your own stages and software, that is the kind of job my AI implementation consultation covers.
Questions from practices trying this on a live brief
Can an AI tool give me a floor plan I can develop in Revit or Archicad?
Not reliably. You can ask for room sizes, adjacency scores and organising ideas, then block the plan yourself in your modelling software. Generated plans tend to look tidy while breaking door swings, stair geometry or wall thickness, so treat any plan image as a conversation sketch and redraw it at scale before anything leaves the studio.
Do I have to tell the client I used AI at concept stage?
Check your appointment terms, your professional body's current rules and your insurer's requirements, because they differ and are changing. A sensible default is to say in the appointment or the first option pack that AI tools may be used for early drafting, that no client personal data is pasted into them, and that every drawing is checked and issued by a named architect.
How many options should I show a client?
Two or three. AI makes ten options cheap to produce, but ten options hand the client your decision and bury the reasoning. Use the model to generate a longer list, apply your must-have tests, then present the survivors with one honest weakness each. Two well-argued options usually move a client faster than six pretty ones.
Which assistant should I use for the brief analysis?
Any current paid assistant from the major vendors can do stages one to four, so choose by your data terms and price. Run the same closed project through two of them and compare the schedules against what was built. Keep model versions out of your internal procedures, because they change every few months.
Further reads
- AI Policy for Architects: A Ready-to-Use Checklist and Clauses — Write down the data rules this workflow relies on
- How Architects Use AI for Fee Proposals, Briefs and Paperwork — Fee proposals and brief paperwork, the stage before this one
- Claude vs ChatGPT vs Gemini for Architects: Which to Pay For — Which assistant to pay for, by practice task
- AI Tools for Architects 2026: A Toolkit From Brief to Handover — A toolkit mapped across every project stage
- Best AI Tool for Construction Design: What It Drafts and What Needs a Stamp — Which AI design outputs a small builder can issue (sketches, option comparisons, specification wording) and which need an engineer's or architect's sign-off.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Anthropic consumer terms (effective 8 October 2025); Claude and ChatGPT pricing pages; Adobe Firefly plans page; OpenAI service terms as last read; a national copyright office's January 2025 report on copyrightability. All house, area and cost figures are invented for illustration.