MSP Guide to Microsoft 365 Copilot Governance: Permissions, Agents and AI Rollouts

That is the pressure behind a new Channel Insider "Partner POV" interview. Managing Editor Victoria Durgin spoke with Peter Fidler, founder and president of WCA Technologies, a New York City MSP. The conversation covers Copilot adoption, AI governance, agents, outcome-based pricing and the Microsoft partner community. Fidler is a practitioner, so his comments are observations and predictions. They are not survey data or measured ROI. They still map neatly onto what Microsoft's own documentation says admins should do.

Who is talking, and a note on names​

Fidler says WCA Technologies was founded in 1988. He says it serves mostly nonprofit, financial, legal, entertainment and real-estate clients, and that as a Microsoft partner it leans toward Copilot.

He is also president of the Americas region of a Microsoft-partner association. The automated transcript renders that name inconsistently, as "AMCP" and "IMCP". The organization appears to be the International Association of Microsoft Channel Partners (IAMCP). The IAMCP describes itself as a professional association of Microsoft partners, organized in chapters across the Americas, APAC and EMEA. It was formed in 1994.

The transcript also spells the guest's surname two ways ("Fedler" and "Fidler"). This article uses Fidler, the spelling he gives when he introduces himself.

The core problem: customers got there first​

Durgin's framing is that AI is one of the first technology waves where customers and employees got hold of the tools before their MSPs did. Fidler describes clients asking for Claude, ChatGPT or other models, while he nudges them toward Copilot. His answer is education rather than prohibition.

His approach, as described:

• Explain the risks before a tool is deployed.
• Make sure the client understands where its data will sit.
• Avoid blocking experimentation outright, because the aim is to keep clients moving forward.
• Treat rollout as change management, not a software install.

He also says WCA is its own "customer zero." It uses AI for presentations and proposals and tunes outputs to the firm's house terminology and his own writing style. He stresses that output has to be verified.

One caution about that claim. The interview does not explain the mechanism behind this customization. Nothing in the source says Copilot automatically learns a user's style just from ordinary use. Treat the "training" language as loose shorthand, probably meaning prompting and instructions, not model training.

Where the "Pandora's box" warning needs sharpening​

Fidler warns that unmanaged AI use is like uploading all your files to the internet. As rhetoric, it lands. As a technical description, it overreaches. The real risk depends on which product, which account type, which configuration and which data source is involved.

Microsoft's documentation draws that distinction clearly. For Microsoft Copilot in a commercial tenant, Microsoft says:

• Copilot is built from large language models, Microsoft Graph content the user has permission to access, and the Microsoft 365 apps. That matches the Microsoft Learn description of Copilot's components.
• Copilot only surfaces organizational data the individual user has at least view permission for. Microsoft explicitly tells admins to use the permission models in services such as SharePoint to make sure the right people have the right access.
• Prompts, responses and data accessed through Microsoft Graph aren't used to train the foundation models, according to Microsoft.
• Interaction history is stored, and admins can manage it with Content search or Microsoft Purview, including retention policies.
• Admins decide whether third-party models such as those from Anthropic and OpenAI are used in Copilot experiences, and additional terms may apply.

The practical takeaway is less dramatic than "Pandora's box" and more useful. Copilot does not override permissions. But if a tenant is full of overshared sites and stale group memberships, an AI assistant that searches everything a user can see will make that sloppiness much easier to find. The cleanup job is permissions hygiene, which is classic MSP territory.

Microsoft's guidance also covers the unmanaged side. When agents extend Copilot, admins should check each agent's privacy statement and terms, because those determine how it handles organizational data. Microsoft says admins can review an agent's required permissions and data access in the Integrated apps section of the Microsoft 365 admin center, and choose which agents are allowed.

From personal assistant to agents​

Fidler's most forward-looking point is that usage will move from the "intern" model, where a person types a prompt, to agents that run on a schedule against business data. His examples are illustrative. They are not confirmed WCA deployments:

• A Monday-morning briefing of the day's priorities.
• A weekly analysis of a financial file, flagging where to focus attention.
• A shared prompt that produces a team-wide priorities summary.
• A news-monitoring agent that watches a firm's top clients for relevant events.

He compares the last example to expensive newswire-style services. He offers no pricing data, so that comparison is opinion, not a verified cost saving.

He also proposes agents for MSP-style workflows such as employee onboarding and offboarding: account creation, group membership, forms, HR guidance and training videos. The interview mentions a possible 15-minute onboarding and a 10% time saving. Both are hypothetical examples of the kind of outcome an MSP could promise. They are not results WCA has reported.

Do customers still need an MSP?​

The interview directly addresses the "SaaS apocalypse" worry that surfaced earlier this year, the fear that customers would replace software and service providers with homemade agents. Durgin says that didn't come to pass to the degree feared. Fidler's response is that some customers will try, and MSPs shouldn't stop them. But doing it well takes process knowledge, expertise and investment. In his analogy, an MSP is like an offensive lineman, noticed only when something goes wrong.

Microsoft's documentation supports the idea that agents are not a snap-your-fingers affair. Its guidance for agent builders says to:

• Map every data and action flow.
• Plan identities, authentication and permissions with least privilege.
• Treat untrusted inputs, such as emails or support tickets, as a prompt-injection risk.
• Avoid giving agents sensitive operations without careful human intervention.

Those are the "known unknowns" that make a managed service valuable.

A practical governance checklist for MSPs​

Pulling Fidler's advice and Microsoft's documentation together, here is a checklist an MSP could run for a Copilot-focused client. It is a synthesis, not a Microsoft-published procedure.

1. Inventory what's in use. Find out which AI tools and agents employees already use or want, and where each sends data.
2. Define approved use cases and data sources. Decide what business tasks the tools are for and what content they may touch.
3. Review permissions first. Check SharePoint, Teams and group access, because Copilot works within what each user can see. This includes access granted to people outside the organization through shared channels.
4. Use existing protection controls. Microsoft says Copilot honors Purview Information Protection usage rights and that existing sensitivity labels, retention, eDiscovery, audit and DLP controls apply. Some capabilities depend on licensing, so confirm what the client owns.
5. Govern agents in layers. In Copilot Studio, admins can use Power Platform data policies to govern authentication, knowledge sources, actions and connectors, HTTP requests, publishing to channels and triggers. Microsoft also describes environment routing to give makers a safe place to build, and the option to disable publishing of agents that use generative AI features.
6. Check what Agent 365 adds. Microsoft says organizations that onboard Agent 365 can have Copilot Studio agents represented as identities in Microsoft Entra. Those identities can then be governed with Conditional Access and access governance workflows. This applies only to organizations that adopt it.
7. Turn on logging. Microsoft says maker audit logs are visible in Microsoft Purview, and agent activity can be monitored through Microsoft Sentinel.
8. Pilot, then decide. Run a contained test with a defined group, watch results, and have a rollback plan.
9. Verify outputs. Microsoft itself says generative AI responses aren't guaranteed to be fully factual and users should review them before sharing.

One limit worth stating. Microsoft's documentation describes controls and configuration choices, not a guarantee against every misuse. The Copilot Studio documentation notes, for example, that Customer Lockbox doesn't cover every category of outbound data, including security audit telemetry and certain Agent 365 governance events. It also says Copilot Studio's content moderation covers predefined harmful-content categories, and that there is no tenant-configurable policy to block conversations on arbitrary admin-defined topics. Third-party and externally hosted services carry their own identity, privacy and compliance responsibilities.

Outcome-based pricing: attractive, hard to do​

Durgin asks whether MSPs should shift from "here's the tech" to "here's what it does for you." Fidler thinks billing may migrate toward outcomes, but cautions that it is hard to measure and price. A couple of badly underestimated jobs can make providers wary of the model. His practical advice is to build repeatable services that can be delivered across many clients, and to treat that as the foundation for any value-based pricing. He says outcome billing works best on something repeatable. That is a sensible, unglamorous point.

His wider message on operations is to pilot small, accept that some pilots will fail, and roll back when needed without letting ego or sunk cost decide. He says he has launched technology that didn't pan out, and that containment kept the damage low.

The community angle​

Fidler argues that peer groups will speed the transition, including peers who are competitors. He names Trusted Tax Alliance and Evolve as groups he has belonged to. He tells a story about asking a competitor who beat him on a large nonprofit deal what he had done differently.

On the partner association, he says:

• It is volunteer-run, with chapters across the US.
• It has more than 3,000 partners worldwide and hundreds of members in the Americas.
• It ran two Microsoft-financed AI and Copilot cohorts, and in the first, one participant closed a $70,000 deal within three to four weeks.

Those are his claims. Public descriptions of the association vary. One partner-published profile says about 1,800 member companies and 2,900 individual members globally, while an AvePoint announcement from October 2025 cites over 5,000 individuals from nearly 2,000 organizations. The numbers differ by date and source, so treat the "3,000" figure as approximate and unverified. The cohort details and the $70,000 sale were not independently confirmed.

He also suggests Microsoft expects partners to do more than collect recurring revenue, with Copilot and AI outcomes in mind, which he says will require retraining many MSPs. That is his perspective on Microsoft's channel, offered with candor about how hard the vendor can be to work with.

What to take away​

• AI work for MSPs is shifting from selling licenses to governance and change management. The risk is as much about permissions and process as about the tool.
• Fidler's examples are predictions. The interview offers no ROI study, adoption survey or evidence that agents are displacing existing applications.
• Microsoft's documentation gives a more precise risk model than the "uploading files" analogy. Copilot respects user permissions, so overshared content is the exposure to fix, while unmanaged or third-party tools carry separate terms and risks.
• Agents raise the stakes. Scheduled, data-connected agents need authentication, connector limits, audit trails and human approval for sensitive actions.
• Outcome pricing is a direction, not a switch. Repeatable services come first.

For WindowsForum readers who administer Microsoft 365 tenants, the lesson from this interview is the same one the tooling already teaches. Before anyone asks for an agent, clean up who can see what.