Agent Mode in Microsoft 365 Copilot

What is it? How does it work? What do you need to do to enable it?

Microsoft is expanding Microsoft 365 Copilot with a new capability called Agent mode. Instead of just responding to single prompts, Copilot can now work toward a goal over multiple steps and across multiple interactions inside Word, Excel, and PowerPoint.

Agent mode allows Copilot to remember context, apply feedback, and continue refining work rather than starting over each time. In a way, this changes Copilot from a reactive assistant into something that behaves more like a task focused agent.

Agent mode is currently beeing rolled out (different timeframes for different products) through the Microsoft 365 Copilot Frontier program. That means that you must activly sign up for the Frontier Program. Either for all users, or just a selection of users. For organizations in the EU, there are also additional approval steps due to the use of Anthropics Claude AI as a subprocessor.

This post explains what Agent mode does, how it works behind the scenes, and what administrators need to do to enable it safely.

What is Agent Mode?

With regular Copilot each prompt is handled on its own. It does not remember what you asked before, and every interaction starts fresh. Agent mode works differently. It keeps track of the conversation, understands the context, and continues working across multiple steps.

That means you can ask Copilot to review a document, give feedback, apply changes, review it again, and keep refining until you are happy. You no longer need to restate everything every time.

Behind the scenes, Agent mode relies on several capabilities working together rather than a single prompt response.

It maintains context across the task so Copilot understands what has already happened and what it is working toward. It breaks down complex requests into smaller steps and executes them in sequence and it integrates directly with Microsoft 365 apps allowing Copilot to make changes in Word, Excel, and PowerPoint as the task progresses.

Agent mode can use more than one AI model when it works on a task. Microsofts own models handle how Copilot works inside Word, Excel, and PowerPoint. In some cases, Copilot may also use approved third party models, such as Anthropics Claude, to help with more complex requests.

This can improve the quality of the results, but it also means that data may be processed by both Microsoft and Anthropic. That is why EU organizations must explicitly approve Anthropic before Agent mode can be used. I will come back to that later.

Performance and quality differences

I have discovered that in real world use Agent mode typically feels slower than regular Copilot, often taking several seconds longer due to the additional planning and execution it performs. That extra time is used to reason, plan, and execute multiple steps. These steps are done simultaneously. As an example, you can see the Copilot working on multiple pages in Word in parralel.

The upside of this is quality. Most users need fewer revisions. Instead of going back and forth several times, Agent mode often gets you where you want to be in one or two iterations.

If you care more about the end result than speed I would say that Agent mode is a clear improvement.

The Frontier program

The Microsoft 365 Copilot Frontier program is Microsofts early access channel for upcoming Copilot features. Capabillities released through Frontier are close to general availability but still under active development. Microsoft uses Frontier tenants to validate performance, usability, and real world behavior before full release.

Enrolling in Frontier is an explicit admin decision. It can be enabled for the entire organization or limited to specific users or groups. Frontier features come without service level guarantees and can change at any time, which makes the program best suited for testing, pilots, and controlled rollouts rather than production critical scenarios.

From a governance perspective, joining the Frontier program is an explicit decision. It is not something that happens automatically. Administrators choose whether to enable it for the whole organization or limit it to specific users or groups. This makes it possible to test new Copilot capabilities like Agent mode in a safe and structured way before broader adoption.

As always, you need to be an administrator to enable the Frointier program.

Head on over to the Microsoft 365 Admin Center. From here, locate Copilot in the left navigation, select Settings and then Copilot Frontier

In the right-hand side panel that opens, choose your settings. The default is “No access”. You can select either “All users”, of “Specific users”.

My recommendation is to start with a few selected “champions” that are aware of the program and understand that features can and most likely will change.

Extra requirements for EU tenants

If your tenant is located in the EU, there is one additional and mandatory step before Agent mode will work.

Agent mode uses Anthropics Claude AI for advanced reasoning tasks. From a GDPR perspective, this means that certain data processed by Copilot may be handled by Anthropic as a subprocessor. Because of this, Microsoft requires explicit approval before any data can be shared with Anthropic for EU based tenants.

The key reason this requires explicit approval is that Anthropic models used in Microsoft offerings are currently excluded from the EU Data Boundary and, where applicable, in country processing commitments. Because of this, Anthropic models are disabled by default for EU tenants and for customers in the UK. They are also not available at all in government or sovereign cloud environments, including GCC, GCC High, DoD, or other sovereign clouds.

As a result, Microsoft requires a deliberate opt-in before any data can be shared with Anthropic for EU based tenants. This approval is not optional and it is not enabled automatically when you turn on the Frontier program. It must be completed separately in the Microsoft 365 admin center by an administrator with the appropriate privacy or compliance role.

Until Anthropic is approved as an AI subprocessor, Agent mode will not function for EU users. In practice, this usually means that Agent mode does not appear at all in Word, Excel, or PowerPoint. There is no partial functionality. Either the approval is in place, or the feature is effectively blocked.

From a governance standpoint this makes sense. Microsoft treats AI subprocessors differently from feature enablement to ensure that data protection decisions are explicit, auditable, and owned by the right stakeholders. In many organizations, this also means that enabling Agent mode requires coordination between IT, security, and privacy teams rather than a single admin acting alone.

After clearing it with your organization, you might need to complete a DPIA (Data Protection Impact Assessment), this can be handled in the Microsoft 365 Admin Center. From here, locate Copilot in the left navigation, select “Settings“, “Data access” and then “AI providers operating as Microsoft subprocessors“

Read the information and then select your preference.

The consent can be revoked at any time by returning to the same panel in the Microsoft 365 admin center.

When all of this is done, the agents becomes available to the Frontier users in the Microsoft 365 Copilot App, among other places.

Agent mode makes Copilot more “intelligent”

Agent mode changes Copilot from something that responds to prompts into something that can actively support real work. Instead of starting over with every prompt, Copilot can keep context, remember decisions, and continue working toward the same goal. This reduces repetition, lowers friction, and leads to more consistent results.

If we use Excel as an example to showcase Agent mode you could say that with regular Copilot, you might ask for a summary of a dataset, then ask for a chart, then ask for adjustments. Each step requires new prompts and repeated explanations. With Agent mode, you can ask Copilot to analyze a workbook, identify trends, create visuals, and refine them based on feedback while keeping the full context in mind all throughout the conversation.

For example, you might ask Copilot to review sales data, highlight key patterns, build a pivot table, and create a chart suitable for an executive presentation. If you then ask for changes, different time periods, or a new layout, Copilot understands that this is part of the same task and continues from where it left off.

This is where Copilot starts to feel less like a feature and more like an assistant. Not just answering questions, but helping you move work forward in a more natural and efficient way. A bit like Work IQ, but for a specific task and not the whole organization.

Are you already in the Frontier program?

And what are your thoughts on allowing Anthropic as a subprocessor? I have heard very different opinions from clients, ranging from full confidence in the safeguards Microsoft has put in place, to strong hesitation driven by data protection, risk ownership, and internal policy.

There is no single right answer here. Decisions around AI governance are mandatory as they directly affect which capabilities your users can access and how work actually gets done.

Please let me know your thoughs in the comments.