An MSP should roll out an AI governance agent the same way it rolls out any agent from its RMM: one install script built once for every client, a short scope conversation with each client before anything is installed, a pilot on a few devices, a health check that alerts when the agent stops, and then the rest of the client's devices. The part that is different is the order. Watch first, decide with the client, and only then turn on any rule that blocks.
Our RMM deployment guide has the commands, the variables and the health check exit codes. This article is about the plan around them: which clients go first, what to agree before you deploy, what to watch in the first two weeks, and the mistakes that cost the most time.
The Short Answer
- Build the script once. One install script in the RMM, with the organization key as a secure variable and each client's ID in a per-site field.
- Agree the scope with each client first. What is recorded, who sees it, and how staff are told.
- Pilot on your own devices, then on two to five devices at one friendly client.
- Add the health check before the rollout, so a stopped agent raises an alert from day one.
- Roll out one client at a time and observe for about two weeks before enforcing anything.
- Review the findings with the client and decide each app together. Enforcement starts after that meeting, not before it.
Build It Once, Reuse It Everywhere
The time saver is in the RMM, not the agent. Keep the organization key in one secure or global variable, and keep each client's ID in a custom field on that client's site. One script then serves every client, and adding a client is filling in one field.
A few habits keep it that way:
- Name the script and the monitor the same way across your RMM, so any technician can find them.
- Never paste the organization key into a script body or a ticket. It is a secret, one per provider. The client ID is not.
- Write down where the key lives. If you rotate it, enrolled devices keep working, but the RMM script needs the new key.
Which Clients Go First
Start where a surprise costs least and teaches most:
- Your own shop. You learn the install, the console and the alerts on devices you own.
- One friendly client that already talks to you about AI. They will tell you what is confusing.
- Clients with the clearest need, such as those handling client records or regulated data, once the steps are routine.
- Everyone else, a few clients per week, so findings reviews do not stack up in the same week.
Plan the review meetings when you plan the installs. Every client you deploy creates a findings conversation about two weeks later.
Agree the Scope Before You Install
The agent inspects AI traffic on the device, so this conversation comes before the script runs, not after. Cover:
- What is recorded. AI use is counted, and prompt text is kept only when a data rule matches: none, a redacted excerpt (the default), or a longer excerpt for investigations.
- Who in your team and theirs can see the findings.
- How staff are told, and when. Have the client confirm with its own counsel what notice or consent applies to its staff. The Partner Agreement makes you responsible for making sure this happens.
- Who decides each app once the findings are in.
A silent install accepts the agent installation terms on your behalf as the provider. Before you deploy, the Partner Agreement requires a written agreement with the client that authorizes the install. Keep it with the client's file.
The First Two Weeks
New clients start in Passive mode: the agent records AI use and sensitive data matches but blocks, redacts and warns about nothing. Use that time well:
- Days 1 and 2: check coverage. Every device the client owns should appear on its Devices page. Compare the count to your RMM, and chase the devices that were off. The client's setup checklist shows All Devices Ready when they are in.
- Day 3 onward: watch the health check. Route its alert to your PSA so a stopped agent opens a ticket. A device that exits 1 or 2 is not protected. Fix those before the review, or the findings will have gaps.
- Week one: skim, do not act. Note which apps appear, on how many devices, and whether people used work or personal accounts.
- Day 14: build the findings for the review meeting.
The Mistakes That Cost the Most Time
| Mistake | What Happens | How to Avoid It |
|---|---|---|
| Hardcoding a client ID in the script | Devices join the wrong client | Keep the client ID in a per-site field and let the script read it |
| Skipping the pilot | A wrong ID or a policy surprise reaches a whole client | Two to five devices first, in Passive mode |
| No health check until later | A stopped agent goes unnoticed for weeks | Add the monitor before the rollout, and test it by stopping the agent once |
| Enforcing on day one | Staff hit blocks nobody explained, and tickets follow | Observe for two weeks, decide with the client, then enforce |
| Not telling staff | Trust problems that outlast the rollout | Agree the notice in the scope conversation |
| Forgetting devices that were off | Coverage looks complete when it is not | Compare the console's device count with the RMM's after a week |
When to Turn On Enforcement
Turn on rules when the client has seen its findings and agreed a decision for each app: approved, approved with work accounts only, or blocked. Switching a client from Passive to Active turns on the rules as configured, with no reinstall. Before you switch, set an uninstall code for the client (on Windows, re-running the install script over an existing install then needs it), and if the client has an MDM, push the proxy and trust settings so people without administrator rights cannot change them.
Tell staff what changes and when. A short note that names the approved tools does more than a list of what is blocked.
Questions MSPs Ask
How long does the whole rollout take for one client? Building the script takes under an hour, once. After that, the install runs from the RMM in minutes per device, and the client spends about two weeks in Passive mode before the review.
Should every client be in Active mode eventually? Not necessarily. Some clients want visibility and reports only. Passive mode is a valid place to stay if the client decides that.
What about Chromebooks? ChromeOS does not allow native agents, so Chromebooks run ProctoredAI as a managed browser extension from Google Admin. It applies the same policy but does not cover everything the desktop agent does.
Do we need a different script per RMM? No. The install and the health check are plain commands. Any RMM that can run a script and read its exit code can run them.
Where to Go Next
The RMM deployment guide has the install commands for Windows, macOS and Linux, the health check and the lock-down steps, and the MSP AI governance guide covers what to do with the findings once the agent is in place. We built ProctoredAI so an MSP can run this rollout from the RMM it already uses, client by client, starting in Passive mode.
Governance starts with visibility. You can't govern what you can't see.
This article is general information, not legal, compliance or security advice, and does not create a professional relationship. Product details are accurate as of the publish date; see ProctoredAI.com/terms for what the service does.