If you’re an MSP, an MSSP, a Microsoft integrator or an ISV, you’ve probably had this conversation already. A client says: “We want Copilot agents to take over the invoice chasing, the onboarding paperwork and the first line of support.” And in the very next sentence: “But is it safe? Who controls what these agents can do?”
That second question is the opportunity. Clients will adopt Copilot agents with or without you. The partner who can answer “is it safe?” across every tenant, with evidence, becomes the one they trust to run it.
This post is for partners. It covers the demand you’re seeing, why security always follows, what makes it hard across many tenants, and a practical model to discover, authorise and secure Copilot agents for every client, and to build a service around it.
The demand is real, and so is the worry
The numbers from the channel point the same way:
- In Augmentt’s survey of 193 MSPs (August 2026), 40% said AI and Copilot governance is their fastest-growing client request, ahead of security and compliance at 24%.
- In the same survey, 47% had an AI or Copilot data exposure or near miss in a client tenant in the past year, and only 10% said more than three-quarters of their tenants were Copilot-ready.
- An AvePoint and Omdia study of 333 MSPs found governance and compliance were the most cited barrier to client AI adoption, at 51%.
In other words: clients want it, partners are being asked to deliver it, and most tenants aren’t ready. That’s exactly where a partner earns its place.
Why security always comes next in the conversation
Copilot agents aren’t like the apps clients bought before. They act on their own. A Copilot agent built to chase invoices reads the finance mailbox, opens files, and sends emails. An onboarding agent creates accounts. A support agent reads customer records.
So the client’s real questions are:
- Which agents do we have? Including the ones staff built themselves.
- Who said yes to each one? And who owns it now?
- What can each one reach? Mail, files, finance, HR?
- What is it talking to? Other apps, other agents, outside services?
- Who’s watching it at 2 am? And who stops it if it goes wrong?
If you can answer those five for every client, you’re not selling a tool. You’re selling confidence.
Why it’s hard in a multi-tenant world
Microsoft is giving every organisation better agent controls. Agents get their own identities through Microsoft Entra Agent ID, Copilot Studio records each agent’s creator as its sponsor, and Microsoft Agent 365 brings agents into a registry. Use all of it.
But partners face problems a single business doesn’t:
- Every client is a separate tenant. Forty clients means forty registries, forty sets of sharing settings, forty lists of app grants.
- Licences differ by client. Some features need extra licences many small clients don’t have.
- Agents come from everywhere. Copilot Chat’s Agent Builder, SharePoint, Copilot Studio, the Agent Store, Microsoft Foundry, and third-party AI apps connected through consent. See what counts as a Copilot agent.
- Some clients run Google Workspace too, with Gemini agents that have their own accounts.
- Your team’s time is the bottleneck. Nobody can review every agent in every tenant by hand, every week.
A practical model: discover, authorise, secure, prove
Here’s a model that works across a client base, whether you’re a five-person MSP or an MSSP running a SOC.
1. Discover
Make one list per client of every agent and AI app with access to their data: Copilot agents, Copilot Studio and Foundry agents, and third-party apps holding app grants. Include who built each one and when.
2. Authorise
Every agent gets three things before it’s “approved”:
- A sponsor: a named person at the client who answers for it.
- A scope: the mail, sites and systems its job needs, and nothing more.
- A review date: every 90 days, the sponsor confirms it’s still needed.
This is where ISVs and integrators come in too. If you build Copilot agents for clients, ship them with a clear sponsor, a documented scope and their own identity. It makes your agents the easy ones to approve.
3. Secure
Watch each agent’s four control layers continuously:
- Identity: does it have its own, and is its sponsor still here?
- Authentication: are its keys and tokens used where and when you’d expect?
- Grants: does its access still match its job? Has oversharing crept in?
- Connections: has it started talking to something new?
Link agent alerts to the sponsor’s account and mailbox, so an agent going off script right after its sponsor’s suspicious sign-in is treated as one incident, not two tickets.
4. Prove
Give each client a plain-English monthly view: agents added and removed, access narrowed, issues found and fixed. This is what turns agent security from a cost into a line on the invoice clients understand.
Turning it into a service
How partners package Microsoft 365 security varies. In Augmentt’s survey, 36% sold it as a project, 32% as a separate line item and 30% inside their standard fee. Agent security fits all three:
- Project: a Copilot readiness and agent review before rollout. Find oversharing, list existing agents and AI apps, set the approval rules.
- Line item: ongoing agent security per client: continuous discovery, approvals, monitoring and the monthly view.
- In the fee: for clients already on a security bundle, agent coverage becomes the reason the bundle is worth more this year.
Start with the project. It’s the easiest yes, and it surfaces the findings that make the ongoing service obvious.
A worked example: one MSP, 30 clients, the first month
This is an illustrative example, not a real client. A 12-person MSP looks after 30 Microsoft 365 clients. Eight are asking about Copilot agents.
Week 1. The MSP runs discovery across all 30 tenants. It finds 64 agents and AI apps with access to client data, including 11 the clients didn’t know about, and two third-party AI apps with access to entire mailboxes.
Week 2. For the eight Copilot clients, it assigns sponsors and scopes. Three agents built by staff who have since left are paused until someone claims them. One Copilot Studio agent can surface an HR folder shared with everyone; the sharing is fixed in one change.
Week 3. Monitoring is on. One night, an AI app’s token is used from a new location while its sponsor’s account shows an unusual sign-in. Both are handled as one case: the session revoked, the token reset, and the sponsor told in the morning.
Week 4. Each Copilot client gets a one-page summary. Five sign up for ongoing agent security as a line item; the other three take it inside their bundle at renewal.
The MSP didn’t add staff. The discovery, linking and first-line checks were automated; the engineers made the decisions.
Checklist for partners
- You can list every agent and AI app across all client tenants in one place.
- Every agent has a sponsor at the client, a scope and a review date.
- Oversharing is checked before each client switches agents on.
- Agent alerts are linked to the sponsor’s sign-ins and mailbox.
- You can stop any agent in any tenant in one step, at any hour.
- Disruptive actions always wait for a person; protective ones can run on their own.
- Each client gets a plain-English monthly view.
- You’ve decided how to package it: project, line item or in the fee.
- If you build agents for clients, they ship with their own identity, a sponsor and a documented scope.
For the client-facing side, use the Copilot governance checklist for MSPs: who decides, who can build and share agents, and the monthly review.
How Centeye helps partners
Centeye Synthwatch is built for exactly this, across every client tenant from one console. Today it lists the AI agents and AI apps holding access to Microsoft 365 and Google Workspace, gives each one an owner and a 90-day re-approval, reads what each grant allows with a risk score, finds oversharing agents could surface, and traces app-to-app chains. It can revoke a grant, lock a service principal or reset an API token, at the setting you choose per client.
Synthagent, the guard agent inside Synthwatch, watches every agent’s four control layers continuously, links what it sees to the sponsor’s account and mailbox in one case, and fixes it in place. Centeye Praxis does the triage and heavy lifting above it, so your engineers focus on the decision, not the digging.
Reading Copilot Studio’s registry, Microsoft Entra Agent ID and Google’s agent identities directly is coming. Synthwatch and Synthagent are in early access.
Talk to us. If your clients are asking about Copilot agents, we’d like to hear how you’re handling it. Book a 30-minute call or ask for early access.
Frequently asked questions
Should MSPs sell Copilot agent security?
Clients are already asking. In Augmentt’s 2026 survey, 40% of MSPs said AI and Copilot governance is their fastest-growing client request. Agent security is the natural next step after a Copilot rollout.
What’s different for MSSPs?
MSSPs already run detection and response. Agents add a new kind of identity to watch. The key is linking agent behaviour to the human sponsor’s sign-ins and mailbox, so agent incidents land in the same case as the rest.
What should ISVs building Copilot agents do?
Ship every agent with its own identity, a clear sponsor at the client, a documented scope, and read access unless write is essential. Agents built that way are easier for clients and their MSPs to approve.
Does Microsoft Agent 365 solve this for MSPs?
It gives each organisation an agent registry and identities. It’s licensed and run per organisation, so partners still need a way to work across every client, and to cover grants, connections and agents outside Microsoft.
Where should a partner start?
Pick the three clients asking most about Copilot. Run discovery, assign sponsors and fix the worst oversharing. That first month usually makes the case for the rest.
New to the terms? See AI agent, non-human identity and OAuth grant in our glossary.
Augmentt survey of 193 MSPs, August 2026, as reported by IT Brief. AvePoint and Omdia study of 333 MSPs, 2026, as reported by ChannelLife. Microsoft Agent 365, Entra Agent ID and Copilot Studio sponsors as described by Microsoft. The worked example is illustrative. Synthwatch and Synthagent are in early access. Microsoft, Copilot, Copilot Studio, Entra, Foundry and Agent 365 are trademarks of the Microsoft group of companies.