From Aryan Vatsa | Product & Market Analysis
Gemini Agent for Microsoft 365 vs Copilot: What Admins Can Verify in Preview
On this page
Google's new Gemini agent can be reached from inside Microsoft 365, and Google's launch post explains that in 2 sentences. Neither sentence says how the agent connects to your tenant. If you run Microsoft 365, Gemini for Microsoft 365 is a real option to watch, and for now it is an option you cannot configure from documentation. Microsoft 365 Copilot, by contrast, publishes its permission model, its training policy and its data boundary exceptions in writing.
Key takeaways
- Google has announced the Gemini agent, not documented it for Microsoft 365 admins. Its launch post names Microsoft 365 as a channel and Office and Teams as connected tools, with no detail on consent scopes, add-ins or tenant setup.
- The two products differ on whose identity does the work. Microsoft says Copilot only surfaces data the signed-in user can already view. Google says each of its agents gets its own identity and its actions are logged to the agent, not a person.
- Claude is not the differentiator it sounds like. Google routes work across Gemini and Anthropic's Claude models, but Microsoft already runs Anthropic models in Copilot, on by default for most commercial tenants outside the EU, EFTA and UK.
- Price, GA date and plan eligibility are not in Google's own text. An analyst report says general availability is expected by the end of October 2026 with no added seat charge. Treat both as unconfirmed until Google publishes them.
Short answer for admins: Google's Gemini agent works inside Microsoft 365 and Slack and routes jobs across Gemini and Claude models. It was announced at Gemini at Work 2026 and is in private preview. Google has published agent-level controls such as identity, audit trail and spend caps. It has not published how the agent connects to a Microsoft tenant, what it costs, or when it ships.
What Google actually announced about Gemini for Microsoft 365
Google Cloud introduced the Gemini agent at Gemini at Work 2026. The blog.google announcement is dated 8 October 2026 and calls it "a universal agent for work". The longer Google Cloud post, adapted from Thomas Kurian's keynote, carries a 9 October date.
Kurian's line is "You give it objectives, not instructions". The agent plans the work, connects to business systems, writes content and code, and chooses the model for each job.
The two sentences that mention Microsoft 365
Read the Cloud post looking for Microsoft and you find two passages. The first says Gemini can be reached from "any channel (command line, Google Workspace, Microsoft 365, or Slack)". The second says it "can securely connect to the software your company already runs, including collaboration tools like Confluence, Microsoft Office, Teams, Slack, and Workspace".
That is the whole of it. Outlook, Word, Excel, SharePoint and OneDrive are not named. There is no mention of an Office add-in, a Teams app, a Microsoft Graph connector or an Entra app registration. Google also says the agent can run "headless" inside third-party surfaces, which suggests an API path, but it does not say which path Microsoft 365 uses.
It is a keynote adaptation, and keynotes sell outcomes. But anyone telling you how Gemini handles your mailbox today is guessing.
What the analyst coverage adds, and how far to trust it
Futurum Group's write-up by Nick Patience, published 9 October 2026, goes further than Google's own text. It reports that the agent is in private preview and that Google expects general availability by the end of October 2026, starting in North America. It also reports no additional seat charge where Gemini Enterprise is available, and says administrators decide which models the organisation can use.
None of those four points appears in the Google post as read for this article. They are still second-hand, so this post labels them as reported rather than confirmed. Android Authority independently describes the agent as limited to enterprise customers in a private preview.
| Question | Confirmed in Google's text | Reported by analysts only | Not found anywhere |
|---|---|---|---|
| Does it work in Microsoft 365? | Yes, as a channel and via Office and Teams connections | Depends on access Microsoft allows (Futurum analysis) | Which apps, and through what mechanism. |
| Which models? | Gemini family and Anthropic Claude today, others later | OpenAI models to follow; admins pick allowed models | Per-tenant model switch in Google's text. |
| Availability | Industry specialisations in preview for financial services and legal | Private preview, GA by end of October 2026 in North America | Eligible plans and regions in Google's text. |
| Price | Hard project spend caps in Cloud Billing | No additional seat charge where Gemini Enterprise is available | Any per-user or per-task price. |
| Data handling | Agent identity, audit trail, sandbox, gateway | Nothing further | Training policy, residency, retention for Microsoft 365 data. |
Columns reflect the Google Cloud blog post of 9 October 2026, the blog.google post of 8 October 2026, Futurum Group's analysis of 9 October 2026 and Android Authority's report of 8 October 2026. A Google admin help page for the agent was not found as of 23 October 2026.
Whose identity the Gemini agent and Copilot act under
This is the difference that matters most to a Microsoft 365 admin. It decides who appears in your sign-in logs, which consent policy applies, and what your audit tool records.
Copilot acts as the signed-in user
Microsoft's privacy documentation states that Copilot "only surfaces organizational data to which individual users have at least view permissions". Grounding runs through Microsoft Graph, and Microsoft says the Semantic Index "honors the user identity-based access boundary".
That model has a known weakness. If your SharePoint permissions are loose, Copilot surfaces the looseness at conversational speed. The permission gap that creates is covered in the comparison of how Glean and Copilot inherit permissions. The model is still legible: one identity, one set of permissions, one audit trail.
Gemini acts as itself
Google describes the opposite design. "Every agent gets its own identity, cryptographically attested and governed like an employee," the Cloud post says. Agents receive "fine-grained, role-based access permissions, approved by your organization's security administrators". Every action "is written to an audit trail and attributed to the agent rather than to a person".
Google's coworker agents go further, with their own @agents.company.com email addresses. A coworker agent "acts under its own identity rather than yours, and it sees only what you share with it". That is a cleaner least-privilege story than inheriting a user's entire Graph footprint.
The unanswered question is which model applies when the agent reaches into Microsoft 365. If it acts under its own identity, you need an app registration and admin-granted Graph permissions in Entra. If it acts on a user's behalf, you need delegated consent, and the agent's activity may be logged under the user's name. The trade-offs between the two patterns are set out in the piece on agent credentials versus user impersonation. Google has not said which one it uses here.
Data handling: where Copilot is specific and Gemini is general
Microsoft's documentation is long, dated and occasionally inconvenient for Microsoft. It names its own exceptions.
On training, Microsoft states that "prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs, including those used by Microsoft Copilot". On residency, it says Copilot is an EU Data Boundary service for EU customers. It then adds that customers outside the EU "may have their queries processed in the US, EU, or other regions," and that calls can move to other regions during high utilisation.
On retention, Copilot interactions are stored in a hidden folder in the user's mailbox. Admins can set retention policies for Copilot interactions in Purview, and eDiscovery can search them. Sensitivity labels apply: where data is encrypted by Purview Information Protection, Copilot honours the usage rights granted to that user.
Microsoft also names two exceptions most buyers miss. Web search queries generated by Copilot fall outside the EU Data Boundary and outside its Data Protection Addendum. Anthropic models used inside Copilot are, in Microsoft's words, "currently excluded from the EU Data Boundary".
Google's Cloud post makes three general data claims. Agents run in an Agent Sandbox with its own network boundary. All agent traffic passes through Agent Gateway, which Google calls an AI network firewall that enforces your policies in real time. And in the Workspace context, "no outside connector holds your data".
What the post does not cover is whether prompts and Microsoft 365 content touched by the agent are used for training, where that content is processed, or how long it is kept. None of that has been stated for this product, on this surface.
Claude now sits behind both Gemini and Copilot
Much of the launch coverage framed Claude support as the surprise. Google says the agent runs each job on the model that fits best, "orchestrating across our Gemini family of models and Claude models from Anthropic today". For a Google product, routing work to a rival lab's model is a real strategic concession.
For a Microsoft 365 admin, it is not new. Microsoft's subprocessor page states that "Anthropic has onboarded as a Microsoft subprocessor". Microsoft enables Anthropic models "on by default for most customers in commercial cloud (excluding EU/EFTA and UK)". Your users may already be sending Copilot work to Claude.
The default-on setting most tenants never reviewed
Microsoft's admin path is documented. In the Microsoft 365 admin center, under Copilot settings, an AI Administrator or Global Administrator can disable Anthropic as a Microsoft subprocessor. Access can also be assigned to specific users or Entra security groups. A separate setting for Copilot in Microsoft 365 apps with Anthropic models arrived for EU, EFTA and UK tenants on 3 April 2026, where it defaults to no users.
Microsoft also offers a variant called Anthropic models with Data Retention, which is off by default. Under that option, Anthropic, not Microsoft, stores most inputs and outputs for up to 30 days. Microsoft's page says Anthropic does not use that retained data for model training without express permission. Provider terms are compared in the side-by-side of model provider terms.
So the honest comparison is not "Gemini offers Claude, Copilot does not". Both route some work to Anthropic. Only Microsoft has published the admin switch, the regional defaults and the data boundary exclusion. My view is that multi-model routing has become table stakes for workplace agents, and buyers should stop paying a premium for it.
Admin controls for Microsoft 365 Copilot and the Gemini agent today
On the Microsoft side, controls are scattered across consoles, but they exist and are dated. Web grounding is governed by an "Allow web search in Copilot" policy in the Cloud Policy service. If you leave it unconfigured, web search is available to users. In US GCC and DoD tenants it is off by default.
Agents inside Copilot are managed in the Microsoft 365 admin center. Microsoft's agent management documentation says admins can enable, disable, assign, block or remove agents, and that users can only reach agents the admin has allowed. It is where a Gemini integration would most likely appear if it ships as a Copilot-compatible agent.
What Google says you can configure
Google's stated controls sit at the agent platform level, not the Microsoft 365 level. You can set a hard limit on a project's AI spend in the Cloud Billing console, and Google says Gemini enforces it. When a cap triggers, the project's agent pauses and can be resumed with one click.
That spend cap is a genuine advantage. Microsoft's seat licence caps cost per user but not the consumption behind agents billed through credits, a pattern traced in the map of where Copilot Credits meter. Google has stated one.
The rest of Google's list describes its own infrastructure. None of it tells a Microsoft admin what shows up in Entra, Purview or the Microsoft 365 audit log when the agent reads an Outlook thread. Your existing controls still apply at the Microsoft boundary. By default, Entra lets users consent to apps for permissions that do not require admin consent, and Microsoft recommends restricting that to verified publishers.
Pricing and GA for the Gemini agent: the questions Google has not answered
Microsoft 365 Copilot has a list price. It is $30 per user per month on an annual plan, on top of a qualifying suite licence. Those seat economics are covered in the comparison of what the $30 Copilot seat buys.
For the Gemini agent inside Microsoft 365, there is no price in Google's text. Futurum reports that Google will not charge additional seats for the agent where Gemini Enterprise is available. Even if that holds, it leaves three questions open for a Microsoft 365 shop that is not a Google customer today.
First, does reaching the agent from Microsoft 365 require a Gemini Enterprise subscription for every user who invokes it? Second, are jobs routed to Claude billed the same as jobs routed to Gemini? Google's smart routing promises the best performance "at the lowest possible cost," but the cost to you is not published. Third, does a project spend cap apply per user, per team or per tenant?
General availability is equally open. Google's own text puts only the industry specialisations, for financial services and legal, into preview. The end-of-October GA date and the North America first rollout come from Futurum alone. If you are planning a renewal around that date, allow for a slip.
Where this comparison is weakest
This post compares a generally available product with a product announced two weeks ago. That asymmetry flatters Microsoft, and you should discount for it.
Preview products are always thin on paper
Copilot launched without most of the documentation quoted here. Microsoft's Anthropic subprocessor page carries dates from 2026, more than two years after Copilot's general release. Google will almost certainly publish admin documentation before or at GA. Judging a private preview on missing documents is partly judging it for being early.
Preview customers may already hold detailed technical guidance under NDA.
Documentation is not the same as a working control
Microsoft's well-written pages do not prove its controls work as written in your tenant. The EU Data Boundary carries named exceptions for web queries and Anthropic models. A vendor that documents more also exposes more of its weaknesses, which is to Microsoft's credit and not a reason to assume Copilot is safer by default.
The counter-case for Google is real too. An agent that acts under its own identity, sees only what is shared with it, and logs actions to itself is arguably the better security design. The alternative inherits a user's full Graph reach. If Google implements that design faithfully at the Microsoft 365 boundary, it could end up the stronger option on least privilege. We cannot tell yet, and neither can you.
A preview evaluation checklist for Gemini in Microsoft 365
If your organisation is in the preview or expects an invitation, put these questions to Google in writing. Do it before any production mailbox is connected. Each maps to a control you already run on the Microsoft side.
| Ask Google | Why it matters on the Microsoft side | The Copilot equivalent you can compare against |
|---|---|---|
| Does the agent use delegated user consent or its own app identity with application permissions? | Decides the Entra consent path and whose name appears in sign-in and audit logs | Copilot acts as the signed-in user through Microsoft Graph. |
| Which Microsoft Graph scopes does it request, by app? | Mail.Read on every mailbox is a very different grant from one user's files | Copilot is bounded by each user's existing view permissions. |
| Is Microsoft 365 content or prompt data used for training? | Your data processing terms with Microsoft do not cover a third party | Microsoft states Graph data is not used to train foundation models. |
| Where is Microsoft 365 content processed and how long is it kept? | Residency commitments stop at the Microsoft boundary | EU Data Boundary, with named exceptions. |
| Can admins block Claude routing per user or group? | Anthropic processing may need its own approval | Disable Anthropic as a subprocessor, per user or security group. |
| Do agent actions appear in Purview or the Microsoft 365 audit log? | An audit trail inside Google alone may not satisfy your eDiscovery duties | Copilot interactions retained in mailboxes and searchable by eDiscovery. |
The right-hand column is drawn from Microsoft Learn pages updated between June and October 2026. The middle column is Zan Digital's assessment, not vendor text.
The identity question shapes every other answer, and the guidance on governing non-human identities for agents applies directly. The audit question decides whether your legal team can reconstruct what the agent did, which is the subject of the specification for what an agent audit log should record.
If I were running a Microsoft 365 tenant this quarter, I would not block the preview and I would not connect it to production mail. I would run it against a test tenant with seeded data and read the Entra sign-in logs myself.
Frequently asked questions
Does Gemini work inside Microsoft 365?
Yes, according to Google. Its Gemini at Work 2026 announcement says the Gemini agent can be accessed from Microsoft 365 and Slack as well as Google Workspace, and can connect to Microsoft Office and Teams. Google has not published which Microsoft 365 apps are covered, or whether access runs through an add-in, a Teams app or a Graph connection. Treat the integration as announced, not documented, until Google publishes admin guidance.
Is the Gemini agent for Microsoft 365 available now?
Not broadly. Independent coverage describes the Gemini agent as limited to enterprise customers in a private preview. Futurum Group reports that Google expects general availability by the end of October 2026, starting in North America. That date does not appear in Google's own launch post, so check with your Google account team before planning a rollout or a renewal decision around it.
How much does the Gemini agent cost?
Google's launch post gives no price for the Gemini agent. Futurum Group reports that Google will not charge additional seats for it where Gemini Enterprise is available. Google does state that admins can set a hard project spend cap in the Cloud Billing console that pauses the agent. Whether routing work to Claude costs more than Gemini has not been disclosed.
Does Microsoft 365 Copilot use Claude models?
Yes. Microsoft documents Anthropic as a Microsoft subprocessor and enables Anthropic models by default for most commercial tenants, excluding the EU, EFTA and UK, where they start disabled. AI Administrators or Global Administrators can turn Anthropic off in the Microsoft 365 admin center or limit it to specific groups. Microsoft states Anthropic models are currently excluded from the EU Data Boundary.
Is Gemini a good Microsoft 365 Copilot alternative?
Not yet for most Microsoft 365 organisations, because it cannot be evaluated against published controls. Copilot documents its permission model, training policy, residency exceptions and audit retention. Google has documented agent identity, audit trails and spend caps at the platform level, but not how its agent connects to a Microsoft tenant. Revisit the question when Google publishes admin documentation and a price.
What permissions does the Gemini agent need in Microsoft 365?
Google has not said. Any third-party agent reading Microsoft 365 data needs consent in Microsoft Entra, either delegated from a user or granted to the app by an admin. By default Entra lets users consent to permissions that do not require admin approval, and Microsoft recommends limiting that to verified publishers. Review that policy before the agent reaches your tenant.
How to prepare your Microsoft 365 tenant this week
Open the Microsoft 365 admin center and check whether Anthropic is enabled as a subprocessor for Copilot in your tenant. If you are outside the EU, EFTA and UK, it probably is, and you should decide that deliberately rather than inherit it.
Then open Entra and read your user consent setting for applications. If any user can consent to any app, tighten it to verified publishers with an admin consent workflow. That one change governs how Gemini, or any agent that follows it, can enter your tenant.
Related in this series
For the seat and suite pricing behind this decision, read what the $30 Copilot seat buys against Gemini in Workspace.
References
- Google Cloud Blog, Thomas Kurian, Welcome to Gemini at Work 2026: Introducing the Gemini agent, 9 October 2026. Used for Microsoft 365 wording, model routing, identity, audit, sandbox, gateway, spend caps and the Fortune 100 claim.
- Google, Google Cloud launches Gemini agent, 8 October 2026. Used for announcement date and description.
- Futurum Group, Nick Patience, Google Gemini Agent Puts the Agent Ahead of the Model, 9 October 2026. Used for reported GA timing, seat pricing and admin model selection, labelled as analyst reporting.
- Android Authority, Google unveils universal Gemini agent for enterprise users, 8 October 2026. Used for private preview status.
- Microsoft Learn, Data, privacy, and security for Microsoft Copilot, updated 8 October 2026. Used for permissions, training, residency and sensitivity labels.
- Microsoft Learn, Anthropic as a Microsoft subprocessor, updated 23 October 2026. Used for default-on regions, admin switch, EU Data Boundary exclusion and dated settings.
- Microsoft Learn, Manage public web access in Copilot, updated 18 August 2026. Used for web search policy and its data boundary exception.
- Microsoft Learn, Configure how users consent to applications, updated 4 August 2026. Used for Entra default consent behaviour and Microsoft's recommendation.
The weakest part of the source base is the Gemini side: one keynote adaptation and analyst reporting, with no Google admin documentation found as of 23 October 2026. Every Gemini claim beyond Google's own post is labelled as reported. Re-check this post when Google publishes help-centre pages for the agent.
Related reading