AI governance for AEC firms: data classes, owners and review steps
An AI policy sets which data each tool may read, who owns each shared agent, and what a person checks before AI output leaves the firm. This guide covers the five steps to write one and what it should contain.
What an AI policy covers
Firms usually start writing an AI policy after staff are already using AI, and a principal asks whether client drawings are going into ChatGPT. The policy that answers that question has five parts:
- Three data classes (open, internal and restricted) define how sensitive each kind of firm data is.
- A list of approved tools shows which tools may read each class.
- Review steps say what a person checks before AI output is sent, issued or published.
- A register of shared agents and custom GPTs names an owner and a backup for each one.
- One named person or group approves exceptions and new tools.
Firms without these parts tend to stall when someone asks to connect the project archive, because nobody can say what the AI would then be able to see. The stall points are covered in why AI implementation stalls.
Five steps to a firm AI policy
1. Take inventory
List every AI tool, license, workspace and custom GPT in use, including personal accounts. A short survey and conversations with 5 to 10 heavy users find most of them. Personal accounts belong on the list because they show which work staff already do with AI, and the shadow AI article covers how to run that survey.
2. Classify the data
Sort firm data into open, internal and restricted, by sector and by client contract. The client contract decides more than the sector does, so read the data and confidentiality terms in current contracts and mark the projects whose data cannot leave the firm's tenant.
3. Test access
Before connecting more sources, test on one sample site with IT. Check what each role can see through the AI, what the citations link to, and what leaves the firm when someone copies an answer into an email. Copilot answers can cite files and chats, and a pasted link may open for people who should not see it.
4. Set rules and owners
Write the firm-wide rules first, then let each business unit add rules for its own sources, reviewers and client restrictions. Every shared agent gets an owner and a backup, recorded in the register.
5. Verify and re-test
A named reviewer approves output before it is reused or published. Keep a fixed set of test questions with known answers for each shared workflow, and run it again whenever a tool or model changes.
Data classes
Three classes are enough for most firms. More classes make the policy harder to follow, and staff stop checking which class a file belongs to. In Microsoft 365, the classes map to sensitivity labels in Purview, and IT applies them to SharePoint sites and files so Copilot results follow them.
Firm-wide rules, unit rules and verified folders
The firm-wide policy is set once. Each business unit, such as marketing, HR, accounting or a sector studio, adds its own sources, reviewers and client restrictions. Each unit keeps a verified folder in SharePoint with a named owner, and agents read only what their permissions allow.
A governed workflow then runs in four steps: the agent reads the verified folder, drafts the output, a named reviewer approves or corrects it, and the approved result is saved back to the folder. The next run reads the approved version, so corrections build up in the folder rather than in someone's memory.
We write AI policies with firms as part of discovery and definition, including the data classes, the business-unit rules and the register of approved agents. To talk through where your firm stands, book a free AI strategy session.
Frequently asked questions
What should an AI policy for an architecture firm include?
The data classes the firm uses, the tools approved for each class, the review steps for AI output, a register of shared agents with an owner and a backup for each, and the person who approves exceptions. Most firm policies fit on two to four pages.
Can we put client project data into ChatGPT or Claude?
Check the client contract first. Enterprise plans of ChatGPT, Claude and Copilot do not train on your data by default, which suits internal data. Projects whose contracts restrict data use, such as some healthcare, government and secure facilities work, belong in the restricted class, and the policy names which tools, if any, may read them.
Do SharePoint permissions carry over to AI tools?
Copilot respects the permissions and sensitivity labels in your Microsoft 365 tenant. An outside tool that reads SharePoint through Microsoft Graph may not apply sensitivity labels the same way, so test one sample site with IT before connecting more sources.
Who should approve AI tools in a firm?
The AI policy names one person or a small group, usually leadership with IT and the design technology lead, who approves new tools and exceptions. Each business unit then sets its own sources and reviewers inside the firm-wide rules.
Get the next AEC tech analysis in your inbox
Original research on AEC tools, AI adoption, and industry data. Free, irregular cadence, easy unsubscribe.
Need help applying this to your firm?
We advise AEC firms on technology and AI strategy. Start with a free 30-minute AI strategy session: your workflows, the right tools, and quick wins you can act on.
Book a free AI strategy session