Where to start governing ChatGPT, Copilot and Claude
Start by finding the AI staff already use, move work to business accounts, then watch before you block. A 30-day plan and checklist for MSPs and IT.

If your client uses ChatGPT, Microsoft Copilot or Claude for everyday work rather than building its own agents, start by finding out what staff already use. Then move that work onto business accounts with the data controls on, write a one-page rule, and watch for two weeks before you block anything.
That order matters because the tools are already in the building. In Microsoft and LinkedIn's 2024 Work Trend Index, 78% of AI users said they bring their own AI tools to work, rising to 80% at small and medium-sized companies. A policy written before you know what is in use describes a company that does not exist. For the wider picture of what governance covers and which rules apply, see the guide to choosing an AI governance framework.
Who this is for and what you get
This playbook is for an MSP (managed service provider) owner or IT lead whose client uses chat AI: ChatGPT in the browser, Copilot in Microsoft 365, Claude, and perhaps a few connectors to email, files or a CRM. Nobody at the client is writing agent code. If they are, the starting point is different.
After 30 days you will have:
- a list of the AI tools in use, who uses each one, and whether it is a personal or work account;
- a business plan for each tool you decided to keep, with staff signed in through work accounts;
- a one-page AI rule that the client's sponsor has signed and staff have read;
- a decision on each connector: approved, restricted or declined;
- two weeks of real usage to base your first warnings and blocks on.
What you need before you start
- A sponsor at the client. One named person who will sign the rule and back it when someone pushes back.
- Admin access to the client's identity provider (Microsoft Entra ID or Google Workspace) and to its device management or RMM (remote monitoring and management) tool.
- A short list of data that must not leave. Client records, health or payment data, source code, contracts, HR files. Five lines is enough.
- A decision to tell staff. People will learn that AI use is being reviewed. It goes better when they hear it from the sponsor first.
The 30-day plan
Step 1: Find the AI already in use (days 1 to 7)
Ask the client which AI tools staff use, then check, because the answer is usually incomplete. The same Work Trend Index found that 52% of people who use AI at work are reluctant to admit using it for their most important tasks.
Look in three places:
- The identity provider. Check which AI apps people have signed into with their work account.
- The browser and the network. Check which AI sites company devices reach, and whether people sign in with a work or personal account.
- The laptops. Check for installed AI apps: the ChatGPT and Claude desktop apps, coding assistants, and local model runners such as Ollama.
For each tool, write down who uses it, which account type, and roughly what data goes into it.
Waxell Endpoints is one way to do this step on managed laptops. Endpoints finds the AI already running on employee laptops, using a curated catalog of 160+ AI services. It runs on macOS (Apple Silicon) and Windows, with no Linux agent and no mobile, so phones need a separate answer. Once the agent and browser extension are installed, browser AI conversations are captured and uploaded by default, with secrets redacted on the device first, and an admin can turn browser capture off. Tell staff that before you roll it out.
Done when: you have one list of the AI tools in use, with users, account type and the kind of data each one sees.
Step 2: Pick the tools you will support and move work onto business accounts (days 5 to 12)
The biggest early fix is usually the account type, not a block. Consumer and business plans handle your data differently, and the providers say so in their own words:
- OpenAI says that on its services for individuals (Free, Plus, Pro) it may use your content to train its models, and that users can opt out in Settings. It also says that by default it does not use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu or its API to improve its models.
- Anthropic says it trains new models using data from Claude Free, Pro and Max accounts when the model-training setting is on, and that this does not apply to services under its Commercial Terms, including the Claude Team and Enterprise plans.
- Microsoft says prompts, responses and data accessed through Microsoft Graph aren't used to train foundation models, including those used by Microsoft Copilot.
So for each tool the client keeps, buy the business plan, sign staff in with work accounts, and ask them to stop using personal accounts for client work. Most clients do not need three chat assistants. Pick one or two, and say plainly in the rule which ones are not supported.
Done when: every tool on the keep list has a business plan, staff sign in with work accounts, and the tools you will not support are named.
Step 3: Write a one-page rule (days 8 to 14)
Keep it to one page. Long policies do not get read, and this one has to be read. Cover five things:
- The approved tools, and the account to use for each.
- Data that never goes into any AI tool, using your short list from the start.
- Data that is allowed, but only in the approved business tools.
- Who to ask before connecting an AI tool to another app.
- That AI use is reviewed, by whom, and what happens when a rule is broken.
Have the sponsor sign it and send it with a short note in plain language. For a longer look at content rules versus tool bans, see the Samsung ChatGPT post.
Done when: the sponsor has signed the rule and staff have acknowledged it.
Step 4: Fix sharing before you switch on Copilot or connectors (days 10 to 20)
Assistants that search company data see what the user can see. Microsoft says Copilot "only surfaces organizational data to which individual users have at least view permissions," and tells customers to use the permission models in services such as SharePoint so the right people have access to the right content. Anthropic says Claude connectors inherit each person's permissions from the connected service, and that custom connectors using a shared credential reach whatever that credential can access.
In practice, old oversharing can turn into search results. Before you switch these features on:
- review folders shared with "anyone with the link" or with the whole company, starting with finance, HR and client files;
- check what any shared or service account behind a connector can reach;
- remove access for people who have left or changed roles.
Then decide connectors one at a time rather than turning them all on. On Claude Team and Enterprise plans, an Owner or Primary Owner has to enable a connector for the organization before members can use it, and owners can limit which actions a connected service can take, for example letting Claude read files but not edit them. OpenAI lists "You control which internal sources are connected" among its commitments for its business products.
For connectors you route through it, the Waxell MCP Gateway is one way to add a check on the tool calls themselves (MCP is the Model Context Protocol, which Claude's custom connectors can use). Admins can configure human approval holds, and new tenants get default deny rules for permanent deletes, bulk deletes and history rewrite. The Gateway governs the calls that go through it: agents holding direct upstream credentials, locally registered MCP servers and unconfigured clients bypass it.
Done when: sensitive folders are no longer shared company-wide, and each connector is approved, restricted to read-only, or declined, with a named owner.
Step 5: Watch, then warn, then block (days 14 to 30)
Treat this like a new firewall. You would not start blocking traffic before you had looked at what goes through it. For two weeks, collect data only: which tools, which accounts, which kinds of data.
Then turn on warnings for the patterns that matter to this client, such as a personal account on a company laptop or client data going into an unapproved tool. Move a warning to a block only after you have seen it fire and checked that it is not stopping legitimate work. Blocking first can push people onto phones and home laptops, where you can see even less.
Done when: a short list of warnings is live, any blocks are based on something you actually saw, and the client knows how to ask for an exception.
Day 30: report and set the rhythm
Give the client a one-page report: tools found, accounts moved, connectors decided, warnings that fired, and what you recommend next. Then set a monthly review, and add any new tool to the list before it goes on the keep list.
Common mistakes
- Writing the policy first. A policy that names the wrong tools is ignored on day one.
- Blocking ChatGPT on day one. Staff move to phones and personal laptops, and you lose sight of the use you were trying to manage.
- Leaving personal accounts in place to save the license cost. The consumer terms above are the reason to pay for the business plan.
- Switching on Copilot or connectors before fixing sharing. The assistant can surface what was already overshared.
- Turning on every connector at once. Decide them one at a time, and start read-only where the service allows it.
- Not telling staff. Monitoring that people discover by accident costs more trust than it saves.
The 30-day checklist
Copy this into the client's onboarding document.
- Sponsor named and briefed
- Admin access to identity provider and device management or RMM
- Short list of data that must not leave
- AI tools found: identity provider, browser and network, laptops
- Each tool listed with users, account type and data it sees
- Keep list decided; unsupported tools named
- Business plans bought; staff signed in with work accounts
- Model-training settings confirmed for each business plan
- One-page rule signed by the sponsor and acknowledged by staff
- Company-wide and "anyone with the link" shares reviewed on sensitive folders
- Each connector approved, restricted or declined, with an owner
- Two weeks of usage collected before any block
- First warnings live; blocks only where a warning fired
- Day-30 report sent; monthly review booked
FAQ
No. Find out what is in use first, then write a one-page rule in the second week. A rule written from real usage names the right tools and gets followed.
Usually not at the start. Blocking before you know what people use it for can move that use onto devices you cannot see. Move work onto an approved business account first, then block specific patterns once you have seen them.
It is better than nothing, but it is not the same as a business plan. OpenAI says that even after an individual opts out, if they give feedback on a response, the entire conversation tied to that feedback may be used to train its models. Anthropic says that when a consumer user allows training, retention extends to five years; otherwise it stays at 30 days.
Either way, the business has no admin control over a personal account. A business plan gives you admin controls and keeps the work under the provider's commercial terms.
Check sharing. Microsoft says Copilot only surfaces organizational data that the user already has at least view permission for, so anything overshared in SharePoint can become easier to find. Review company-wide and "anyone with the link" shares on sensitive folders first.
A tool installed on laptops does not see what happens on a phone. Waxell Endpoints, for example, has no mobile agent. The practical answer is the account: ask staff to use the approved business account, signed in with their work identity, on any device they use for work. Put "work AI happens in the work account" in the one-page rule.
Sources
- Microsoft and LinkedIn, "AI at Work Is Here. Now Comes the Hard Part," 2024 Work Trend Index. Opened 2026-11-12.
- OpenAI Help Center, "How your data is used to improve model performance". Opened 2026-11-12.
- OpenAI, "Enterprise privacy at OpenAI". Opened 2026-11-12.
- Anthropic, "Updates to Consumer Terms and Privacy Policy," August 28, 2025. Opened 2026-11-12.
- Claude Help Center, "Use connectors to extend Claude's capabilities". Opened 2026-11-12.
- Microsoft Learn, "Data, Privacy, and Security for Microsoft Copilot". Opened 2026-11-12.
When you are ready to see the AI on your client's laptops in one place, book a 30-minute demo.