People who search for "AI for architects" usually arrive with one of two ideas. Either AI is about to produce finished renderings at the push of a button, or it is about to take the job away. Neither holds up once you try it on a real project. What AI does well right now is take hours of writing, organizing and first-pass imagery off your plate. What it does badly is exactly the part that carries your license: dimensions, code, local zoning and coordination.
This guide follows a project that lands on small-firm desks every month: a kitchen and primary bath remodel in a 1,050 sq ft condo in Chicago, for a couple with a toddler, construction budget around US$ 165,000. Phase by phase, I will say what I would hand to AI, with which tool, what it costs and where it will fool you if you do not check.
AI for architects: the map, phase by phase
Before tools, the map. Picture the project as a line from the first inquiry email to the punch list. AI plays a different role along that line, and in some stretches it should stay out entirely.
| Phase | What AI does well | What stays with you |
|---|---|---|
| Inquiry and programming | Organize scattered answers, transcribe voice memos, flag what is still missing | Reading the clients, noticing when the couple disagrees, setting scope |
| Fee proposal | Write scope, phases, schedule and terms from your own rules | The number. Your hourly cost and your market set it, not a chatbot |
| Schematic design | Material and mood variations from a sketch or a SketchUp massing, mood boards | Layout, clearances, dimensions, existing structure |
| Design development and CDs | Checklists, sheet notes, general notes cleanup | The drawings, coordination with MEP and structural, details |
| Specifications | A complete first draft in minutes, in your format | Checking products, manufacturers, referenced standards, quantities |
| Permits and approvals | Checklists for the permit set and the condo alteration agreement | Sealing the drawings, the permit application, answering plan review |
| Construction administration | Summarizing field notes, a weekly client update, RFI log cleanup | Site visits, technical decisions, approving pay applications |
| Marketing | Captions, short scripts, turning one project into six posts | Real photos and the way you explain your decisions |
Notice the right column never shrinks. The gain comes from the middle column, which happens to be where a two-person firm loses its evenings: proposals, specs, client emails, captions.
Programming: where AI pays for itself in the first month
A remodel program shows up in pieces. A four-minute voice memo from one client, a Pinterest board with 40 pins, the condo's original floor plan as a scanned PDF and one line from the other client: "we want it open, but we don't want to spend crazy money." Pulling that into one page takes an hour. With AI it takes ten minutes, and you also get a list of what is missing.
What I do: transcribe the memo (ChatGPT and Gemini both handle audio), paste it with my notes and ask for three separate things: the program by room, the contradictions and the questions I still need to ask. The contradictions section is the valuable one. On the Chicago condo, the AI caught that the couple asked for "an open kitchen" and, in the same memo, "no cooking smells in the living room." That becomes a question at the second meeting (can the hood vent to the exterior, and will the association allow a new penetration in the facade?) instead of an argument during construction.
If you get more than five inquiries a month, it is worth turning this into an automatic flow. I walk through it step by step, with monthly cost, in ChatGPT for architects: automating the client questionnaire and fee proposal.
Renderings from a sketch: which tool for which moment
This is where most of the confusion lives. There are two very different kinds of "AI rendering," and mixing them up creates rework.
Type 1: images generated from scratch or from a loose reference
Midjourney, ChatGPT's image model and Gemini belong here. They are great for mood boards and for the "is it more this or more that?" conversation. The problem is that they ignore geometry. The window moves, the island grows 16 inches, a door appears where there is a structural column. For a client who does not read drawings, the image turns into a promise.
Type 2: images that respect your geometry
Veras (a plugin for SketchUp, Revit, Rhino and others), LookX, the AI features inside D5 Render and SketchUp Diffusion work on top of your model or your sketch. The base is yours; the AI only dresses it. For schematic design, this is where time really drops: you push the massing in SketchUp, run eight material variations in half an hour and bring two to the meeting.
My pick: for client studies I would use Veras or LookX on the model, not Midjourney. Midjourney stays in the mood board for the first meeting. For the final presentation image, the one that goes into the portfolio and sometimes into the contract exhibits, I would stay in D5 (or whatever engine you already use) with the full 3D model and use AI only for atmosphere and planting. A final rendering with invented geometry is how you end up hearing "but the picture had a window here."
One detail few people remember: on Midjourney, images are public in the gallery unless you are on a plan with Stealth mode, which only comes with the pricier tiers. If the client does not want their home on the internet, use a tool with a private option or work only with generic references.
Specifications and outline specs
Specs are the clearest win. You have an old project manual, you swap products, adjust rooms and always forget something. AI writes the whole first draft from your room list and finish schedule, using your headings and, if you want, CSI MasterFormat section numbers.
The risk is that AI fills gaps with plausible fiction. It will write "per ASTM" with the wrong standard, invent a tile thickness and name a manufacturer you never chose. The rule that works: in the prompt, forbid any standard, brand or dimension that is not on your list, and ask it to tag everything it inferred with [VERIFY]. The full prompts, with a spec section actually generated, are in AI prompts for architects.
Fee proposals: AI writes, you price
AI does not know what your work is worth. Ask ChatGPT what to charge for a 1,050 sq ft condo remodel and it returns a generic range scraped from blogs, often "8 to 15 percent of construction cost," which tells you nothing about your overhead or your market. The serious basis is still your own numbers: hourly cost, hours per phase on your last five similar projects, and what the local market accepts.
Where it helps: turning "US$ 18,500 fixed fee, three phases, two rounds of revisions" into a clear proposal with scope, exclusions (that section prevents half of all disputes), deliverables per phase and payment terms. If you use AIA contract documents such as B105 for small residential work or B101, AI can draft the cover letter and a plain-English summary of the scope, but the contract itself stays the AIA document you license. It also helps you answer the client who asks for a discount without sounding defensive.
Permits, seals and the condo board
No AI applies for a permit or seals a drawing, and it should not. Your seal means you are the architect of record under your state's licensing law, and your state board, not the software vendor, is the one that hears complaints. What AI does well is keep you from forgetting things. A per-project checklist, generated from the scope, with reminders such as: confirm which code edition the city has adopted, check whether the scope triggers a full permit or a lighter review track, get the condo association's alteration agreement and insurance requirements before mobilizing, schedule required inspections.
On the Chicago condo, that checklist is what prevents the classic scene: the contractor stopped at the loading dock because the building manager never received the certificate of insurance, and the client calling you at 7:45 a.m.
A real example: a 12-week schedule in one minute
I asked ChatGPT for a high-level schedule for the condo, giving it: 1,050 sq ft, new flooring throughout, kitchen opened to the living room by removing a non-bearing partition (confirmed by the structural engineer), primary bath gutted and rebuilt, custom millwork in the kitchen and two bedrooms, full repaint. Client target: 12 weeks of construction. Output, unedited:
Week 0 (before construction): building permit issued, condo alteration agreement signed and contractor's certificate of insurance on file with management, final field measure for millwork, order tile and fixtures with lead times over 3 weeks.
Weeks 1 and 2: protect common areas and elevator, demo the kitchen partition, remove flooring and bathroom finishes, debris removal during the building's permitted hours.
Weeks 3 and 4: rough plumbing and electrical in kitchen and bath, rough-in inspections, waterproofing membrane in the shower with a flood test.
Weeks 5 and 6: drywall and patching, tile in bath and kitchen backsplash prep.
Weeks 7 and 8: flooring throughout, first coat of paint.
Weeks 9 and 10: cabinets and millwork, countertops template and install, plumbing fixtures.
Week 11: trim, lighting, final paint coat, appliance hookup.
Week 12: final inspections, cleaning, walkthrough with clients, punch list.
Schedule risks: countertops depend on cabinet install for templating; if the fabricator needs 3 weeks after template, week 10 becomes week 12. Consider templating from shop drawings only if the fabricator agrees.
Is it good? It is good as a draft. I would change two things. Custom millwork in eight weeks from order is optimistic for most shops right now, and it left out the client's sample approvals, which always slip. That is the pattern: AI gets the logical sequence right and gets your subcontractors wrong. You fix in two minutes what would take half an hour to build.
AI tools for architects and what they cost
| Tool | What I would use it for | Price (late 2026) |
|---|---|---|
| ChatGPT (OpenAI) | Programming notes, specs, proposals, mood-board images, photo edits of existing conditions | Free with limits; Plus US$ 20/month |
| Claude (Anthropic) | Long technical text: project manuals, spec sections, proposal language | Free with limits; Pro US$ 20/month |
| Gemini (Google) | Transcribing memos, editing a room photo while keeping the space | Free with limits; Google AI Pro US$ 19.99/month |
| Midjourney | Mood boards and concept imagery | From US$ 10/month; Stealth mode only on higher plans |
| Veras (EvolveLAB) | Study renderings on your SketchUp, Revit or Rhino model | Subscription with a trial; check the site |
| LookX | Renderings from sketches and simple models, architecture-trained styles | Free plan with credits; paid plans, check the site |
| D5 Render | Final renderings from the 3D model, AI tools for atmosphere | Community version free; Pro by subscription, check the site |
| SketchUp Diffusion | Quick variations inside SketchUp | Included with paid SketchUp plans |
Starting at zero dollars, I would use free ChatGPT for text, the free LookX credits for studies and D5 Community for final images. With budget for a single subscription, I would take ChatGPT Plus, because it covers text, transcription and images in one place. If your days are mostly specs and proposals rather than images, swap it for Claude Pro.
Where AI gets architecture wrong
- Zoning and local amendments. Ask for a setback, a FAR or a height limit on a specific lot and it answers confidently, often wrong or from another city. Read the zoning ordinance and the city's code amendments at the source, every time.
- Building code. It mixes IBC and IRC, cites sections that moved between editions and does not know which edition your jurisdiction adopted. Use it to remember a topic exists (egress, accessibility, energy), never as the final citation.
- Dimensions in images. A 24-inch-deep island, a 28-inch bathroom door, a stair with 9-inch risers. The image convinces and it is wrong.
- Structure. "Can we take this wall out?" is not a chatbot question. It is a question for a structural engineer and, in a condo, for the association.
- Quantities. It can lay out the spreadsheet, but it does not reliably compute tile area from a description. Take quantities from your model.
Common mistakes when a firm starts using AI
- Sending a generic AI image before the layout is settled. The client falls in love with the picture and the project becomes a chase after it.
- Uploading a plan with the client's name, unit number and address to any tool. Strip identifying data first and turn off training on your conversations. Several states now have consumer privacy laws, and the AIA Code of Ethics already expects you to protect client information.
- Asking for "a spec section" without your template. You get textbook filler. Paste one of your own sections as the model.
- Letting AI set the fee. It does not know your overhead or your city.
- Thinking you need fifteen tools. Two used well cover most of the lost hours.
Where to start this week
Pick one task you dislike that repeats every month. For most architects I know it is the spec or the proposal. Do it three times with AI, adjusting the prompt each time, and save the version that worked. By the fourth time it is routine. Then move to programming, and leave renderings for last, because they take the most testing.
If you would rather skip building prompt by prompt, you have two ways out: grab the pack of tested prompts linked on this page, or wait for the ready-made agents, which already handle the questionnaire, the proposal draft and the spec draft in your format without you writing anything.
Want to skip this part?
Ready-made AI agents for architects: nothing to build, no prompts to write, they work for you from the first inquiry to the spec.
- Takes the inquiry and returns an organized program, with the questions still missing
- Drafts the fee proposal using your own pricing and timeline rules
- Writes the outline spec from your finish schedule and tags what you need to verify
Frequently asked questions
What is the best AI for architects?
It depends on the phase. For text (programming, specs, proposals), ChatGPT or Claude. For study renderings that respect your geometry, Veras or LookX on top of the model. For final images, your rendering engine (D5, for example) with the full 3D model.
Can AI make a rendering from a sketch?
Yes. LookX, Veras and the AI features in engines like D5 generate images from a sketch or a simple model. The result is good for studies; for the final image, check dimensions and openings, because AI tends to change details.
Can I use ChatGPT to write specifications?
Use it for the first draft. Give it your room list and products, forbid any standard or brand you did not supply and ask it to tag inferred items with [VERIFY]. Review and responsibility stay with you.
Will AI replace architects?
No. It removes hours of repetitive work, but it does not take professional responsibility, cannot seal drawings, does not reliably know your local zoning and gets dimensions wrong in images.
How much does AI cost for a small architecture firm?
You can start at zero with the free tiers of ChatGPT, LookX and D5 Community. With one subscription, ChatGPT Plus at US$ 20 a month covers text, transcription and images.