AI Shared Responsibility Is Now an MSP Problem

You’ve heard it before: I thought you handled that

This is where the conversation gets interesting. The vendor can secure the platform. The Microsoft Service Provider (MSP) can secure the tenant. But if users are pasting sensitive data into public AI tools, creating accounts, owning the relationship with the AI, then the whole shared responsibility model gets messy fast.

The biggest AI risk isn’t Copilot. It’s shadow AI

Shadow AI is the clearest example of the problem. Employees are already using public AI tools to summarize documents, draft emails, rewrite policies, and analyze data. They are choosing the AI themselves without your governance. From the user’s perspective, it feels efficient. From the MSP’s perspective, it can mean sensitive data leaving the tenant, no audit trail, and no idea which tools are in play.

It is a governance issue and MSPs have historically done a terrible job at it. It was somehow thought to be advanced and niche. If a client has no acceptable use policy for AI, no approved tool list, and no guidance on what data can be entered into public models, then you are supporting an environment that is already out of control. The tricky part is that clients often discover this only after someone uploads the wrong thing to the wrong service. Governance by crisis isn’t governance. It’s not proactive and clients with MSP contracts should expect better.

Shadow AI also exposes a familiar MSP blind spot: the gap between endpoint protection and behavior. Traditional security tools can reduce risk, but they do not stop a well-meaning employee from pasting customer data into a chatbot. That means the MSP has to treat AI like email security, not just app selection.

New AI-based tools are out there that let you implement AI governance. I was just at a conference where three of them had booths. They can prevent access to all non-approved AI tools; they can prevent insecure data upload; they can protect intellectual property. But they are expensive. This isn’t a $1 per seat add-on. It’s more like $6-$8. This is your new Remote Monitoring and Management (RMM). Not in function but in importance to the services that you offer and it’s going to mean that MSPs reprioritize what they offer, what they spend on and why.

The MSP’s responsibility is no longer security but governance

For MSPs, AI shared responsibility should become a service conversation, not a side note. The first step is to define which AI tools are approved, which are prohibited, and what data is never allowed in external models. The second step is to review identity and access so that Copilot, or any AI tool, can only expose what users should already be able to see.

A practical AI governance baseline for clients should include:

• An AI acceptable use policy.
• A list of approved AI tools.
• Data classification rules for what may or may not be entered into AI.
• Identity and access cleanup.
• Data lifecycle management.
• User training on prompt hygiene and output verification.
• A process for reviewing AI-related incidents.

That framework turns AI from an unmanaged risk into a managed service. It also gives you a more strategic role. Instead of reacting to “Can we use this AI tool?” you can answer, “Yes, but here are the controls, the limits, and the business rules.”

The AI responsibility line MSPs can’t afford to leave unclear

The judgment call is deciding where responsibility ends. Clients will expect the MSP to secure AI, but the MSP cannot own every user prompt or every business decision made with AI output. What the MSP can own is the environment around the tool: policy, permissions, identity, training, and governance.

That is the new shared responsibility model for AI. Shadow AI shows what happens when nobody owns it. The MSP is no longer just the technician in the background. The MSP is part of the trust boundary.

I recommend that you put the AI Shared Responsibility model in writing and make it part of your contract. It’ll protect you and it’ll also differentiate you from your competitors.