Before you write AI policy, find out what's connected
Author
Audrey
NobleCloak's AI Correspondent · AI-drafted, fact-checked against our sourced evidence before publishing.
Date Published
Most small compliance programs write their first AI policy the same way. Someone, often an outside consultant, interviews the department heads and asks which AI tools their people use. The answers go into a spreadsheet, and the spreadsheet sets the policy's scope.
The weak spot is the starting list. It holds what people remembered to say. Your identity provider already holds a better one.
This post compares two ways to build that list: consultant interviews, and a read-only scan of the apps connected to Google Workspace or Microsoft Entra. They work together. The scan's output is what you hand the consultant, or what you write the policy from yourself.
What interviews find
Interviews are good at the tools people chose on purpose: the licensed copilot, the chatbot the marketing team pays for, the vendor IT approved last year.
They miss three things.
- Apps someone connected in a minute. A meeting-notes app that asked for calendar access. A writing assistant that signed in with Google. Clicking "Allow" on a consent screen doesn't feel like buying software, so nobody mentions it.
- AI that arrived inside an existing tool. If the CRM added a summarizer, the person using it thinks of it as the CRM.
- What each tool can reach. Even when people name a tool, they rarely know whether it can read mail, files or calendars. That is the detail a policy decision turns on.
None of this is dishonesty. It is the limit of asking people to remember.
What a read-only scan finds
When someone signs in to an app with their work account and clicks "Allow", the grant is recorded. Both major identity providers let an administrator review those grants.
- Google Workspace lists "Apps that have accessed Google data", with the "Number of users accessing the app" and the OAuth scopes each app uses. It sits under Security, Access and data control, API controls, Manage App Access.
- Microsoft Entra shows the permissions granted to each enterprise application, with separate tabs for admin consent and user consent.
You can read these yourself. Our read-only scan reads Google Workspace and Entra and gives you every AI app your people authorized, the data scopes it holds, who's using it, since when.
A scan has limits too. It sees what SSO and OAuth can see. It does not capture personal accounts, personal devices or direct API access. Someone pasting member data into a chatbot on their own phone won't show up, and neither will an AI feature running inside a vendor's own systems. Interviews and a written acceptable-use rule still cover that part.
The scan's output is the input to policy
An AI policy is easy to write in general terms and hard to apply. "Staff may use approved AI tools" leaves every real question open.
A list of connected apps turns the policy into decisions you make one at a time. Take a single row: a meeting-notes app has held calendar and drive read access for eight people since March. You have three choices.
- Allowed. It fits how you work, the vendor passed your review, and the scopes are acceptable. Record it as approved.
- Needs approval. It is fine for some uses or some people. Name who signs off.
- Denied. The access is too broad, or the vendor won't answer your questions. Revoke the grant and record why.
Do that for every row and you have the core of the policy: a dated list of tools, a decision for each, and an owner. The general language around it can come from a template or from your consultant.
The same split shows up in the OCC's November framework for examiners. It says a system's purpose statement "may distinguish actions the system can take autonomously from actions that call for review, approval, abstention, or escalation." The framework is a resource for examiners. In its own words it "does not establish supervisory expectations for banks." It still shows the questions an examiner is equipped to ask. More in what the OCC's examiner framework means for AI agents, and on reading OAuth grants as your first policy draft.
Where the consultant still earns the fee
A good consultant brings judgment and continuity. They know your program, your regulator and your exam history. Keep them.
What changes is the brief you give them. "Find out what AI we use" sends them into interviews. A list of connected apps, with scopes and users, lets them spend the hours on what a list can't do:
- deciding which tools touch customer or member information and need full vendor diligence;
- writing the policy language and fitting it into your existing program;
- for RIAs and broker-dealers, tying it to Reg S-P. The amended rule at 17 CFR 248.30 calls for written policies and procedures "reasonably designed to require oversight, including through due diligence and monitoring, of service providers." An AI app with access to client information can meet the rule's definition of a service provider. See the Reg S-P AI vendor file.
The same holds if you are pricing a GRC platform instead of a consultant. A platform gives you an organized place to keep the inventory. Someone still has to fill it in, and the first column stays blank until someone reads what's connected.
A sequence for a one- or two-person program
- Pull the list of connected apps from Google Workspace or Entra, or have it pulled for you.
- Sort it by scope. Mail, files and calendars first: that is where customer information lives.
- Decide each tool: allowed, needs approval, or denied. Revoke what you deny.
- Write the policy around those decisions, with your consultant or from a template.
- Read the list again on a schedule. New grants appear between reviews. Point-in-time assessment vs continuous monitoring covers how often.
Where NobleCloak fits
Two places, and some things it doesn't cover.
Reading what's connected. Our read-only scan reads Google Workspace and Entra: every AI app your people authorized, the data scopes it holds, who's using it, since when. Coverage is limited to what SSO and OAuth can see. Our team runs the first report with you. Self-serve connect is coming soon.
Holding the decisions for agents. When your staff use AI agents through NobleCloak, policy marks each tool as allowed, needs approval, or denied. Approval goes to a named person or role, and one approval covers one action. The Record of Decisions keeps who asked, which agent, which tool, the decision, who approved, and when. See how human-in-the-loop works.
Coming soon: mapping scanned apps to tools you can connect, and drafting a starting policy from the scan. Today you make those decisions yourself, with our report in hand.
What it doesn't do: replace your consultant, review your vendors' contracts, or see AI used on personal accounts and devices.

