AI Floor Plan Generation for Architects: Where It Fails a Real Brief

Yes as a sketching aid, no as a drawing. AI produces plausible floor plan options in minutes, but each one is an unchecked diagram until someone tests it against the brief, circulation, daylight, structure, services and regulations. Before a client sees one, score it against a written brief, then get an architect's sign-off.

The trouble isn't that the plans look bad. They look fine, which is the problem: a layout can have well-proportioned rooms and still move the staircase, build over an inspection chamber or leave the accessible shower room a few centimetres too narrow. Vendors say as much. Maket's page describes its outputs as design-stage documents and says a licensed architect or engineer should review the design against site conditions and local building codes.

Follow me on Instagram@sagnikteaches

Write one real brief as a ledger the plan can fail

Don't start by trying tools. Start with a brief you already know the answer to, ideally a closed project from your own files, and turn it into a ledger: every requirement as a numbered line, marked either hard (a failure stops the review, and some hard lines are simply facts about the site) or soft (scored 0 to 2). The ledger is the real test. Without it, the plan is judged on whether it looks nice.

Connect on LinkedInSagnik Bhattacharya

The brief below is a composite I made up for illustration. A couple with two children and a parent who sometimes uses a wheelchair want to extend a detached two-storey house. The ground floor is 62 m² gross, the rear faces south-west, the staircase sits in a central hall, and a chimney breast on a side wall stays. They want an open-plan kitchen, dining and living space of at least 38 m², a 4 m² utility room, a ground-floor guest bedroom of 11 m² with an accessible shower room of 5 m², and an 8 m² home office with its own door.

Subscribe on YouTube@sagnikteaches

If your practice already writes room data sheets from a brief, you have most of this in another form, and room data sheets built from a client brief shows how to produce them. A ledger differs in one respect: each line is something a plan can visibly pass or fail.

What the tool can be told, and what it can't

Now read the generator's input form against the ledger. For Maket, the home page describes its inputs as the rooms you need, your target floor area and your preferred house shape, and says you can upload an existing plan to rework it. It doesn't mention a site boundary, orientation, an inspection chamber or an accessibility standard. That isn't a criticism of the product; it is a description of what it was built for. But it tells you that H1, H2, H5 and H6 will have to be checked after the event, by you, because the tool can't be given them. H3 and H4 sit in between: an uploaded plan shows the staircase and the chimney breast, but nothing on the page says the tool treats them as fixed, so whether they survive is the first thing a run should tell you.

Write down, for each tool you test, which ledger lines it can accept as inputs and which it can't. This list is more valuable than any feature comparison, because it shows where the review work lands. If a vendor claims more, test the claim. Finch's pricing page, for example, lists a "code and compliance check" under its Enterprise plan from three seats, with AI-powered floor plan generation. It is a reasonable thing to ask a vendor which rules the check covers, for which jurisdictions, as of what date, and who is responsible when it misses one. Its prices are in euros on the page, so I haven't converted them here.

It helps to see what the brief becomes once it is squeezed through inputs like those. For the extension above, the generator's form would get something like this:

Rooms: open-plan kitchen, dining and living; utility; guest bedroom; accessible shower room; home office
Target area: 66 m2 of new and reworked space
House shape: rectangular
Existing plan: uploaded ground floor (stair, hall and chimney breast drawn, nothing marked as fixed)

Look at what has gone: the boundary, the chamber, the turning circle, the separate office door, and the fact that the stair and the chimney breast must not move. The uploaded plan shows those two, but the form has nowhere to say they are fixed. What remains is a room list and a picture of the existing house, which is the easy part of the job. If a generator returns something good from that, it has been lucky rather than informed, and the ledger is how you find out which.

Run the test in an afternoon: one brief, three runs

The method is deliberately dull, so you can repeat it for every tool and every upgrade.

1. Prepare (45 minutes). Write the brief and the ledger. Prepare the site overlay: existing plan, boundary, chamber and the rear wall line, at one scale.
2. Generate (15 minutes). Run three options with the inputs the tool accepts, including one from the uploaded existing plan if the tool takes one. Save every output with the settings used.
3. Overlay and score (25 minutes per plan). Place each output on the overlay. Tick hard lines pass or fail, then score soft lines 0 to 2.
4. Compare with your own answer (20 minutes). Put the best generated plan beside the design you actually produced. What did the tool see that you didn't? What did it miss that you'd never miss?
5. Record (10 minutes). Hard fails, soft score, minutes spent. Keep the record for the next tool.

Mind the credits. Maket's pricing page charges 20 credits per floor to generate or upload a floor plan, and 10 for a render. The Free plan gives 30 credits once, with single-floor generation only, which covers one generation, and the page says downloadable exports are for paid plans only, so on Free you would be scoring from screenshots and the overlay becomes guesswork. Homeowner, at $20 a month, lists 300 credits a month, multi-floor generation and PDF, DXF and render exports; Pro, at $100 a month, lists 3,000. This test, one upload plus three single-floor options, costs 80 credits, so a month of Homeowner covers it about three times over. The home page mentions PDF and image files rather than DXF, so check what you actually receive before you plan an overlay around a DXF.

A generated plan with typical faults, scored line by line

To show how the ledger works, here is a plan I have invented as an illustration of typical faults. It is not the output of any named product. Imagine the generator returns this ground floor: the extension is 3.8 m deep and clears the boundary. The living, kitchen and dining space runs across the back with 36 m². The staircase has been rotated and moved 1.2 m towards the side wall to enlarge the hall. A utility room sits in the rear west corner with its own external door. The shower room is 1.5 m by 2.0 m beside the guest bedroom, with an inward-opening door. The home office is entered through the utility room.

The tally is 4 hard passes and 4 hard fails, and 8 soft points out of 12. Any hard fail means the plan isn't client-ready, and four means it is a diagram and nothing more. The salvageable part is the idea: a utility room beside the rear door, with a kitchen and dining spine along the back. The geometry gets discarded and redrawn. That is a typical, useful outcome. The tool gave you an arrangement to react to, and the ledger stopped you from sending it to the client.

Notice the sums as well. If hand-sketching an option takes you about 30 minutes and scoring a generated one takes 25, the saving is five minutes an option before you count fixes. The value of generation lies in the options you wouldn't have thought to draw, not in time saved on a tight brief. On a blank-sheet brief with few hard lines, the sum looks much better.

On the plan itself, write a short note so the next person knows what to do with it:

Plan R2: diagram only. Keep: utility beside the rear door; kitchen and dining spine along the back. Discard: stair position, utility location over the chamber, shower room geometry, office access. Next step: redraw on the overlay with H3 and H5 fixed, then re-score.

Ten checks, in the order I'd run them

1. Overlay on the survey. Put the output on your site and existing plan at the same scale. Typical miss: the footprint crosses a boundary or a setback line.
2. The hard lines. Run through them first and stop at the first fail, because a plan that fails H3 isn't worth scoring further.
3. Area schedule. Measure every room and compare it with the programme. Typical miss: a room labelled home office that measures 4 m².
4. Circulation. Trace the route from the front door to every room with doors swung. Compare circulation as a share of gross area with your last three similar jobs. Typical miss: a bedroom reached through a kitchen, or a corridor that eats a living room.
5. Doors and clear floor space. Draw swing arcs, turning circles and furniture. Typical miss: a door that hits a toilet, or an island that closes off the aisle.
6. Structure. Mark walls to keep, walls removed and any new opening. Typical miss: long unsupported spans and openings in walls that may carry load, which need an engineer.
7. Services and wet rooms. Locate stacks, drainage routes, flues and ventilation paths. Typical miss: a wet room with no external wall and no path for ventilation.
8. Daylight and orientation. Count windows to each habitable room and check the sun side. Typical miss: windowless rooms, or living space on the dark side.
9. Escape, accessibility, fire and ventilation. Use your own jurisdiction's checklist. The generator hasn't seen it. Typical miss: a bedroom whose only exit passes through a kitchen, which can fail escape rules in many places.
10. Buildability and budget. Ask what the structure would need to look like and whether it fits the cost plan. Typical miss: cantilevers and awkward roof forms that nobody priced.

Treat checks 6, 7 and 9 as the ones you never delegate to a tool. They are where responsibility sits with you, whatever the software says. If you want a structured way to check a whole set before issue, AI drawing review before you issue an architectural set covers the next stage.

Upper floors: stacking faults a single-floor test misses

Multi-floor generation is listed on Maket's paid plans, and it adds a class of failure the ground-floor test never shows: each floor can look tidy on its own while the two don't stack. If your brief takes in the first floor, or you test the tool on a two-storey new-build, score each floor separately, then lay the upper floor over the lower one and run four more checks.

• Stair and landing. The stair arrives where it leaves, the landing has room to turn, and no upper wall crosses the stair void or cuts the headroom over the flight.
• Wet rooms over wet rooms. Bathrooms and en-suites sit over the kitchen, the utility or another wet room, or over a route where a soil pipe can drop without passing through a living space.
• Walls over walls. Every upper partition lands on a wall below or is marked as needing support. A wide ground-floor opening under a bedroom wall is a question for the engineer, not a detail.
• Escape from upper rooms. Bedrooms meet the escape rules your jurisdiction sets for upper floors, which the generator has not been told.

A made-up two-storey run shows the pattern. The ground floor scores reasonably, and the first floor offers three bedrooms, a family bathroom and an en-suite in a tidy arrangement. Overlaid, the en-suite sits above the middle of the new open-plan living space, so its soil pipe would have to drop through the ceiling the clients care most about, and the main bedroom's new partition lands over a 5 m opening with nothing beneath it. Neither fault shows on either floor alone. Add them to the ledger as hard lines for any two-storey brief, H9 for wet rooms stacking over wet rooms or a service route and H10 for upper walls over walls below, so every later tool is tested against them from the first run.

Using a chat assistant to attack the plan, not draw it

A chat assistant is poor at drawing and quite good at reading. Give it your ledger and a written description of the plan, and ask it to find the lines the plan breaks. It can't measure from text, but it can catch logic faults you stop seeing after your fifth plan of the day.

Here is a brief ledger (H = hard, S = soft) and a written description of a generated plan.
List every ledger line the description appears to break. For each, quote the sentence in the
description that causes it. Where the description doesn't contain enough information to decide,
say exactly which dimension or position you would need. Do not redesign anything and do not
say a line passes unless the description proves it.
LEDGER: [paste]
PLAN DESCRIPTION: [paste]

A reasonable reply to the plan above might be this (shortened):

H3: breaks. "The staircase was rotated and moved towards the side wall." The ledger says the stair stays.
H5: cannot decide. "Utility room in the rear corner." I need the chamber's position against the rear corner.
H6: likely breaks. "Shower room 1.5 m by 2.0 m with an inward-opening door." I need the turning-space dimension from your standard to be sure.
H7: breaks. "Home office is entered through the utility room." The ledger asks for its own external door.

Read it critically. On H5 the assistant says it can't decide, but you can: you have the chamber on your overlay and the utility sits over it. Use the output as a list of questions to check on the drawing, not as a verdict. It is also only as good as the description, which is why you write that description yourself rather than letting the tool summarise its own work.

Keep a test log across briefs and tools

One brief tells you about one afternoon. Five briefs tell you about a tool. Keep a simple log, with a row per run, so patterns show up. This extract uses invented figures:

The pattern is the useful part. If your own log showed something like this, with many hard fails where lines tie the layout to an existing building and few on a nearly blank sheet, it would tell you where to point the tool: early options for new-build or open-plan conversions, not alterations to a tightly constrained house. Yours may show something different, and that is the reason to keep one.

What the client sees, and what stays in the office

Nothing leaves the practice straight from a generator. A sensible sign-off ladder has four steps: the generated output, your ledger score, a reviewer who has checked the ten items above, and a named person who decides whether it is shown. Keep a short record per plan: plan ID, brief version, tool and settings, ledger score, reviewer, date and status (diagram only, developed or rejected).

When a plan does go in front of a client, label it: "Concept layout diagram. Not to scale. Not checked against structure, services or regulations. For discussion only." Send a picture or PDF, never the editable file or the DXF, so it can't be mistaken for a drawing. Don't send a generator's output with its branding or credits visible without checking the tool's terms for commercial use. The Maket pricing page I read says nothing about commercial use, so ask the vendor.

The way this usually goes wrong is mundane. A junior sends a client a generated option as "a quick idea" on a Friday afternoon. The client forwards it to a builder friend, who prices it from the picture, and on Monday the client asks why the fee proposal covers design work for a scheme that "already exists". Nothing in that image was checked, yet it now anchors the conversation, and its moved staircase has a price attached. The fix is procedural rather than technical: generated outputs live in a folder the client never sees, and only a plan with a reviewer's initials and the concept label leaves the office.

Three practice-level questions need answering before any of this is routine. Who owns and may reuse the output, which copyright and client data risks of AI in architecture practices discusses. What does your professional indemnity insurer need to know, which what to tell your PI insurer about AI design covers. And what client information, such as a brief, a site plan or photos, you are willing to upload to a tool you don't control. Free tools deserve extra caution, and the limits of free AI architecture design software lists what to look for.

Where generated layouts earn a place in a small practice

None of this means the tools are useless. They're useful at the start. In the first hour of option-finding, a generator can show arrangements you hadn't drawn, test whether a room programme physically fits the footprint, and give a hesitant client something to react to. A client who says "not that one, I hate the long hall" has told you more in a minute than a page of brief.

My decision rule has three lines. Use generated layouts for early options only. Never show one you wouldn't be prepared to develop. Never issue one: whatever goes out as a drawing was drawn, checked and signed in your own system. For turning the brief itself into schematic options, drafting design schematics from a client brief shows where AI helps before a plan exists.

Finally, repeat the test. Every vendor update changes what the tool does, and the ledger lets you rerun the afternoon in an hour. When a new version passes five briefs with no hard fails, you will know it from evidence, and your clients will see only plans that have earned it.

Questions architects ask about generated floor plans

Can I use a generated plan as the base for a permit application?

No. Treat it as a sketch and redraw it. A permit set needs dimensions, structure, services and compliance information with an accountable author, and a generated plan has none of that behind it. If a generator's output is the starting geometry, trace and verify it on your own CAD or BIM base, so the drawing you issue is demonstrably yours.

Do generators take account of the existing building's structure?

Don't assume so. Maket's home page lists the rooms you need, a target area and a preferred house shape as inputs, and says you can upload an existing plan to rework it, but it mentions no site boundary or orientation. An uploaded plan is not the same as walls marked as fixed, so test whether the stair and any structural walls survive a run. List every ledger line the tool cannot accept; you will check those yourself, afterwards, by overlay.

How many generated options should a client see?

One to three, each reviewed by you and each labelled as a concept. More than three turns the meeting into a menu, and every extra option is one more plan you have to have checked. If the client wants to see everything the tool made, say that you only show options you would be willing to develop.

Should I tell clients I used an AI tool?

Yes, in plain words in the fee proposal or appointment terms: early option-finding may use AI tools, and every design is reviewed and issued by named staff. Check how your professional indemnity cover treats AI use, and keep a record of what was generated and what you changed. Being open early is easier than explaining later.

Further reads

• AI Tools for Architects 2026: A Toolkit From Brief to Handover — A toolkit from brief to handover, stage by stage.
• AI Software for Architectural Drawings: Notes, Schedules and Sheet Checks — Notes, schedules and sheet checks on a drawing set.
• AI Construction Design Software: What Renovation Contractors Can Draw — What a contractor can draw with AI, and where the architect steps in.
• Best AI Tool for Construction Design: What It Drafts and What Needs a Stamp — What AI drafts for construction design and what needs a stamp.
• AI Policy for Architects: A Ready-to-Use Checklist and Clauses — Clauses for a practice AI policy you can adapt.
• AI for Retail Store Design: Layout and Display Ideas — Plan a shop re-layout with AI, safely: measure, get three options, tape-test the best, check exits and access, then measure whether sales moved.
• Best AI for Shop Design: Floor Plan and Display Brief for Your Fitter — A shop owner's guide to AI for a refresh: measure, draft a layout, mock up windows and signs, then hand the fitter a brief they can price and check.
• Can AI Design a Restaurant? Layouts, Concept Images and Limits — AI can build a restaurant mood board and a design brief in a weekend. A pre-designer checklist covers what its layouts miss on fire, access and kitchen flow.
• Best AI for Architecture and Interior Design: Concept Image Tools by Cost — What concept-image tools cost a small architecture or interior design studio a month, and what each licence lets you show a client.
• 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 and home page (inputs, outputs, credits, review advice), Finch pricing page. Pages read October 2026. The sample brief, ledger and generated plan are invented for illustration and are not the output of any named product.