Governing AI agents at scale: Lessons from our journey at Microsoft

Empowering employees and protecting your organization through agent governance

Welcome to the agentic frontier

Agents are expanding the frontier of enterprise AI. By creating tools that surface knowledge, take actions, and even reinvent workflows, organizations can apply the power of AI to business processes in new and innovative ways.

But this shift raises questions for business and IT leaders: How do you get the benefits of agents without putting your organization and employees at risk? How do you encourage citizen developers to create agents freely while maintaining control, security, privacy, and compliance?

At Microsoft Digital, the company’s IT organization, we’re putting practical governance structures in place to ensure our internal agents are useful, safe, and properly scoped. Through a deliberate strategy of empowerment with established guardrails, we’re unlocking the potential of agentic transformation while maintaining the trust that defines our work.

The AI maturity model and frontier transformation

Agentic AI has made a new operational model possible, one that blends machine intelligence with human judgment, creating AI-operated, human-led teams.

We call organizations that enact this model Frontier Firms.

As organizations move toward this new operational state, they progress from foundational AI assistance through escalating levels of agentic maturity and complexity. First, humans operate with help from an AI assistant like Microsoft 365 Copilot. Then, human-agent teams work together. But the future lies with humans leading teams of agent users: AI agents that perform core labor with relative autonomy.

Capturing the benefits of this model relies on many factors, but in our experience as Microsoft Digital, two main tenets are instrumental to a successful transformation:

1. Empowering employees and teams to create and experiment with their own agents
2. Properly governing those agents to protect the enterprise

It’s a balance. If you set agent builders free without the proper guardrails, you risk data overexposure, agent sprawl, and security vulnerabilities. However, being too restrictive about governance stifles individual imagination, workflow reinvention, and innovation that can come from agentic AI.

We’re here to help you find the right balance for your organization.

This guide shares what we’ve learned along the way. As you read, you’ll follow our journey as Customer Zero at Microsoft, and you’ll gain access to tips and resources that we’ve assembled to help you apply our expertise to your own agent governance practice.

Every organization is different, and your experience will differ from ours in terms of risk tolerance, technical capability, resourcing, and more. This guide highlights some principles and best practices you can apply to your own business context, needs, and objectives.

“At Microsoft, we’ve moved beyond envisioning the agentic future into operating within it every day,” says Brian Fielder, vice president of Microsoft Digital. “Our experience as Customer Zero gives us a unique perspective on what it takes to govern AI agents at scale, turning early lessons into proven practices that help organizations innovate with confidence.”

Now is the time to seize this opportunity. Follow along to start your own journey toward frontier transformation and capture the benefits of trusted, connected agentic intelligence.

Learn from our experience governing agents

Within Microsoft Digital, we’ve been acting as Customer Zero for frontier transformation by creating the tools, infrastructure, and processes that power agents at Microsoft.

Our goal is to make it easy for employees to engage with agentic tools freely and adaptably while maintaining safety and responsibility. The path to this objective relies on a three-pronged approach to governance:

• Embedded governance functionality: Agent creation and publishing tools should incorporate good guidance, governance, and guardrails out of the box, making agents people create essentially self-governing.
• IT oversight: This is a new space and a new way of working, so it isn’t feasible for all agents to self-govern at this point. As an IT organization, we fill gaps in governance through reviews and oversight. We establish risk-based policies around types of agents, exposure and sharing, and other pivots.
• User education: It’s almost impossible to predict every governance gap and need, so educating our users helps them avoid accidentally increasing risk. Our Agents at Microsoft team and individual change managers are the guides for these efforts. Employees can also refer to resources like Microsoft Learn courses and the Agent Builders SharePoint hub.

Throughout this journey, we’ve empowered our employees to create all kinds of agents, ranging from simple personal tools built by people working in every function, with every level of technical skill, all the way to AI-powered enterprise tools designed by professional developers for use across lines of business and even the entire company.

As part of the process, we’ve incorporated guardrails to ensure less technical employees are limited to tools that simply retrieve enterprise knowledge, such as SharePoint Agent Builder or Copilot Studio, while software engineers get the full power of any tool they need that can take action or automate workflows, including Microsoft Foundry and Microsoft 365 Agent Toolkit.

Over the course of this journey, we’ve learned valuable lessons about effective agent governance, including:

• How to build an impactful but flexible governance strategy
• Strategies for creating an AI-ready data ecosystem
• Ways to apply appropriate policies and controls for highly diverse agents
• Approaches for tracking the impact and value of agents

Chapter 1: Building your agent governance strategy

Thinking through your organizational needs and building a framework to govern agents

As we’ve incorporated agents into different aspects of our organization, we’ve also deepened their involvement in employees’ daily workflows and core business processes. Because of this, we’re diligent about the governance guardrails and policies that protect our organization.

We’ve accumulated a wealth of knowledge and insights in this area through our efforts governing Microsoft 365 Copilot. Based on this experience, some of the key priorities that we made sure to adhere to included:

• Effectively applying controls to ensure users and apps don’t get access to privileged information
• Preventing employees from creating agents that violate company policies
• Balancing the freedom for employees to share their creations with the need to prevent agent sprawl
• Delineating which agents are authoritative and applicable for enterprise functions and which ones are meant for employees’ own personal use.
• Inventorying agents to provide lifecycle management
• Securing and protecting confidential data while respecting our responsible AI principles: Fairness, reliability and safety, privacy and security, transparency, accountability, and inclusiveness
• Unlocking telemetry that enables us to govern agents effectively

By focusing on each of these dimensions, our governance team has centered its efforts on the value these agents provide to the company while also ensuring organizational safety and trust. To realize this value, we emphasize three key principles that help protect both our employees and the organization:

We incorporated elements of our tenant’s minimum bar for governance into how we secure agents. Those include Microsoft Purview Information Protection, a functional inventory, activity logging, lifecycle management, and the ability to properly isolate agents so that they don’t cross data boundaries.

Our overarching tenant governance strategy is to govern items like documents and data at the container level. However, within a SharePoint site, for example, the added functionality of agents demands that we introduce further controls like sharing limits, breadth of knowledge sources, agent metadata, and information about an agent’s behaviors.

Turning priorities into principles

To operationalize governance, we developed six principles that guide our approach to agents. They form the governance foundation for a wide matrix of agent creation and usage opportunities.

1. We ensure a strong data hygiene foundation so we can trust our data estate as employees build and use agents.
2. We empower employees to build personal agents that can access permitted services and data sources to help automate and accelerate their tasks.
3. We empower teams and lines of business to build agents with known lower-risk patterns to accelerate impact.
4. We provide a smooth release path for engineering teams to develop agents designed for enterprise functions so they can access all the services and sources they need. This includes the same software development lifecycle (SDLC) reviews and certifications as other enterprise software, which we outline in Chapter 3.
5. We accelerate innovation through agent and automation templates while maintaining an AI Center of Excellence (CoE) to help teams think through their opportunities.
6. We reimagine employee experiences and task execution to simplify and optimize productivity.

Securing control through agent lifecycles

As we strategized to operationalize good governance, agent lifecycles became one of our most crucial tools. We superimposed the enterprise lifecycle on top of these policies, with both user-based and attestation-based lifecycles.

This means we treat agents owned by individual employees like any other user app and delete them when they leave the organization. Meanwhile, we ensure that agents owned by teams have a lifecycle that’s defined by the tenant and tied to attestation, our internal enterprise SDLC, and accountability confirmations.

This approach helps us combat sprawl by eliminating agents that no longer serve a purpose. It provides a solid foundation for more fine-tuned, matrixed policies and practices.

Governing amid real-time technology acceleration

One recent development illustrates how the rapid advancement of AI technology requires us to stay ahead of policy for new features.

Model Context Protocol (MCP) adds new capabilities, but also new risks and challenges. It’s a simple standard that lets AI systems communicate with the right tools and data without custom integration work. Instead of building a new connection or API every time, teams plug into a common pattern.

That standardization delivers speed and flexibility, but it also changes the security equation. We’ve extended our security and governance practices to account for MCP servers.

Our practices and policies help us govern agents effectively in this new environment. First, we assess security across four layers: Applications and agents, the AI platform, data, and infrastructure. We establish a secure-by-default strategy by positioning every remote MCP server behind our API gateway and establishing practices for vetting, identity management, automation that slows agents at the right moments, context trimming, and server isolation.

As you define policies for governing your own agentic ecosystem, you can take inspiration from our process. Start by asking questions about what you want to accomplish and what you want to protect, then move on to establishing your most important priorities. From there, you can cement those priorities into policies.

Learning from our approach to agent governance strategy

Chapter 2: Establishing a solid data foundation for agent governance

Setting agents up for success using a secure, robust data foundation

Operating according to an escalating maturity model means we’ve done the foundational work to secure and govern our data estate for Microsoft 365 Copilot. Many of the same principles apply to agents, with the added complexity of incorporating additional data sources.

To lead these efforts, we established a cross-functional team of data professionals within our AI CoE. This team is mostly comprised of Microsoft Digital employees who support corporate functions like Corporate, External, and Legal Affairs (CELA) and Global Workplace Services. Together with our AI CoE, this team helped us define what it means to have AI-ready data.

In essence, AI-ready data just means information we’ve certified for AI workloads. We certify those data sources using Microsoft Purview to identify defects in our core data products, and we’ve also built AI-powered assessments to certify which data lakes are AI-ready.

In most ways, governance is tool-agnostic and rooted in basic principles. With robust data labeling, data hygiene, and permissions in place alongside our AI tools, which respect labels by default, we can confidently give every employee the ability to build basic agents and trust in our governance guardrails. For decades, the challenge of data analysts and engineers was maintaining a consistently reliable source of truth despite inconsistent data quality, insufficient governance, and years of collecting data in silos. Microsoft Fabric and Microsoft Purview can help resolve these issues.

We’re embracing a more balanced, federated approach to data management today. We call this approach a data mesh. Rather than allowing unchecked decentralization or forcing all our data into a single centralized system, the data mesh formalizes domain ownership while embedding governance, quality, and interoperability directly into shared platforms.

With this approach, our domain teams publish data as well-defined, discoverable products, while common standards for security, metadata, and compliance are enforced through automation rather than manual processes. This model preserves enterprise trust and consistency without sacrificing speed or autonomy. By adopting a data mesh mindset, we can scale analytics and AI more effectively across the organization while still keeping ownership closely connected to the business focus.

Confidentiality labels, the practical framework for data protection

To operate according to Zero Trust principles, we needed a coherent system that lets us see, label, and protect data. Otherwise, the burden of data loss prevention would fall solely on employees, who would have to exercise individual discretion whenever they decided how to house and share potentially sensitive content.

With labeling, it’s important to strike a balance between the depth necessary for supporting an array of data governance controls and the simplicity to ensure labeling isn’t burdensome for users.

We decided on four overarching labels for container and file classification, each with its own sub-labels. The highest-level schema looks like this:

1. Highly confidential: We only share our most critical data with named recipients.
2. Confidential: Any items crucial to achieving our goals feature limited distribution.
3. General: Employees can share daily work–like personal settings and postal codes–internally throughout Microsoft.
4. Public: We share unrestricted data meant for public consumption freely. That includes information like publicly released source code and openly announced financials.

For our risk tolerance and organizational needs, we made the decision to protect data designated confidential or higher. As a result, we contain data flows to their tenants and only trust suitable storage destinations for content. That suitability depends on a storage location’s ability to gate which connectors can work with particular source data and sensitivity labels.

The administrators responsible for workspaces like SharePoint sites set default labels. These labels serve as a foundation for appropriate access and circulation for objects within those containers. It takes the burden of labeling off of employees. The sensitivity labels that administrators apply map to several different categories of policies that can anticipate and help to mitigate data loss and risk.

They communicate four key areas:

1. Breadth of availability: Labels determine whether the workspace is broadly available internally or is a private site.
2. External permissions: We administer guest allowance via the group’s classification, allowing specified partners to access teams when appropriate.
3. Sharing guidelines: We tie important governance policies to the container’s label. For example, can an employee share this workspace outside of Microsoft? Is this group limited to a specific division or team? Is it restricted to specific people? The label establishes these rules.
4. Conditional access: While we haven’t implemented this policy at Microsoft, tying identity and device verification to container labels can introduce additional governance controls.

Within Microsoft Digital, we’ve put a lot of thought into how each of our labels aligns with relevant policies. You can see more of the logic behind our sensitivity labels and their policies in this graphic:

If a container owner needs different policies for a set of files to provide greater external access, they can self-service new groups without accidentally violating our governance practices.

At Microsoft, we use Microsoft Purview, which is our suite of data estate management tools, but you can use your tool of choice to apply labels in your environment. Microsoft tools will respect them. Microsoft Purview helps us accomplish three important tasks: mapping our labeling structure onto the relevant policies, verifying them against our standards, and backstopping self-service data loss prevention practices through automation.

Automation is particularly useful. We’ve configured Microsoft Purview Information Protection to scan automatically for wayward credentials, malicious user behaviors, and other sensitive information in items without the proper protections. When Purview detects a violation, our governance team receives alerts that prompt them to contain the risk by upgrading an item’s sensitivity label or requiring employees to remedy the issue.

The result is a system that allows flexibility for employees to self-manage their digital workspaces while providing guardrails that help our governance experts take appropriate actions without overtaxing their time and resources.

Our approach within Microsoft Digital is just one way to create an AI-ready data estate, but aspects of our story will hold true for almost any organization. Consider establishing a body to take over responsibility for AI-ready data, developing your primary goals for AI-ready data, unifying your data estate, and implementing a system of confidentiality labels.

Learning from our approach to agent governance strategy

Chapter 3: A matrixed approach to agent governance

Governing different types of agents for different contexts, built with different toolsets

Our customers have expressed a strong desire to start building agents, but they’re concerned about where to begin and how to manage those agents once they’re built. They worry about persistent problems such as hallucinations and agent sprawl. These concerns are especially pronounced on IT teams.

During our Customer Zero journey, we’ve learned that the diversity of agent types and creation methods means there’s no one-size-fits-all approach to governance. Generalized approaches will only get you so far.

We’ve found it helpful to think about different kinds of agents along an escalating spectrum of development complexity:

There’s an entire matrix of different parameters that apply to an agent at any level of this spectrum, and they all require different policies. Those parameters include:

• Level of reach: Personal agents, limited sharing (like development environments or team boundaries), or enterprise-wide distribution
• Agent-building tool: SharePoint agent builder, Agent Builder in Microsoft 365 Copilot, Microsoft Copilot Studio, or tools geared to more professional developers (such as Microsoft Foundry or Microsoft 365 Agent Toolkit)
• Knowledge sources and content accuracy: Public sites, SharePoint and OneDrive, directly uploaded files, enterprise apps and systems, or third-party knowledge bases

Each of these parameters creates a pivot that we need to govern, and we’ve carefully assembled a set of policies and controls to account for them. As our understanding and use of agents advances, we’re continually updating how we match their characteristics and capabilities with relevant policies and any applicable reviews.

Within Microsoft Digital, we’ve adopted a risk-based approach that helps us establish a matrixed model for agent governance. The foundational idea is that we identify potential harms for each kind of agent, then assign policies for the level of review and oversight they require.

For example, simple agents that can only read and present data tend to be low risk. Because their access is tied to their creators’ identities and access, our data governance structures and guardrails can prevent overexposure. But for agents that have capabilities like writing data, taking action, or creating items, more reviews are necessary.

A matrix of agent governance policies, pivoted by parameter

The following matrix enumerates the factors that determine how we govern different kinds of agents created using different tools. This matrix helps our employees understand the agent creation process and helps us maintain safety and control.

In addition to mapping out our policies for governing agents, the matrix illustrates how we see their relative utility across the organization. It demonstrates an escalation from personally useful to organizationally useful agents. Their governance policies and controls escalate accordingly.

Regionality is an additional concern. Regulatory compliance might vary, but it’s important to keep in mind that certain kinds of data access and actions might be perfectly permissible in one region, but not in another.

One example is our Employee Self-Service Agent, a central resource employees can turn to for help with IT support, HR questions, and facilities requests. Because it can access potentially sensitive personal information, this agent required additional review from European works councils to ensure it met all relevant workplace standards.

As you facilitate the experimentation and innovation with agents across your workforce from citizen developers to pro developers, consider adopting a similar matrixed approach to agent governance. It starts with understanding your organization’s needs, your risk tolerance, and the different employee populations you want to equip with agent-building capabilities.

Learning from our matrixed approach to agent governance

Chapter 4: Tracking, impact, and value

Managing agents and assessing their business impact for the organization

It’s clear that agents bring astonishing capabilities to the enterprise. For many organizations, what remains unclear is exactly how to measure their impact. Without that information, businesses are at a loss for ways to articulate value and drive improvement.

Tracking agents is also a crucial component of preventing sprawl: We need to understand what agents we have, how employees are using them, what critical processes they’re supporting, and if they’re contributing value or need to be retired.

We’re at the beginning of our impact-tracking journey, but our work can provide a starting point for your own efforts to measure the value of AI initiatives at your organization.

Managing our agent catalog through comprehensive tracking

Microsoft Digital partners with other internal organizations to ensure we’re prioritizing the right agents and avoiding agent sprawl. Ideally, these engagements take place before teams start building their agents so we can avoid wasted effort or duplicated work.

Still, ongoing management efforts are crucial to keeping our agent ecosystem healthy. Telemetry is the key to assessing usage and ensuring compliance. We’ve developed our own internal tooling to ensure that:

• Metadata is complete and available
• The tooling tells us the right information about our agents
• The tools connect properly with other compliance tooling, like Microsoft Purview

This telemetry also reveals agent behaviors, shows how agents do their work, and tracks events, actions, and policy baselines.

These capabilities help us gain visibility into policy adherence and violations, and then to conduct enforcement actions. We also track the speed of reaction and mitigation. AI-ready data and robust guardrails mean we head off most violations before they occur.

A robust inventory, an agile policy framework, and an automated workflow for enforcement are cornerstones for successfully governing agents at scale.

The release of Microsoft Agent 365, now in early access, represents the next step in agent observability and management, two key aspects of agent governance and sprawl mitigation. This control pane for agents incorporates many of our learnings as we’ve bridged governance gaps through IT intervention.

Some of the key aspects of the control pane:

As Customer Zero for Agent 365, we’re excited to have a platform for observability and telemetry that encompasses everything from agentic creation through usage.

Tracking governance from agent inception

Professionally developed agents add a new dimension of tracking and governance, because we need standards in place for ensuring compliant agent-building and to remediate any issues.

We use our Azure DevOps instance to catalog apps on our tenant, and we’ve applied this practice to agents created professionally for lines of business and enterprise agents. This tool contains our service tree with product and app log registration, which is tied to our KPI dashboard and scoring system that validates agent data against our policies.

Our expectation is that all new apps and agents start from a place of compliance. Any new agent is registered through this platform, and we expect adherence within the first 14 days. In our experience, the introduction of new metrics, policies, or timeframes as our governance policies evolve is where agents tend to drop out of compliance. The priority is restoring compliant status.

We’ve established a series of metrics to help track and manage these expectations:

• Enablement velocity
• Renewal velocity
• Agents in compliance
• Time to remediation of noncompliance

Through a DevOps process built on our preexisting software development lifecycle practices, we’ve applied governance not only to agents themselves, but to the process of building them professionally.

Measuring progress and unlocking value

Properly measuring value depends on concrete definitions of success and metrics that support it. Articulating AI’s impact came with several challenges. First, we had to land on a consistent taxonomy for different measurement areas. Then we needed to make the relevant data accessible, ensure its quality, and confirm it made sense.

The Microsoft Digital AI Value Framework is our flexible, modular tool for measuring the impact of our AI initiatives. With tools for measurement firmly in place, we can effectively demonstrate value and guide further decision-making.

We plan to use the following capabilities to improve the overall ecosystem:

• Filtering our agent inventory on specific criteria like the type of agent or how it was built
• Enhancing governance-specific actions we can take with agents in areas like ownership and quarantining
• Gaining visibility into trends like agent usage
• Ingesting agent blueprints and defining policy templates

We’re still in the midst of our agentic measurement journey at Microsoft, but the blueprint for tracking already exists. Your organization might be in the early stages of agent readiness and deployment. If that’s the case, it could be helpful for you to internalize the lessons we’ve learned as Customer Zero and apply them as early as possible in your own journey toward AI maturity.

Learning from our agent adoption experience

Governing the frontier to scale innovation

AI agents are rapidly becoming core contributors to how work gets done. As our experience within Microsoft Digital demonstrates, realizing their full potential demands more than powerful tools or enthusiastic builders. It requires thoughtful governance that evolves alongside your AI maturity, protects what matters, and gives employees the confidence to innovate responsibly.

As you consider your own strategy for managing agents, it can be helpful to keep one truth in mind: Governance is a catalyst for progress, not a barrier. By embedding guardrails into tools, grounding agent creation in AI‑ready data, applying risk‑based and matrixed policies, and reinforcing all of it through adoption and education, we’ve been able to expand agentic capability without sacrificing security, privacy, or trust.

From our experience, we’ve learned that governance works best when it’s:

• Proportional, scaling with risk and agent complexity
• Embedded, not bolted on after the fact
• Human‑led, recognizing that accountability and judgment remain essential
• Iterative, adapting as technology, regulations, and business needs evolve

When you design governance this way, it allows experimentation, learning, and impact at scale. Employees feel empowered to build agents that solve real problems, while IT and compliance teams gain visibility and control without becoming bottlenecks. Crucially, leaders can measure value, manage risk, and make informed decisions about where to invest next.

This is the foundation of the Frontier Firm: Organizations where humans lead and agents operate, guided by clear principles and trusted systems.

As you continue your AI maturity journey, remember that there is no single, correct governance model. Your approach will reflect your risk tolerance, regulatory environment, data maturity, and organizational culture. The practices outlined here provide a proven starting point informed by real-world deployment at enterprise scale.

“At Microsoft, we believe the future of agentic AI depends on governance that empowers people first,” says Vijaya Alaparthi, principal group product manager in Microsoft Digital. “The structures should be invisible when they’re working, intentional when they’re needed, and trusted by everyone they serve.”

Now is the moment to act. Start with strong foundations. Empower your builders. Measure what matters. And treat governance not as a constraint, but as a strategic advantage that allows your organization to move faster, innovate safely, and lead confidently on the agentic frontier.

Get started building and managing agents at your company with Microsoft Agent 365.

Want more information? Email us and include a link to this story and we’ll get back to you.