An architecture practice's AI policy should name the approved tools and accounts, say which client information must never enter them, treat AI-generated images as concept illustrations unless a person authored the final work, and require a named person to check anything AI touched before issue. One page is enough if every clause has an owner and a consequence.
A policy is not a ban. Most practices already use AI somewhere, often on personal accounts nobody approved, so its first job is to move that use onto accounts the practice controls. Its second is to make professional responsibility explicit: an AI tool never signs a drawing, so a person does.
Architects hold two kinds of material that a generic AI policy handles badly: clients' confidential information, often covered by appointment terms or a non-disclosure agreement, and documents issued with professional liability attached. The clauses below are written for those two things.
The text is written to be pasted into a document. The ten clauses run to about 670 words, which is roughly a page in a normal font, with Schedule 1 on the back. Square brackets mark decisions only your practice can make, and each clause carries a note on why it is there and how to check it is followed. Treat the wording as a draft for your own lawyer and insurer to read, not as legal advice.
The policy, clause by clause
1. Scope and owner
This policy applies to every use of generative AI on [Practice]'s work, including chat assistants, image generators, transcription and summary tools, and AI features built into design software. It applies to staff, directors, contractors and freelancers working for us. [Name] is the AI lead and owns this policy and the approved-tools list. A breach is handled under [our disciplinary procedure]; telling the AI lead promptly counts in the person's favour.
Why and how to check: naming design-software features closes the gap where nobody thinks of a built-in assistant as "AI". The check is simple: the AI lead's name is on the policy and on the staff noticeboard.
2. Approved tools and practice-managed logins
Only tools on the approved list may be used on client work. At the date of this policy the list is [tool and plan]. Each person uses their own practice-managed login, and logins are never shared. Free and personal accounts are not used for anything connected with a client or project. The AI lead adds a tool to the list only after reading the vendor's data and training terms.
Why and how to check: vendors forbid sharing credentials. Anthropic's consumer terms say you may not share your login information or credentials with anyone else, and OpenAI's terms say the same, so one login for the whole office breaches the terms as well as your own policy. Check by asking the AI lead for the list of seats and comparing it with the headcount.
3. Information bands
Information is green, amber or red. Green information may be used in any approved tool. Amber information may be used only in an approved business-plan tool, with the project lead's agreement, after names, addresses and other identifying details are removed. Red information never goes into any AI tool. The bands, with examples, are in Schedule 1.
Why and how to check: three bands are easy to remember and hard to argue with. Check by asking any team member to place three of their current documents in a band; hesitation tells you where the examples in Schedule 1 are too thin.
4. Permitted uses
Approved uses are: drafting and editing our own text; summarising public documents; producing concept imagery from our own sketches and models; transcribing and summarising our own meetings once attendees know; and checking our own spreadsheets and formulas. Any other use needs the AI lead's agreement first.
Why and how to check: a short permitted list is easier to keep than a long prohibited one. Recording rules differ from place to place, so the attendees-know condition is a minimum, not a legal opinion. Check by reading the AI log (clause 7) for uses outside the list.
5. Client confidentiality
We treat anything covered by a client's confidentiality terms or a non-disclosure agreement as red, unless the client has agreed a specific use in writing; the AI lead records that agreement and the band changes for that use only. We tell a client when we use AI tools on their project information if their appointment or brief asks us to, or if the information is amber. We do not claim work is AI-free or AI-made unless that is true.
Why and how to check: confidentiality terms usually bind the practice whatever tool it uses, and a tool is a third party. Check by reading new appointment letters and NDAs for restrictions before the project starts, and recording any restriction in the project file.
6. Copyright and ownership
AI-generated images and text are treated as illustrative. We do not describe them as our original design work and do not promise a client exclusive rights in them. Work we present as ours is authored by a person: drawn, modelled, composited or substantially edited, with the source files and edit history kept. We do not put another designer's name, or their work, into a prompt. We do not upload third-party copyright material, including consultants' drawings and manufacturers' CAD files, without permission. We use only image tools whose terms allow commercial use, and we assume no vendor indemnity unless our contract says so.
Why and how to check: this clause rests on the points in the copyright section below. Check by asking for the layered file and the original sketch behind any render that goes into a competition entry or a client presentation.
7. Checking before issue
Nothing AI has touched is issued until a named checker, who is not the person who prompted it, has verified every figure, reference, dimension, product claim and statement against its source and recorded their initials in the AI log. AI does not produce or alter content on drawings, specifications, schedules, calculations or compliance statements without that check. A person, not a tool, signs everything we issue.
Why and how to check: a second pair of eyes catches what the prompter has already stopped seeing. Check by sampling five issued documents a quarter and tracing each AI-touched item back to its log entry.
8. Telling clients, consultants and our insurer
Our appointment letters say that we may use AI tools on a business plan to draft and summarise, that we do not enter confidential information into tools that train on it, and that a person checks every output before it is issued. We tell our professional indemnity insurer about our AI use as the policy requires. Consultants and freelancers working for us follow equivalent rules, written into their appointment.
Why and how to check: clients and insurers dislike surprises more than they dislike AI. The insurer disclosure tutorial covers what to say and when. Check by confirming that the standard appointment letter now carries the sentence.
9. If something goes wrong
If red information goes into an AI tool, or amber information into the wrong one, the person tells the AI lead within the hour. We record what was entered, delete the chat, check the account's training setting, decide with our adviser whether the client must be told, and fix the cause. Reporting an honest mistake is never a reason for discipline; hiding one is.
Why and how to check: people hide mistakes when reporting is punished, so the last sentence does real work. Run one practice drill a year so the steps are familiar before a real incident.
10. Training and review
Everyone who uses AI tools on practice work completes our briefing before they start and once a year. The AI lead reviews this policy and the approved list every [quarter], and also when a tool's terms change, a client objects or an incident occurs.
Why and how to check: if you have clients in the EU, the EU AI Act asks organisations that use AI to take measures to support staff AI literacy (the wording was softened in July 2026), so a dated record of each briefing is useful as well as sensible. Check by looking at the briefing record.
| Clause | Owner | If it is broken |
|---|---|---|
| 1 Scope and owner | Director | AI lead position vacant; policy reviewed within a month |
| 2 Approved tools and logins | AI lead | Account closed; person re-briefed |
| 3 Information bands | Project lead | Treated as an incident under clause 9 |
| 4 Permitted uses | AI lead | Use stopped until agreed |
| 5 Client confidentiality | Project lead | Client informed if our adviser says so |
| 6 Copyright and ownership | Director | Image withdrawn from use |
| 7 Checking before issue | Project lead | Document recalled and rechecked |
| 8 Telling clients and insurer | Director | Letters corrected; insurer told |
| 9 If something goes wrong | AI lead | Disciplinary procedure for concealment only |
| 10 Training and review | AI lead | No AI use until the briefing is done |
Schedule 1: three information bands, with architects' examples
| Band | Examples | Rule |
|---|---|---|
| Green | Published planning policy; manufacturer literature; our own marketing copy; our own concept sketches with no client identifiers | Any approved tool |
| Amber | Client briefs with names removed; our own drawings with the address removed; meeting notes with attendee names removed; site photos with no people or vehicle plates | Approved business-plan tool, project lead agrees, identifiers removed |
| Red | Anything under a non-disclosure agreement, unless the client has agreed a specific use in writing; layouts of secure or security-sensitive buildings; personal data such as a homeowner's name with address and floor plan, or details of children at a school; client budgets and funding; consultants' drawings; competition entries before submission; anything a client has asked us not to share | Never |
Removing a name is not always enough. A description unique to a site, such as the only protected historic building on a named street, can identify a client without a name in sight, so the project lead decides whether an anonymised prompt still points at one person. Here is the same request as it might be written carelessly and as policy allows.
Careless: Summarise this email from [client name] about the rear extension at [address], budget [amount], and list their objections to the planning officer's comments.
Allowed: Summarise this email from a client about a rear extension to a two-storey house and list their objections to the planning comments. [Names, address and budget removed; project lead agreed; business-plan tool.]
The anonymising tutorial shows the removal step in detail. Neither version belongs in a free personal account.
Copyright in renders: what clause 6 rests on
A national copyright office published a report on AI outputs in January 2025. Its announcement says the outputs of generative AI can be protected by copyright only where a human author has determined sufficient expressive elements, that prompts alone do not establish authorship, and that protection can arise from a human-authored work perceptible in the output or from a human's creative arrangement or modification of it. It also says using AI to assist creation does not by itself bar protection. Other places may differ, so ask your adviser, but the direction explains clause 6: a render someone typed a sentence to produce is weak ground for claiming authorship, and a render built on your own model and then composited and overpainted by a person is stronger.
Vendor terms do not fill the gap. Anthropic's consumer terms assign to you whatever rights Anthropic has in the outputs, with the words "if any", and OpenAI's terms use the same formula. Indemnity (a vendor's promise to defend you if an output is challenged) varies by plan. OpenAI's output indemnity in its service terms covers API and enterprise-type customers rather than ChatGPT Business, and Adobe's commercial-safety claim covers its own Firefly models, not the partner models that appear in the same apps; check each vendor's current terms before relying on either. Clause 6 therefore assumes no indemnity.
Keep a provenance record for any image that matters. One row per image is enough.
| Field | Example entry |
|---|---|
| Image and project | Concept view 3, [project reference] |
| Source of the starting point | Massing model by [name], exported on [date] |
| Tool and plan | [Image tool], [plan], commercial use permitted |
| Human work after generation | Overpainted entrance and planting in [software]; layered file kept |
| Labelled as | "AI-assisted concept image" on the presentation slide |
| Checked by | [Initials], [date] |
What "checked" means for each kind of AI-touched item
Clause 7 says "verified against its source". The table turns that into a minimum for the items architects actually issue.
| AI-touched item | Minimum check | Checker |
|---|---|---|
| Specification clause | Product names, standards and performance values against the manufacturer's current data | Specifier |
| Area schedule or room data sheet | Totals recomputed in a spreadsheet and compared with the brief (see the room data sheets tutorial) | Project architect |
| Meeting minutes | Decisions and actions confirmed against the recording or notes; actions agreed with their owners | Meeting chair |
| Planning statement | Every policy reference and measurement traced to the source document | Project lead |
| Consultant letter or email | Names, dates, references and instructions | Project lead |
| Render caption or marketing text | Claims, credits and the "AI-assisted" label | Director |
For drawing sets, the drawing review tutorial goes further. The principle for all of them is the one in clause 7: the checker is never the person who wrote the prompt, because the prompter reads what they meant rather than what is there.
Five scenarios to test the policy on your team
Hand these out and ask staff to rule on each before you reveal the answer.
| Scenario | Ruling | Why |
|---|---|---|
| A director wants a 40-page client brief summarised before a meeting. | Amber, if allowed | Approved business-plan tool, identifiers removed, project lead agrees. Red if it carries personal data or falls under an NDA. |
| A job architect pastes a structural engineer's drawing PDF into a free chatbot to list clashes. | Breach | Free account, third-party copyright, no permission, and the output is unverified. |
| A freelance visualiser generates three façade options from a screenshot of our massing model. | Allowed | If bound by equivalent rules by contract, labelled as concept images, with the layered files kept. |
| Marketing wants an AI image of a "completed project" using a description of a real client's building. | Not allowed | It presents as built work, uses client information and misleads. Use real photography with consent. |
| A placement student drafts a planning letter on a personal account, with the site address in the prompt. | Breach | Personal account and an identifiable site. The approved tool is the route, and the letter is amber. |
What approved tools cost a seven-person practice
Moving people off personal accounts means paying for business seats. List prices, as checked in October 2026, are worth confirming on each vendor's page. A ChatGPT Business seat lists at $25 a month, or $20 billed annually, with a two-seat minimum, and a Claude Team Standard seat lists at the same figures with the same minimum. For seven people that is 7 x $25 = $175 a month on monthly billing, or 7 x $20 = $140 a month on annual billing, which is $1,680 a year. Microsoft 365 Copilot Business lists at $21 a user a month on annual billing on top of a Microsoft 365 business plan, so seven users come to 7 x $21 = $147 a month, or $1,764 a year, before the base licences. Microsoft has also offered a lower promotional price on annual plans until the end of 2026, so check what your own quote shows.
The reason to pay is in the terms rather than the features. Anthropic's privacy centre says that by default it does not use inputs or outputs from its commercial products to train its models, with an exception for content you explicitly report as feedback. OpenAI's pricing page says ChatGPT Business data is not used for training by default, and other vendors' business plans make similar commitments. Confirm the current wording, and keep the vendor terms page in the policy folder with the date you read it.
The AI log, with three filled rows
The log is what makes clauses 4, 7 and 10 auditable. It can be a shared spreadsheet with one row per use.
| Date | Project | Tool and plan | Use | Band | Checker | Outcome |
|---|---|---|---|---|---|---|
| 12 Oct | [P-2207] | [Assistant], business plan | Summarised public planning policy extracts | Green | [Initials] | Quoted clauses checked against source |
| 14 Oct | [P-2193] | [Assistant], business plan | Drafted client email from our own notes | Amber | [Initials] | Sent after edit |
| 17 Oct | [P-2193] | [Image tool], business plan | Three concept façade images from our massing sketch | Green | [Initials] | Labelled "AI-assisted concept" in the client deck |
Review the log at the quarterly meeting. Three questions are enough: which uses were outside the permitted list, which tools changed their terms, and which checks found errors. The answers decide whether the permitted list grows or shrinks. For a wider check of what you already do, the drawings and client data checklist works as an audit before you adopt the policy, and the copyright and client data risks tutorial gives the background.
Adopting the policy in a week
- Day 1, current use. Ask every person, anonymously if you prefer, which AI tools they use and for what. Expect surprises. Twenty minutes each.
- Day 2, tools. Choose the approved tools, read their data and training terms, and buy the seats.
- Day 3, drafting. Paste the ten clauses, fill every bracket and name the AI lead.
- Day 4, outside readers. Send the draft to your lawyer and your insurer's contact, and ask whether the appointment letter wording needs changing.
- Day 5, briefing. Run a 60-minute session using the five scenarios, have everyone sign, and start the log.
That is about two working days of a director's time and an hour and a half per person. The policy will be wrong in places, which is why clause 10 exists. A page the practice follows beats a perfect one in a drawer.
Questions practices ask when adopting an AI policy
Do we need a policy if only one person uses AI?
Yes, a short one. The risk comes from what goes into the tool, not how many people use it. One person pasting a client brief into a personal account is enough to breach a confidentiality clause. Start with clauses 2, 3 and 7 (accounts, information bands and checking before issue) and add the rest as use grows.
Does the policy cover AI features inside our design software?
Yes. Clause 1 names AI features built into tools you already use, and the approved list should include them with their data settings. Check what each feature sends to the vendor's servers and whether your plan lets you switch it off, because vendors add such features in routine updates.
What about consultants and freelancers?
Bind them by contract. Add a sentence to their appointment saying they follow rules equivalent to yours on client data, copyright and checking, and ask visualisers which image tools they use before you commission them. A freelancer's personal account is outside your controls, so the contract is the control.
Can a policy stop staff using AI on personal phones?
A ban alone rarely works. Give people an approved tool that is at least as convenient, make it safe to report mistakes, and never punish honest disclosure. Bans with no alternative push use out of sight, which is the outcome the policy exists to prevent.
Further reads
- Shadow AI: How to Stop Staff Pasting Client Data Into Free Tools — Stop staff pasting client data into free tools.
- How to Roll Out an AI Policy So Staff Actually Follow It — Get staff to follow the policy once it is written.
- Claude vs ChatGPT vs Gemini for Architects: Which to Pay For — Pick the assistant to put on the approved list.
- Best AI Rendering Tools for Small Architecture Practices (2026) — Choose image tools whose terms you can live with.
- AI Renders vs Outsourced Visualisation: What Practices Save — Weigh AI renders against outsourced visualisation.
- AI Acceptable Use Policy for a Small Professional Firm — A general acceptable-use policy to compare.
- AI Consultant for Architects: What a First Engagement Costs — Seven consultant days priced deliverable by deliverable for a ten-person practice: audit, policy, one workflow, team session and a 30-day check.
- AI Training for Architects: What a Practice Workshop Covers and Costs — A half-day AI workshop agenda for an architecture practice, modules by role, and the cost per person once staff time is counted, set against self-taught routes.
- 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.
- Is ChatGPT Plus Worth It for an Architect? A Practice Cost Check — Break-even minutes for a $20 plan, what Business really costs with its two-seat minimum, and a two-week log to prove the hours saved.
- Responsible Use of AI in Surveying Practice: A Small Firm's RICS Checklist — The RICS responsible-AI standard turned into a checklist for a six-person practice, with named owners, filled-in examples and a 90-day plan.
- AI Automation for Architects: Enquiry to Fee Invoice, Step by Step — Six draft-only automation steps from enquiry form to overdue-invoice reminder, with costs, task counts and the architect approving every document.
- AI Floor Plan Generation for Architects: Where It Fails a Real Brief — A test protocol for architects: write one real brief as a ledger, score a generated plan against it, and set the sign-off rules before clients see anything.
- AI Marketing for Architects: Case Studies, Website Copy and LinkedIn — Turn a site-visit voice note into a case study, service-page copy and LinkedIn posts that sound like your practice, with permission and claims checks.
- AI Tools for Architects 2026: A Toolkit From Brief to Handover — Which AI tool to use at concept, design, tender, site and handover, what each costs on its own pricing page, and three stacks for a five-person practice.
- AI Tools for Landscape Architects: Planting Plans, Renders, Proposals — A landscape practice's workflow: plant lists with a four-point check, schedule sums, honestly labelled concept renders and staged fee proposals.
- 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) and privacy centre page on commercial data (dated 18 August 2026); Claude and ChatGPT pricing pages and Adobe's Firefly product page (checked October 2026); a national copyright office's January 2025 announcement of its report on AI outputs; OpenAI's terms of use and service terms and Microsoft's Copilot Business pricing as summarised in October 2026.