What MSP AI Governance Is
MSP AI governance is a managed service in which a managed service provider sees which AI tools a client's staff use, decides which are allowed, keeps sensitive data out of the ones that are not, and gives the client evidence that the policy works, for every client from one console.
It is the same job an MSP already does for endpoints, email and backups, applied to a new channel. Staff now send documents, client records and source code to AI services every day. The client needs someone to run the controls, and the MSP already runs the devices.
AI governance is not a policy document. A written acceptable-use policy that nobody enforces tells an auditor what the client intended, not what happened. Governance is the policy plus the enforcement plus the record.
Why Clients Are Asking Now
- Staff adopted AI before IT did. Chat assistants, meeting recorders, browser extensions and coding tools arrive with a personal sign-up and no ticket. Most clients cannot list the AI apps in use on their own devices.
- The questions have arrived. AI now shows up in security questionnaires, vendor reviews, audits and board conversations, and the client turns to its IT provider for the answer.
- Rules already apply. HIPAA, PCI DSS, GLBA and the GDPR cover data whatever tool it is pasted into, and the EU AI Act adds duties for organizations that use AI, including AI literacy under Article 4.
- AI tools now act, not just answer. Agents and MCP servers give AI clients access to repositories, CRMs and file systems. That turns a data question into an access question.
An MSP that offers the service turns those questions into a recurring line. One that doesn't will be asked anyway, and will have nothing to show.
The Five Controls of AI Governance
Every AI governance program, from a ten-seat dental practice to a regulated lender, comes down to five controls. Frameworks name them differently, but auditors look for the same five things.
1. Inventory: See Every AI Tool
List every AI app in use, by device and account: the web apps, the desktop apps, the browser extensions, the command-line tools and the local models. Telling products apart matters. Claude and Claude Code, or ChatGPT and Codex, share domains but carry very different risk. ProctoredAI recognizes the 230 AI apps in its catalog by domain or by client signature, and reports AI-looking services the catalog doesn't know yet.
2. Decision: Approve, Warn or Block
For each client, every app gets a status: approved, unapproved or blocked. Unapproved use can ask the person for a business reason first, and an exception can be granted for a set time on one device with the reason on record. Account rules can require a work account rather than a personal one.
3. Data Protection: Stop Sensitive Data Before It Leaves
Inspect prompts and uploads on the device, and act on what they contain: block credentials, redact government ID numbers, warn on bulk contact lists, and keep patient data to the AI tools covered by an agreement. The text in screenshots is read too, so a picture of a spreadsheet gets the same treatment as the spreadsheet.
4. Agent Governance: Control the Tools Behind the AI
Model Context Protocol servers connect AI clients to real systems. Governance means knowing which servers are configured on which devices, what each can reach, and deciding which may run.
5. Evidence: Prove It Every Month
A monthly report per client shows what was used, what was stopped, which devices were enforcing, what changed in the policy, and which framework controls have evidence. The report is frozen with a digest so an auditor can match it to the record. The client's AI Use Policy is generated from the configuration it describes, and each employee's acknowledgment is recorded.
Why the Endpoint Is the Control Point
There are three places to put AI controls: the network, the browser and the device. Only the device sees everything.
| Where the Control Sits | What It Sees | What It Misses |
|---|---|---|
| Network or DNS filter | Which AI domains are reached, on the office network | Content, laptops off the network, and products that share a domain |
| Browser extension | AI web apps in the managed browser | Desktop AI apps, command-line tools, SDK traffic, other browsers and MCP servers |
| Endpoint agent | Every AI app on the device, with content inspected on the device | Apps that pin their certificates are seen but cannot be inspected by any on-device proxy. The coverage view says so. |
An agent also fits how MSPs already work. It installs from the RMM, a script monitor checks its health, alerts open tickets in the PSA, and it runs on every client's devices whatever browser they prefer. ProctoredAI runs on macOS, Windows and Linux, with a managed extension for ChromeOS. A protected system service owns enforcement on Windows and macOS.
Rolling It Out to a Client in Four Steps
- Observe. Deploy the agent from the RMM in observe-only mode. Nothing is blocked; every AI app, account and sensitive-data finding is recorded. Two weeks gives a true picture.
- Review. Walk the client through what was found: the apps in use, the personal accounts, the data that reached AI services. This meeting sells the service.
- Decide and enforce. Approve the tools the client pays for, block the ones it won't accept, apply the policy packs for its industry and frameworks, and publish the AI Use Policy for employees to acknowledge.
- Report monthly. The AI Governance Report goes to the client each month, through a share link with no sign-in needed. Alerts open tickets in the PSA in between.
Industry templates set sensible starting policies for healthcare, finance, legal, defense and government clients, so step three takes an afternoon rather than a project.
Packaging and Pricing the Service
AI governance sells best as a per-device line next to endpoint protection and backup. Clients already understand that model, and it grows with the client.
- Lead with the assessment. The observe step produces a findings report that justifies the service in the client's own data.
- Offer two levels. Governance of AI use, and governance plus endpoint data protection (removable media, clipboard, data at rest) for regulated clients. ProctoredAI's Govern and Protect plans map to those two levels.
- Price per device, monthly. ProctoredAI's partner price starts at $3.00 for Govern and $4.75 for Protect per endpoint per month, and drops automatically as the fleet grows. The MSP sets its own client price. See partner pricing.
- Bill on usage. ProctoredAI bills partners monthly on device-days actually used, with no order forms or minimums. The MSP's own devices are free up to 100 endpoints.
How AI Governance Maps to Compliance
Selecting a client's frameworks applies the matching policy packs, gives the MSP a checklist for the steps outside the rules, and adds control evidence to every monthly report. Each framework has its own page:
- HIPAA: protected health information only to AI tools under a business associate agreement.
- PCI DSS: card numbers blocked from AI services and removable media.
- SOC 2: restricted data transmission and monitored endpoints, with verifiable reports.
- CMMC: CUI and export-controlled content kept out of every AI app.
- ISO/IEC 42001, the NIST AI RMF and the EU AI Act: an AI register, impact assessments and AI literacy evidence.
- GDPR and GLBA: personal and customer financial data redacted or stopped.
Questions to Ask Any AI Governance Platform
| Ask | Why It Matters |
|---|---|
| Does it cover desktop apps and command-line tools, or only the browser? | The most capable AI tools now run outside the browser. |
| Is content inspected on the device, or sent to a cloud to be checked? | Inspection in the cloud means a copy of the sensitive data leaves anyway. |
| Can it tell products apart on shared domains? | Approving a chat assistant should not approve the coding agent from the same vendor. |
| What happens when the agent stops? | A control that fails open is a control that turns off when it matters. |
| Does it fit the RMM and PSA, and run every client from one console? | A service that needs its own tools does not scale across clients. |
| Can an auditor verify the reports? | Evidence needs to be checkable, not just printed. |
| Is it built for multi-tenant MSP use, with white-label reports? | The client should see the MSP's brand, and every client's data must stay separate. |
Questions MSPs Ask
What is the difference between AI governance and AI security?
AI security protects AI systems from attack. AI governance controls how people use AI: which tools, with what data, under what policy, with what evidence. For most MSP clients, who use AI rather than build it, governance is the need.
Is shadow AI the same as shadow IT?
It is shadow IT for AI: tools staff adopt without approval. The difference is what goes into them. A shadow file-sharing app holds files someone chose to upload; a shadow AI tool receives whatever a person pastes in to get help, which is often the most sensitive material they have.
Should an MSP just block all AI?
Blanket blocks push staff to personal phones and laptops, where nothing is visible. Most clients get better results approving one or two tools with work accounts, blocking the risky categories, and stopping sensitive data everywhere.
How long does it take to onboard a client?
The agent installs from the RMM in minutes. A client is typically observed for two weeks, reviewed in one meeting, and enforced the same afternoon.
Does AI governance read everything employees do online?
It shouldn't. ProctoredAI routes only AI services, the web channels a client opts into, and decided MCP servers through the agent. Other sites go direct and are never seen, and prompts are redacted in events by default.
Can the MSP use it internally first?
Yes. Running it on the MSP's own devices is the best way to learn it, and the first 100 internal endpoints are free.