---
title: "An AI policy for a small architecture practice: what to write down"
url: https://notionforarchitecture.com/blog/ai-policy-small-architecture-practice
updated: 2026-10-30
author: Notion for Architecture
---

# An AI policy for a small architecture practice: what to write down

A small architecture practice's AI policy is a one-page document that names the approved tools, says what client and project data may be entered, names who reviews AI output before it leaves the office, states when AI use is disclosed to clients, and says where records are kept. Write decisions, not principles, and review it every quarter.

## Key takeaways

- A policy is a list of decisions: which tools, which data, who checks, what you tell clients, where you record it.
- Adoption is running ahead of governance: RIBA's 2026 report says few practices have formal AI policies.
- Ask your professional indemnity insurer and read your appointments before you finalise the wording.
- Map rules to work stages: low-risk drafting early, human sign-off wherever the output becomes a commitment.

## What is an AI policy for an architecture practice, and do you need one?

An AI policy is a short internal document that says how the people in your practice may use AI tools on client work. You need one as soon as anyone in the office, including you, pastes project information into a chat assistant, transcribes a meeting with an AI tool or generates images for a client. Without it, each person makes private decisions about confidentiality and checking, and you find out only after something goes wrong.

The need is current. According to the RIBA AI Report 2026, 74% of responding practices use AI, up from 59% the year before, and RIBA's Head of Research has noted that governance has not kept pace, with few practices holding formal AI policies. RIBA has also said it intends to publish an AI overlay to the Plan of Work and a responsible-use policy during 2026. Check riba.org for whether those have appeared before you finalise yours, and treat this page as organisational information, not legal or insurance advice.


## What should the policy decide? A six-row skeleton

Each row below is a decision you must make once and write down. The right-hand column gives the question to ask yourself, not a recommended answer, because the right answer depends on your contracts and your insurer.

| Decision | What to write | Question to settle it |
| --- | --- | --- |
| Approved tools | A named list with the plan or tier you pay for | Does the vendor state whether our inputs are used for training, and can we switch that off? |
| Data classes | Green (public), amber (internal), red (client confidential) | What may go into which tool? Which clients' appointments restrict it? |
| Review rule | Named person checks anything that leaves the office | Who signs off text, images, summaries and calculations? |
| Never-delegate list | Code interpretation, structural or fire conclusions, contract terms | What would we be unable to defend if the output were wrong? |
| Disclosure | When and how clients are told | Does the appointment or client brief say anything about AI? |
| Records | Where prompts, outputs and approvals are kept for key decisions | If challenged in two years, can we show who checked what? |


## How do you classify data before it goes into a tool?

The most useful single device is a three-colour data rule. It takes ten minutes to write and removes most judgement calls.

- Green: public information such as published planning documents, product literature and your own marketing copy. Any approved tool.
- Amber: internal but not client-identifiable, such as a generic specification clause with names removed or a fictional example. Approved tools on a business plan only.
- Red: anything identifying a client, a site address, unreleased design information, fee data or third-party drawings you hold under a confidentiality term. Not entered into a general AI tool unless the appointment allows it and the tool's data terms have been read.

Revisit the rule whenever a vendor changes its terms. Hosted AI tools change data and retention settings often, and the setting that protected you last year may not exist now.


## Where does AI sit across the design stages?

Stage names differ (RIBA Plan of Work 2020 in the UK, AIA phases in the US), but the risk logic is the same: the closer the output is to a commitment, the more review it needs.

| Stage group | Reasonable uses | Review needed |
| --- | --- | --- |
| Briefing and feasibility (RIBA 0-1, AIA pre-design) | Structuring a brief, question lists, summarising public documents | Light: confirm facts against the source |
| Concept and developed design (RIBA 2-3, AIA SD/DD) | Early image ideas, option wording, meeting summaries | Medium: label images as exploratory; check rights |
| Technical design and tender (RIBA 4, AIA CD/bidding) | Drafting standard text, checking a schedule for inconsistencies | High: a named person reads every line that enters a document |
| Construction and handover (RIBA 5-7, AIA CA) | Minutes, action lists, snag-list tidying | Medium: check actions and instructions before issue |

- Track AIA and RIBA stages in Notion: https://notionforarchitecture.com/blog/design-phases-notion-aia-riba-stages
- What AI tools actually work for architects: https://notionforarchitecture.com/blog/ai-tools-architects-what-works-2026


## What about client disclosure and professional indemnity?

Disclosure first. Read each appointment and client brief for any term on technology, confidentiality or intellectual property, and decide a default: for example, tell clients in the appointment letter that AI-assisted tools may be used for drafting and note-taking under human review. Whether you must disclose is a contract and professional-conduct question, so check your contract and your professional body's guidance.

Insurance second. The RIBA report coverage mentions architects' worry about whether professional indemnity insurance responds to AI-related errors. Do not assume. Ask your broker in writing whether AI-assisted work changes your cover or your proposal-form answers, and file the reply with the policy.


## How do you roll it out in a month?

1. Week 1: inventory: Ask everyone which AI tools they already use, including on personal accounts. List them without blame.
2. Week 2: decide: Fill the six-row table above, choose the approved list and the data colours, and name the reviewer.
3. Week 3: tell people: One page, one meeting. Show a worked red, amber and green example from your own work.
4. Week 4: record and review: Start the register (below). Diary a quarterly review and a check when any approved tool changes its terms.


## Where do the policy and the AI register live in Notion?

We make ArchitectOS, a Notion template, so weigh this accordingly. It does not contain an AI policy or an AI register, and it is not a compliance system. What it does have that a policy can lean on is a Template Vault for standing documents, a Pin Board for rules the team should see daily, and an Approvals section for recording sign-offs against a project.

A practical set-up: pin the one-page policy, keep the approved-tool list as a small database (tool, plan, data colours allowed, owner, last terms check), and note any AI-assisted deliverable and its reviewer in the project's approvals or decisions log. Last reviewed: 30 October 2026.

- See what's inside ArchitectOS: https://notionforarchitecture.com/#templates
- ArchitectOS pricing: https://notionforarchitecture.com/#pricing
- Decisions log in Notion: https://notionforarchitecture.com/blog/meeting-minutes-decisions-log-notion-architects


## FAQ

### Does an architecture practice need a written AI policy?

If anyone uses AI on client work, yes. It need only be one page. It records which tools are approved, what data may be entered, who reviews output and what clients are told, so decisions are consistent and defensible if challenged.

### Can I put client drawings or names into ChatGPT or similar tools?

Only if your appointment allows it and you have read the tool's data and retention terms for your plan. A safe default is to treat client-identifying and confidential information as off limits to general tools until you have checked.

### Do I have to tell clients I use AI?

It depends on your contract, the client's brief and your professional body's guidance. Many practices add a short line to their appointment letter. Check your own terms and, where unsure, take advice before relying on a default.

### Will professional indemnity insurance cover AI-related errors?

Do not assume. Policies differ. Ask your broker in writing whether AI-assisted work affects your cover or proposal-form answers, and keep the reply on file.

### Has RIBA published an AI policy or Plan of Work overlay?

RIBA said it intended to develop an AI overlay to the Plan of Work and a responsible-use policy during 2026. Check riba.org for the current position before you cite either in your own documents.

## Sources

- [RIBA AI Report 2026](https://www.riba.org/work/insights-and-resources/ai-report/)
- [RIBA news: AI adoption reaches tipping point](https://www.riba.org/news/ai-adoption-reaches-tipping-point-across-uk-architecture/)
- [RIBA artificial intelligence resources](https://www.riba.org/work/insights-and-resources/position-your-practice-for-a-fast-changing-world/artificial-intelligence-resources/)