Free AI Tools for Architecture: Concepts, Admin and Where Free Stops

Free ChatGPT, Claude, Gemini and Copilot Chat can handle concept text, image ideas, meeting minutes and first-pass document summaries for a small practice. They stop at upload caps (ChatGPT Free allows 3 file uploads a day), image resolution, usage resets every few hours and, above all, data terms: free consumer plans can use or review what you paste.

The cap that bites first is rarely a paywall. A client's brief, a fee letter and a consultant's report are all confidential, while free consumer plans were written for individuals, not for practices that hold other people's projects. Paying for an individual plan doesn't change that default either: OpenAI's help centre says data sharing is switched on by default for Free, Plus and Pro personal workspaces. The useful question for a practice is which jobs can live on a free tier safely, and where the limits make them awkward.

Follow me on Instagram@sagnikteaches

What a free plan really caps in a practice

Vendors publish few hard numbers for free tiers, and the ones they do publish move, so every figure below carries the date I read the page: 1 November 2026. Where a vendor gives no number, the table says so rather than guessing.

Connect on LinkedInSagnik Bhattacharya

Two details in that table matter more than the rest. First, Anthropic's upload page says Claude reads charts and images inside a PDF only up to 100 pages; for PDFs of 101 to 1,000 pages it processes the text alone. A long drawing set or specification can therefore lose its drawings. Second, ChatGPT Free's three uploads a day are a document budget rather than a chat limit: they cover one planning bundle, one structural report and one drawing set, and after that you are pasting excerpts by hand until tomorrow.

Subscribe on YouTube@sagnikteaches

Gemini sits in the middle on documents (10 files per prompt, 100 MB each), but Google's help page doesn't state the free-tier context size. It says only that paid plans get a larger 1 million token window. For a long specification that is the number to test on your own file rather than assume.

Concept stage: free tools for briefs, options and image ideas

At concept stage the best free use is words, not pictures. A chatbot is good at turning a rambling client email into a structured list of requirements, assumptions and questions, which is the first half hour of every project. The second best use is arguing layout strategies in plain text, a courtyard plan against a linear plan for instance, so the team reaches the sketch table with positions rather than blank paper.

Here is a brief digest as it might run, with the identifying details already removed. The prompt goes in as written.

You are helping an architect prepare for a first meeting. Below is an
anonymised client brief.
1. List the requirements that are stated.
2. List the assumptions I would need to confirm.
3. Write five questions for the first meeting.
Do not invent facts. If something is not stated, write "not stated".
Brief: Client A owns a detached house from the 1930s and wants a
single-storey rear extension for a kitchen and dining area with level
access to the garden. A budget band has been discussed. A neighbour
shares a boundary wall on one side. Client A works from home two days
a week.

What comes back, written out here as an illustration rather than a recorded output, lists the stated items (single-storey rear extension, kitchen and dining area, level garden access, shared boundary wall, home working two days a week) and marks size, budget figure and programme as "not stated". Its assumptions to confirm would include whether the rear wall can be opened up, where the drains run, and whether the extension needs planning consent. Its questions would cover how the household uses the space, what the client wants to keep, and how the neighbour has reacted so far.

What you fix is the gap-filling. If the reply quotes a budget number that isn't in your text, the tool invented it. If it states that no consent is needed, delete the sentence, because that is a legal conclusion drawn from nothing. The digest is a checklist for you to complete, never a briefing note to send.

Image tools and the licence line

Free image generation exists in ChatGPT and Gemini and in dedicated tools. For a practice the question is less about quality than about what you may do with the result, so here is what each vendor says on its own pages.

• Leonardo.Ai. Its help centre says free users get 150 tokens a day, resetting every 24 hours. Ownership and commercial-use rights differ between its free and paid plans, and its pricing page wasn't readable when this was checked, so read the current terms on Leonardo's own site before a free-tier image goes anywhere near client work.
• Adobe Firefly. The free plan offers limited daily generations that refresh each day. Adobe calls its own Firefly models commercially safe, says it is the creator's responsibility to decide whether partner-model outputs suit a project, and offers contractual IP indemnification only to enterprise customers who buy that entitlement.
• Stability AI Community License. Free for organisations under $1M in annual revenue, counted regardless of the source of that revenue, with commercial use allowed and outputs owned by you. Above $1M you need an enterprise licence. Running the models yourself typically needs your own hardware and setup time, so this one is free in licence terms but not in effort.
• Gemini and ChatGPT. Gemini's image help page gives no ownership or commercial-use terms and tells you to use your discretion before you publish. OpenAI says its ad-free option for Free accounts comes with lower usage limits and no access to some tools, such as image generation.

Microsoft's Copilot documentation states a point worth applying to every tool here. Microsoft doesn't claim ownership of Copilot output, but says it makes no determination on whether a customer's output is protected by copyright, because similar prompts from different customers can produce similar results. Treat concept imagery the same way: useful for internal mood and massing conversations, and a poor basis for anything you want to claim as the practice's own design work.

Two habits keep free image tools harmless. Feed them only your own massing sketches or public-domain references, never a client's site photographs, a neighbour's windows or a drawing with a title block. And keep a note of which tool made which image, so you can answer the question a client's lawyer will eventually ask. The longer comparison of paid image tools sits in AI for architecture and interior design concept images, by cost.

Admin that fits free limits: minutes, RFIs and fee letters

Meeting minutes without an AI note-taker

Free note-taker plans are built to be tasted, not lived on. Otter's free Basic plan gives 300 transcription minutes a month, stops each conversation at 30 minutes, allows 3 lifetime imports and keeps the 25 most recent conversations. A practice that holds a 60-minute design review every week and two 45-minute client calls works out like this: 60 plus 90 is 150 minutes a week, about 650 minutes across a month. The free plan covers less than half of that, and not one of those meetings fits under the 30-minute cap.

The free route that does work is your own typed notes. Take bullets during the meeting, paste them into a chat afterwards, and ask for an action table. A rough set of site-meeting notes, with role labels in place of names, shows the shape.

Notes: site mtg wed. M&E says duct clash at grid C, wants 150 drop.
We said no, min ceiling 2.4. Agreed to look at moving riser.
Contractor to confirm window lead time by fri. Client wants
programme update.
Task: write minutes as a table with columns Item, Decision or
note, Owner, Due. Use "date not given" where the notes give no date.
Do not add anything that is not in the notes.

The reply should give four rows: the duct clash (mechanical engineer asked for a 150 mm drop, practice declined because the minimum ceiling height is 2.4 m, riser to be reviewed), window lead time (contractor, "by Friday" as noted), the programme update (client request, owner not given) and a decision-log entry for the declined drop. Fix these before issuing: owners the notes never named, and any date the tool converts from "fri" into a calendar date you didn't give it. The instruction "date not given" is there to stop that, but it isn't foolproof, so check every row against your own memory before issuing. The fuller routine is in recording, summarising and issuing architect meeting minutes the same day.

RFI logs, fee proposal skeletons and chasing emails

Three more admin jobs fit the free tier because the input is small and mostly generic.

• RFI chasing. Paste last month's RFI log with project names removed and ask for every item older than ten working days, grouped by the party who must respond, with a one-line chaser for each. The log is your data, so the tool only sorts and words it.
• Fee proposal skeleton. Take your last proposal, strip the sums and the client's name, and ask for a scope and exclusions list in the same order as your template. The principal sets the fee; the tool sets the headings. Compare the exclusions line by line with your standard appointment terms, because a plausible-looking exclusion that you never agreed is worse than a missing one.
• Polite pushback. A contractor asks for an extension of time or a client asks for "just one more option". Give the tool the facts in three bullets and ask for two versions, firm and friendly. You choose the tone and you send it.

One job to keep away from free chat: calculating areas, quantities or fees. A language model can produce a plausible-looking total without adding the numbers, so any sum it gives needs a spreadsheet to confirm. For a deeper walk through the paperwork side, how architects use AI for fee proposals, briefs and paperwork covers it on paid tools.

Document limits: what a free plan can read from a project folder

Take an illustrative house-extension job with four large documents: a 38-page planning statement draft, a 62-page structural report, a 240-page specification and a 30-sheet drawing set exported as one PDF. The question is which free tool can take each one, and what you lose on the way.

Count the uploads. Sending all four documents to ChatGPT Free takes four uploads against an allowance of three a day, and the specification is the one most likely to be left out. On Claude the same job is six files in one chat after splitting the specification, well inside the limit of 20 files per chat, although usage still resets on a five-hour clock and a long specification can use a lot of it. Gemini would take all six in one prompt at 10 files, with the open question of free-tier context on the longest.

OpenAI's file-upload FAQ confirms that ChatGPT can analyse text and images in a PDF attached directly to a chat, but notes that embedded images may not be included when a PDF is saved as project knowledge. It doesn't address scanned PDFs, which is a gap for old measured surveys and signed-off drawings that were printed and rescanned. If a scan matters, test it on a page with a known answer before relying on anything it says about the rest.

A verification routine keeps these tools honest on drawings. Ask for every figure with its sheet or page number. Pick three at random and check them against the sheet. If one is wrong, stop using the output for numbers and use it only to find where to look. Models and file readers change often, so the habit matters more than any single result, and it costs about ten minutes per document.

A note on formats: export to PDF before uploading anything. Anthropic says that for documents other than PDFs, such as Word files, Claude extracts the text only and can't read embedded images, so a Word report full of annotated photographs loses them. A free chat is not a CAD or BIM viewer, so test a copy of any native file on non-confidential content before you build a routine around it. Google's help page says the Gemini app supports "most file types" without listing them, which is a reason to test, not an assurance.

Client data and licences: the checks before anything is pasted

Three checks decide whether a document goes into a free tool. Run them in order, because the first failure ends the conversation.

Check one: the data terms on the account you are logged into

The table above shows the pattern. ChatGPT Free, Claude Free and Gemini Free are consumer services. Each lets you restrict training or retention in settings, and each keeps some data for safety or legal reasons whatever you choose. OpenAI says a deleted chat is scheduled for permanent deletion within 30 days unless it must be kept longer for security or legal obligations. Anthropic says chats flagged under its usage policy can be kept for up to 2 years. Google says reviewed Gemini chats are kept up to three years and aren't removed when you delete your activity. A setting is a preference, not a contract.

Ads are the other consumer-plan difference. OpenAI says ads may appear on Free and Go, that advertisers receive only aggregated, non-identifying performance information, that Temporary Chats show no ads, and that ads aren't eligible to appear near sensitive topics. Plus, Pro, Business, Enterprise and Edu accounts have none. For an architect the practical point is that a chat window with ads in it is not a workplace tool, and nobody in the practice should be screen-sharing one in front of a client.

Check two: what your contracts and the client's terms allow

The legal position that matters most isn't the vendor's. It's your appointment document and any non-disclosure agreement. Some clients, such as developers, public bodies and owners of sensitive buildings, restrict disclosure to third-party processors. If the contract says nothing, assume an expectation of confidentiality still applies, treat a consumer chatbot as a third party, and ask your adviser or insurer if you are unsure. When in doubt, ask the client in writing before you paste, and keep the reply with the project file. Documents already public, such as a lodged planning application, carry less risk than a fee letter or a structural report, but the tidy rule is to test the document against the client's terms, not its public status.

Check three: who can use the output, and how

The licence notes in the concept section apply here too. The test for any output is whether you could explain to a client where it came from and who may use it. Write that into the appointment terms as a short clause: whether AI tools are used for concept work, whether outputs are original to the practice, and who owns what.

For a deeper treatment of copyright and client-data exposure across a whole practice, see copyright and client-data risks of AI in architecture practices. The rest of this section is the short version a free-tier user needs.

Anonymising a brief before it goes in

Anonymising works when you remove what makes the project recognisable, not just the name. The replacement rules I'd use:

• People become labels: Client A, neighbour B, contractor C, structural engineer.
• Addresses become a building type and an era: a detached 1930s house.
• Sums become bands or percentages, or disappear.
• Dates become seasons or "week 3".
• Company names become roles.
• Photographs and drawings are cropped or left out.

Then run the re-identification test: could a neighbour, a planning officer or a competitor guess which project this paragraph describes? A gatehouse with a shared well and one distinctive roof defect is identifiable with no name in it. If the answer is yes, cut a detail or keep the task off the free tool.

A filled-in traffic-light list for a house extension job

Practices that write the rule down once stop having the argument every week. This is the list as it might look for one small house-extension project. Adapt the colours to your own contracts.

Notice what the list does to the free tier: it shrinks to public material and anonymised thinking, which is exactly where free tools are strongest anyway. The red rows aren't an argument for paying for a chatbot. They are an argument for using Copilot Chat with a work sign-in, if you have Microsoft 365, or for keeping that material out of AI altogether until your policy catches up. The policy itself is the subject of an AI policy for architects, with ready-to-use clauses.

What staying free costs a five-person practice

Free has a cost even when the invoice is zero. The cost is time spent splitting files, re-explaining context in every new chat and checking that nobody pasted a red-row item. A quick sum shows where it overtakes a paid seat. Suppose your internal cost rate is $40 an hour and one person loses 20 minutes a week to those workarounds. Twenty minutes across 4.33 weeks is about 87 minutes a month, or roughly $58. A seat at $20 a month breaks even at 30 minutes a month. The workaround time is nearly three times that, before counting any risk.

The paid options below are list prices from the vendors' own pages on 1 November 2026.

If your practice is already on Microsoft 365 or Google Workspace, the first move costs nothing: use the work sign-in for Copilot Chat or the Gemini app and keep the consumer accounts for public material. If it isn't, two paid seats are the smallest step, $40 a month for ChatGPT Business or Claude Team at the annual price. A one-person practice has to buy two seats or stay on an individual plan with training switched off and the same traffic-light rule.

Which assistant suits an architect's documents better is a separate question, covered in Claude, ChatGPT and Gemini compared for architects.

Five signs a practice has outgrown free

Don't pay on a hunch. Pay when one of these is true for a month, because each has a number you can check.

1. Red-row material is in prompts. If anyone pastes client names, budgets or reports more than occasionally, the problem is the plan, not the person.
2. You hit the cap twice a week. Three ChatGPT uploads, or a five-hour Claude reset that lands on a Friday deadline, is a paid-seat problem.
3. Two people keep rewriting the same instructions. Shared Projects belong on a business plan; a shared personal login is a worse answer and breaks the terms of some services.
4. A client or insurer asks about your AI use. A written answer that names the plan and its data terms is far easier with a business agreement behind it.
5. The routine is stable. If the same three jobs run every week, the saved minutes are measurable and the seat cost is easy to justify.

The practical starting kit for a three-person practice is modest: the work-account Copilot Chat or Gemini app for anything confidential, a consumer free plan with training off for public and anonymised thinking, a one-page traffic-light list pinned where everyone works, and a monthly ten-minute check that nobody has drifted into the red rows. If you'd rather map the jobs and the rule with someone, that is what my AI implementation consultation covers.

Free AI for architects: practice questions

Can free AI check a design against building regulations?

Not reliably. A chatbot can point you to the section of a code you should read, but it can quote a superseded edition or blend requirements from different places with equal confidence. Ask for the clause reference, open the current edition yourself and read it. Never put a compliance statement in a drawing note or report on the strength of a chat answer alone.

Can I record a client meeting with a free note-taker?

Recording needs the consent of everyone on the call, and some clients forbid third-party recording tools in their terms, so ask first and keep the answer in the project file. Free plans also cap each conversation (Otter Basic stops at 30 minutes). For most practices, typed notes pasted into a chat afterwards are simpler and easier to defend.

Can staff use their personal AI accounts for practice work?

Treat it as a policy decision rather than a habit. Work done in a personal account sits in a service the practice cannot audit or delete from, and it leaves with the person. Decide which jobs may use a personal free account (public and anonymised material), which need a work account, and write the split down where everyone can see it.

Is Copilot Chat really free for a practice?

Microsoft says Copilot Chat is included at no extra cost with eligible Microsoft 365 plans, and its documentation says enterprise data protection applies to work and school accounts. It answers from the web plus any file you upload with the prompt. Copilot working inside Word, Excel and PowerPoint needs the paid Copilot licence, and Microsoft shows a label in each app saying which experience you have, so read the label before assuming.

Further reads

• AI Renders vs Outsourced Visualisation: What Practices Save — What rendering costs against outsourcing it.
• Is ChatGPT Plus Worth It for an Architect? A Practice Cost Check — A cost check on the first paid plan most practices try.
• AI Software for Architectural Drawings: Notes, Schedules and Sheet Checks — Notes, schedules and sheet checks on drawing sets.
• How Much Does AI Cost a Small Architecture Practice per Month? — A monthly budget for the AI tools of a small practice.
• Shadow AI: How to Stop Staff Pasting Client Data Into Free Tools — Stopping staff pasting client data into free tools.
• Free AI Tools vs Paid Plans: What a Small Business Should Pay For — The general free-versus-paid decision for any small firm.
• Free AI Architecture Design Software: Limits Before You Rely on It — The design-software side of free AI, with its limits.
• Top AI Tools for Architects: Eight Compared by Job and Price — Eight AI tools matched to five practice jobs, with USD entry prices, plan limits and a monthly cost for a five-person practice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help articles (data controls, file uploads FAQ, ads in ChatGPT), Anthropic pricing, privacy and upload help pages, Google Gemini Apps help and Workspace privacy pages, Microsoft Learn Copilot Chat pages and the Microsoft 365 Copilot pricing page, Otter pricing, Adobe Firefly plans, partner-model and business pages, the Leonardo.Ai help centre, the Stability AI licence page and Community License, Google's Gemini Notebook plans page, read on 1 November 2026 and checked again on 4 November 2026.