Microsoft recently published guidance on building an Agentic Center of Excellence (CoE) as organizations move from experimenting with AI agents to deploying them more broadly across the enterprise.
As organizations move beyond copilots and begin deploying agents that can reason, orchestrate tools, and participate in business processes, many are asking whether they need an Agentic CoE. The bigger question may not be whether we need another CoE, but whether our existing CoE model is ready for a world where AI does more than assist people.
When I first went through the guidance, one question immediately came to mind:
Do we really need another CoE?
Many large organizations already have Centers of Excellence around Cloud, Microsoft 365, Power Platform, Security, Data and AI. More recently, Copilot has also become part of the CoE conversation.
Having worked with an M365 CoE model, where we incubated new requirements and projects, brought together stakeholders across technology and business teams, evaluated emerging capabilities, and more recently focused heavily on Copilot evaluation and adoption, much of the Agentic CoE concept initially felt familiar.
But there is an important difference.
The fundamentals of a good CoE haven't changed.
What is changing is the autonomy, reach and responsibility of the technology being introduced into business processes.
Unlike traditional applications, agents can reason over information, invoke tools, interact with systems and potentially perform actions on behalf of users. As a result, governance, ownership and operational oversight become significantly more important.
What Does an Agentic CoE Actually Do?
Microsoft describes the Agentic CoE around four core functions:
Govern → Enable → Optimize → Scale
The objective is to help organizations move from isolated agent experiments toward a repeatable enterprise capability.
That means bringing together stakeholders across business, technology, architecture, security, data, Responsible AI, and platform teams rather than allowing every agent initiative to develop independently.
This part isn't necessarily new.
A mature M365 CoE already works across organizational boundaries.
When a new capability comes in, the job isn't simply to enable a feature. We first need to understand the requirement, identify the business value, evaluate the technology, involve the right stakeholders and determine how it fits into the wider enterprise environment.
Copilot reinforced that approach.
Agents take that evolution one step further.
From M365 CoE to Agentic CoE
In an M365 CoE, a typical incubation might start with:
Business requirement → Technology evaluation → Stakeholder alignment → Architecture → Pilot → Enterprise implementation
The CoE provides a place where something new can be evaluated before it becomes another enterprise service or capability.
That model worked well when evaluating capabilities such as Microsoft 365 Copilot.
But consider what happens when the requirement changes from:
"Can Copilot help our employees find and summarize information?"
to:
"Can an agent perform part of this business process for us?"
The conversation immediately becomes broader.
Consider an employee onboarding scenario.
A Copilot might help a manager summarize onboarding guidance, locate documentation, or answer questions about the onboarding process.
An agent, however, could coordinate the process itself by requesting accounts, initiating equipment provisioning, assigning mandatory training, updating HR systems, and tracking completion status across multiple platforms.
The conversation immediately moves from productivity assistance to business process execution.
That is where governance, architecture, security, lifecycle management, and operational accountability become significantly more important.
An agent might retrieve information, reason across different sources, invoke tools, communicate with other agents or systems and potentially perform actions as part of a business process.
So while the CoE model remains familiar, the scope of the technology being incubated is changing considerably.
What Changes When We Move to Agents?
This distinction is important.
A Copilot-focused CoE is largely about enabling people to work more effectively with AI.
The human remains at the center:
Human → Copilot → Better outcome
An Agentic CoE needs to consider scenarios where agents participate more directly in accomplishing the outcome:
Human → Agent → Tools / Systems / Data → Outcome
And increasingly:
Agent → Agent → Tools / Systems → Outcome
That doesn't mean humans disappear from the process.
It means we need to start thinking about where humans need to remain involved and where agents can take on parts of the work themselves.
That is a meaningful evolution from the Copilot conversation.
What About an Existing AI CoE?
This raises another obvious question.
If an organization already has an AI CoE, why establish an Agentic CoE?
I don't think the answer should automatically be to create another organizational structure.
An AI CoE may already cover AI strategy, architecture, Responsible AI, data, models and enterprise AI standards.
Similarly, an M365 or Power Platform CoE may already provide platform expertise, incubation and stakeholder engagement.
The Agentic CoE can build on these capabilities.
What it introduces is a stronger focus on agents as participants in business processes.
That includes understanding how agents are designed, where they operate, how they interact with enterprise platforms and how multiple agent capabilities can be scaled across the organization.
So I would see the relationship more like this:
AI CoE
Enterprise AI strategy and standards
↓
M365 / Power Platform / Application Platforms
Platforms where AI experiences and agents are delivered
↓
Agentic CoE
A repeatable model for turning agent opportunities into enterprise capabilities
These don't necessarily have to be three different teams.
They can be different capabilities within the organization's wider technology operating model.
Don't Start by Creating Another Team
This is probably the part of Microsoft's guidance that resonates most with me.
An Agentic CoE doesn't have to mean creating a new department with another organizational chart.
Microsoft describes different operating models including centralized, hybrid and federated approaches.
That matters because every organization starts from a different level of maturity.
For an organization just beginning its agent journey, a small centralized team may make sense. Expertise is scarce and concentrating it helps establish patterns and reusable approaches.
As adoption grows, a hybrid model becomes more practical.
The central CoE provides expertise and common patterns while business and technology teams start building solutions closer to their own domains.
Eventually, mature organizations may operate in a more federated way, where individual teams can deliver agents while the CoE focuses more on shared capabilities, guidance and enterprise alignment.
That evolution feels very similar to what many of us have already seen with cloud, Microsoft 365 and Power Platform.
The technology changes.
The CoE matures with it.
The CoE Should Help Decide Whether You Even Need an Agent
One area I think will become increasingly important is technology choice.
Not every business requirement needs an AI agent.
Sometimes a Power Automate flow is enough.
Sometimes a traditional application is the better solution.
Sometimes Microsoft 365 Copilot already provides what the user needs.
Sometimes a Copilot Studio agent is appropriate.
And some requirements may justify a more sophisticated agent built using Microsoft Foundry or other development frameworks.
A good Agentic CoE shouldn't exist simply to create more agents.
It should help answer:
What problem are we trying to solve, and is an agent actually the right solution?
That is exactly the kind of conversation a CoE should facilitate.
Otherwise, "agent" risks becoming another technology label attached to every new requirement.
From Projects to Reusable Agent Capabilities
Another interesting aspect of the Agentic CoE model is the opportunity to avoid rebuilding the same capability repeatedly.
Imagine several business units independently needing agents that:
- access similar enterprise knowledge
- interact with the same internal systems
- perform similar employee-support activities
- connect to common services
- use the same enterprise capabilities
Without a CoE approach, each project could solve those problems independently.
A CoE can identify common patterns and reusable capabilities that make the next implementation faster.
This is where the value starts moving beyond individual agent projects.
Instead of:
Requirement → Build Agent → Finish Project
the model becomes:
Requirement → Evaluate → Build → Learn → Reuse → Scale
Over time, the organization builds an agent capability, rather than simply accumulating individual agents.
Agents Should Be Treated More Like Products Than Experiments
This is another part of Microsoft's guidance that I think is particularly important.
Early AI initiatives naturally start as experiments.
We try something.
We evaluate the output.
We demonstrate the capability.
And then we move to the next use case.
That approach doesn't work once agents become part of real business processes.
An enterprise agent needs an ongoing purpose and ownership. The underlying business process can change. The systems it interacts with can change. New capabilities become available.
That means the mindset needs to move from:
"We built an agent."
to:
"We operate an agent capability."
Agents should be treated more like products than projects. They need defined owners, success metrics, lifecycle management, support models, performance monitoring and continuous improvement roadmaps. As business processes evolve and underlying systems change, agents must evolve as well.
This is one of the biggest mindset shifts organizations will need to make as they move from experimentation to enterprise-scale adoption.
For anyone coming from a service or product operating model, this is a familiar transition.
The technology might be new.
The need for clear ownership and continuous improvement isn't.
Where I See the Agentic CoE Fitting
Looking at Microsoft's framework through my own experience with an M365 CoE, I don't see Agentic CoE as replacing what organizations already have.
I see it as an evolution.
An M365 CoE helped us evaluate and incubate new platform capabilities.
Copilot expanded that conversation into AI-assisted work.
Agents now extend it further into AI participating in the execution of work.
That progression can be thought of simply as:
Traditional Technology
Applications help people perform work.
↓
Copilot
AI helps people perform work.
↓
Agents
AI can participate in performing the work.
↓
Agentic Organization
People and agents increasingly work together across business processes.
And somewhere between those stages, organizations need a mechanism to decide what to build, bring the right stakeholders together, establish repeatable patterns and turn experimentation into an enterprise capability.
That is where I think the Agentic Center of Excellence fits.
Final Thoughts
After going through Microsoft's Agentic CoE guidance, I don't think the most useful question is:
"Should we create another CoE?"
A better question is:
"Is our existing CoE model ready for agents?"
For organizations that already have mature Microsoft 365, Power Platform or AI Centers of Excellence, the answer may not be another standalone team. It may simply be the next evolution of capabilities that already exist.
The experience gained from governing cloud platforms, incubating Microsoft 365 services, enabling Power Platform solutions and driving Copilot adoption doesn't suddenly become irrelevant. In many cases, it becomes the foundation for what comes next.
The difference is that agents introduce a new dimension. We are no longer focused solely on how technology helps people perform work. We are beginning to explore how technology can participate in the execution of work itself.
That shift requires stronger governance, clearer ownership, reusable architectural patterns and a disciplined approach to scaling successful implementations.
Viewed through that lens, an Agentic CoE is not really about creating another team.
It is about creating an organizational capability that can answer three critical questions consistently:
- What business problem are we solving?
- Is an agent the right solution?
- How do we scale successful patterns responsibly across the enterprise?
As organizations move toward more agent-enabled ways of working, the companies that succeed will likely be the ones that treat agents not as isolated experiments, but as enterprise capabilities that are governed, operated and continuously improved.
And for me, that is where the Agentic Center of Excellence becomes truly valuable.
Update (September 18, 2026): What Microsoft is Seeing Across 40,000 Agents
Microsoft recently shared insights from its analysis of 40,000 agents built with Copilot Studio, and it adds some useful real-world context to this discussion.
What I find interesting is that the conversation is clearly moving beyond simply building an agent. As organizations create more agents and start using them across business processes, questions around governance, ownership, quality, security and scaling become much more important.
This also reinforces why I see an Agentic CoE as an evolution of the existing CoE model, rather than just another governance layer. The challenge isn't only helping teams build agents. It is creating a repeatable way to take the right agents from experimentation into governed enterprise use.
Update (September 25, 2026): Microsoft Foundry Continues to Evolve
Microsoft Foundry is evolving quickly as organizations move from experimenting with agents to running them in production.
Recent updates bring broader model choice, including the GPT-6 family and Claude Opus 5.5, along with voice agents, support for long-running workloads and capabilities to continuously evaluate and optimize agents.
What I find interesting is that model choice is becoming less of a one-time architecture decision. Different agents may need different models depending on the task, performance and cost.
For an Agentic CoE, this adds another responsibility: helping teams decide not only where agents should be built, but also which models to use and how those agents should be governed, optimized and operated at scale.
References
Microsoft Learn — Build an agentic Center of Excellence
https://learn.microsoft.com/en-us/agents/center-of-excellence/
Microsoft Learn — Agentic AI adoption
https://learn.microsoft.com/en-us/agents/adopt-overview
Microsoft Learn — Agentic CoE operating models
https://learn.microsoft.com/en-us/agents/center-of-excellence/operating-models
Read the latest Microsoft Foundry updates below:
