RagnarPitla The authentication setting in Copilot Studio looks like three radio buttons. It is actually three decisions bundled into one control, and two of them are invisible at the moment you choose:
- How users sign in.
- Which channels you can publish to.
- Whether your agent can ever get a token to call a downstream API as that user.
Almost every painful auth conversation is someone discovering decision 2 or 3 weeks after making decision 1. The choice is cheap to make and expensive to reverse, so make it deliberately.
Ask what "auth" means here first
The word covers two unrelated problems, and people routinely bring the second one described as the first:
- User authentication - how the person talking to the agent signs in, and what the agent knows about them. That is this skill.
- Downstream authentication - what identity the agent uses to reach SharePoint, Dataverse, an API or an MCP server. That is connections, connection references and service principals, and it is mostly a different problem with different failure modes.
A useful separating question: does the answer need to differ per user? If two users must see different data, you need user authentication. If every user should see the same thing, you almost certainly want a service identity and the user auth question is a distraction.
For the downstream half, the ALM failure that actually bites at go-live -
connection references that arrive unresolved in the target environment and
leave flows deactivated - is covered by
power-platform-alm-connection-refs. Do not re-derive it here.
The three options and what each one forecloses
Verify the current picker in-product before advising: the option names and the admin controls around them have changed more than once. What follows is the shape of the decision, not a promise about today's labels.
No authentication
Anyone with the link can talk to the agent, and you cannot control who. That is the whole trade. It is defensible for genuinely public content and indefensible for anything else.
Worth knowing: a data policy in the Power Platform admin center can require authentication, which removes this option entirely. If someone insists it was available last month and is not now, look at admin policy before looking for a bug.
Authenticate with Microsoft
The convenient option, and the one chosen by default under time pressure. It sets up Entra ID authentication with no manual configuration, and because Teams already identifies the user, people are not prompted to sign in again.
The two things it costs you:
- Channel reach. This option is oriented at Teams and Microsoft 365. If the roadmap includes a public website, a bespoke channel or an embedded experience, confirm the channel is supported before building on it.
- No access token.
User.AccessTokenandUser.IsLoggedInare not available. You get identity, not delegation. An agent that must call an API as the signed-in user cannot do it from here.
That second point is the single most common late discovery in this area. If there is any prospect of calling a downstream API on the user's behalf, this option will have to be unwound.
It is also not available for agents integrating with Dynamics 365 Customer Service, which is worth checking early in a customer service engagement rather than late.
Authenticate manually
More configuration, and the only option that gives you a token. Supports Entra ID - including certificate and federated-credential variants - and generic OAuth2 providers, which is what you need for a non-Microsoft identity provider.
Choose it when any of the following is true:
- The agent must call an API as the signed-in user.
- You need to publish beyond Teams with real authentication.
- Your identity provider is not Microsoft.
One governance trap that is easy to miss: with Entra ID you can require sign-in and then use agent sharing to control which people in the organisation may chat with the agent. With generic OAuth2 you can require sign-in, but you cannot restrict which users - anyone who successfully signs in with that provider can talk to the agent. If the requirement is "only our finance team may use this", a generic OAuth2 provider does not deliver it, and this surprises people at review time.
The failure modes worth recognising on sight
A User. variable shows as Unknown after an auth change. Switching from
manual authentication to Authenticate with Microsoft removes
User.AccessToken and User.IsLoggedIn. Topics referencing them do not
rewrite themselves - they break, and they break quietly until publish. After
any change to the authentication mode, search the topics for the variables that
no longer exist and fix them before publishing.
"I changed the setting and nothing happened." Authentication changes take effect only after the agent is published. Test-pane behaviour after an auth change is not evidence of anything until you publish.
"It works for me and not for them." Different answers for different users is a permissions symptom far more often than an auth-configuration symptom. The agent may be authenticating perfectly and simply retrieving what each user is allowed to see. Establish whether the difference tracks the user or the session before touching auth config.
"The option is greyed out." Tenant admin controls can lock the manual authentication option so it cannot be changed in Copilot Studio at all. That is a governance decision made elsewhere, not a defect - find who owns the policy.
An MCP server that will not appear or work. MCP in Copilot Studio requires generative orchestration to be turned on. Confirm that before debugging credentials, because the symptom looks like a connection problem.
Design it in this order
- Does behaviour need to differ per user? No means a service identity and a much simpler design. Say so plainly - it removes work.
- Where must this be published? Get every channel on the table now, including the ones that are "maybe next quarter". Channel reach is the constraint most likely to invalidate the easy choice.
- Will the agent call anything as the user? If yes, or plausibly yes later, you need the option that yields a token.
- Who is allowed to talk to it? If the answer is a specific group rather than "anyone in the tenant", check that your chosen provider can actually express that.
- Only then configure, and publish before you believe any of it.
Say when you do not know
Field-level configuration for manual authentication - URL templates, grant types, scope delimiters, certificate settings - is genuinely fiddly and changes between providers. Do not reconstruct those templates from memory. Send people to the field reference and let them read the current table.
The same applies to which channels support which authentication mode today. That matrix has moved, and a confident wrong answer here costs someone a rebuild.
Verified references
- Configure user authentication - the three options, the variables each one exposes, the sharing behaviour and the manual configuration field reference.
- Extend an agent with Model Context Protocol - MCP integration, including the generative orchestration prerequisite.
- Power Platform ALM overview - for the downstream identity and connection reference half of the problem.