AI agents often stay at the edges—chatbots, assistants, or isolated use cases.

But the real value starts when agents are embedded inside core business processes.

I attended Microsoft partner architect seminar in Hilton Kalastajatorppa. I had one session about developing Copilot Studio agents with Claude Code and then I participated other sessions. In the session “Agent Stories Part 3: Using Agents in Business Processes (Dynamics MCP)”, Erno Soinila and Jani Sahlman explored exactly this:
how agents can connect systems, support process execution, and bring automation directly into Dynamics 365 Finance & Operations scenarios.

From Assistants to Process Participants

A key shift highlighted in the session:

Agents are no longer just interfaces—they are becoming part of the process itself.

Instead of:

  • Users asking questions
  • Agents returning answers

We move toward:

  • Agents triggering actions
  • Agents interacting with systems
  • Agents supporting or executing process steps

This represents a clear evolution:

From chat → to process orchestration.

MCP: Opening Systems for Agents

At the center of this shift is MCP (Model Context Protocol).

MCP enables agents to:

  • Connect to business systems (like ERP and Dynamics)
  • Access data and operations
  • Execute actions through controlled interfaces

In practice, this means:

  • Dynamics APIs and services become accessible to AI
  • Business logic can be exposed securely
  • Agents can move beyond read-only interactions

Dynamics as a Connected Process Layer

The session showed how Dynamics 365 Finance & Operations fits into this architecture.

Agents can:

  • Read financial and operational data
  • Trigger transactions
  • Support decision-making

Examples include:

  • Financial queries and reporting
  • Order handling and validation
  • Process support across procurement and accounting

The important part:

Dynamics is no longer just a system users work in—it becomes a system agents work with.

Balancing Speed vs Accuracy

One practical challenge highlighted was performance.

When agents call systems directly via MCP:

  • Data is accurate and real-time
  • But response times can be slower

To solve this, a hybrid pattern is used:

  • Frequently used data is replicated (e.g., to AI Search)
  • Agents query replicated data for speed
  • Critical updates still go through MCP for accuracy

This introduces a trade-off:

Freshness vs speed—and solutions must balance both.

Architecture Pattern in Practice

The typical architecture looked like this:

  • User interface (often Teams or Copilot)
  • Copilot Studio as orchestration layer
  • MCP connectors for system access
  • Data replication for performance (e.g., AI Search, Fabric)

This layered approach allows:

  • Fast user interactions
  • Reliable system-level operations
  • Flexibility to extend across multiple systems

Combining Low-Code and Pro-Code

A strong theme continued from earlier sessions:

Real solutions combine:

  • Low-code tools (Copilot Studio, Power Platform)
  • Pro-code layers (Agent Framework, Python, APIs)

For example:

  • Complex workflows can be implemented in code
  • Exposed as agents to Copilot Studio
  • Used seamlessly by business users

This creates a powerful pattern:

Developers build the complexity → business users consume it through agents.

Multi-Agent and Workflow Flexibility

The session also showed how agent solutions can evolve:

  • Simple pipelines → AI-enhanced processes
  • Autonomous agents → limited scope ownership
  • Multi-agent architectures → complex orchestration

However, an important point:

Not every process needs a complex multi-agent system.

Many valuable solutions remain:

  • Simple
  • Focused
  • Integrated into existing workflows

Cost and Efficiency Considerations

A practical (and often overlooked) topic was cost.

Two optimizations discussed:

  • Reducing token usage with structured “skills”
  • Minimizing unnecessary system calls

These become critical when:

  • Scaling to large volumes
  • Running frequent automated processes

This highlights a key operational reality:

AI architecture is not just about functionality—it’s also about efficiency.

A Key Insight: Don’t Wait for Perfect Solutions

One subtle but important takeaway:

The ecosystem is still evolving rapidly.

  • MCP capabilities are expanding
  • New integrations appear frequently
  • Best practices are still forming

Which means:

Waiting for a “perfect architecture” will only slow you down.

Instead:

  • Start building
  • Learn from real use cases
  • Iterate continuously

Final Thoughts

Agent Stories Part 3 shows what happens when AI moves deeper into operations.

The shift is clear:

  • From assistants → to process participants
  • From isolated tools → to connected ecosystems
  • From experiments → to production workflows

And the key enabler:

The ability to connect systems, data, and logic through MCP.

For organizations working with Dynamics:

  • The opportunity is not just better interfaces
  • It’s rethinking how processes are executed

Because ultimately:

The real value of agents is not in answering questions—but in helping the business actually run.