Implementing Model Context Protocol in Dynamics 365 Business Central: Unlock AI Agents

What Is Model Context Protocol?

AI chat assistants and agents have matured from simple question‑answer bots into powerful tools that can manipulate business data, generate orders and trigger actions on behalf of users. Most organisations, however, still access ERP data through traditional REST APIs or custom connectors. These interfaces expose data but don’t tell the AI how the data is structured, what fields mean or which actions are allowed. That gap makes it hard for agents to behave responsibly without extensive custom code.

Microsoft’s Model Context Protocol (MCP) aims to solve this problem. MCP is an open standard that describes how AI applications should communicate with business systems. An MCP server sits in front of a system such as Dynamics 365 Business Central and exposes the system’s tables, pages and business actions as discoverable tools. AI clients — including Copilot Studio, GitHub Copilot, ChatGPT and other large language models — can connect to the server and ask for data or run operations through a consistent, context‑aware interface. Because the server describes relationships between tables and enforces permissions, agents can read and act on data with far less risk of hallucination or policy violations. MCP is essentially a bridge between your ERP and emerging AI agents.

Why Businesses Should Care About MCP

Operational leaders and ERP teams often struggle to get value from AI because it lives outside the day‑to‑day business processes. MCP changes that by letting AI agents interact with your core systems just like a user would. Once enabled, you can:

Expose business logic to AI assistants. Agents can create sales orders, post journals, update inventory records or run approval workflows without developers writing custom API wrappers. MCP provides the metadata and action definitions needed for the agent to choose the correct operation.
Maintain security and governance. Permissions, role‑based access and audit logs are respected. The same rights a user has in Business Central apply to the agent, reducing the risk of unauthorised actions. Administrators decide which tables and actions are available.
Reduce integration complexity. Instead of wiring up dozens of endpoints, MCP aggregates them into one discoverable interface. Agents can navigate relationships (customer → orders → line items) on their own. This lowers the cost of connecting AI tools to your ERP.
Enable cross‑system automation. With Azure Logic Apps or other MCP‑compliant platforms, you can create remote MCP servers that expose workflows from multiple systems — ERP, CRM, e‑commerce and custom applications. Agents can then orchestrate processes across these systems through a single protocol.

For companies already investing in AI‑powered CRM and e‑commerce experiences, MCP is the missing layer that ties those experiences back into the operational backbone.

Preparing Your Dynamics 365 Business Central Environment

Before turning MCP on, confirm that your Business Central environment is on version 27 or later — earlier versions do not support the protocol. Upgrading can involve schema changes and testing; treat this as a mini‑project rather than a last‑minute toggle.

Once you are on a compatible version, follow these steps to enable MCP:

Enable the MCP feature. In Feature Management, look for “Enable MCP Server access” and toggle it on. This registers the MCP server with your environment.
Create an MCP configuration. Search for MCP Configurations and click New. Give the configuration a meaningful name (e.g., “SalesAgent”). Use the option to Add all standard APIs as tools to expose common tables such as Customers, Items and Sales Orders. Remove any tables you do not want to expose. Pay particular attention to sensitive entities — you may not want agents seeing payroll or vendor bank details.
Set permissions. By default, MCP is read‑only. To allow agents to create or modify records, assign create/modify/delete permissions to the relevant API page objects. Limit write permissions to the smallest set of operations necessary; this reduces the risk of unwanted changes.
Activate the configuration. Mark the configuration as Active so it becomes available to MCP clients. You can create multiple configurations (for example, separate ones for Sales, Purchasing and Inventory) to control what each agent can access.
Review data quality. AI agents depend on clean, consistent data. Before exposing tables, audit your master data for duplicates, missing fields and mismatched identifiers across systems. Agents will struggle if your item numbers differ between your ERP and your e‑commerce platform or if customer addresses aren’t standardised. Fixing these issues up front avoids confusing the AI later.

Connecting MCP to Copilot Studio for AI Agents

Once your configuration is active, it’s time to connect MCP to an AI agent. Copilot Studio provides a no‑code environment for building conversational agents and is currently the simplest way to test MCP. Here’s a practical flow:

Create or open an agent in Copilot Studio. Navigate to the Tools section and click Add a tool.
Search for the Business Central MCP connector (preview). Choose Create connection and sign in with your Business Central credentials. You’ll be prompted to select your environment, company and the MCP configuration you created earlier.
Choose a credential model. Copilot Studio offers several options:

Maker‑provided credentials — good for internal testing but risky in production because all users operate under the maker’s permissions.
Prompt the user for permission — the agent asks the end user to sign in when needed. This builds transparency but interrupts the flow.
Delegated access (recommended) — the agent uses the end user’s identity, respecting their permissions. If a user cannot see a particular table, the agent can’t either.

Test your agent. Use the Test pane to ask questions like “Show me all open sales orders for customer X” or “Create a new sales order for customer Y with item Z and quantity 10.” Copilot Studio will invoke the MCP tools behind the scenes, assemble the right API calls and return structured data. Pay attention to how the agent interprets your prompts; if it misidentifies a field or picks the wrong tool, adjust your prompt or refine the configuration.

Pitfalls and Handoff Issues

During real implementations, several challenges often surface:

Ambiguous entity names. If multiple entities share similar names (e.g., “Customer” exists in both CRM and ERP), agents may select the wrong one. Use clear naming conventions in your MCP configuration and limit the scope of each agent.
Permission errors. Delegated access means the agent inherits the end user’s rights. If the user lacks permission to create sales orders, the agent will fail. Align agent capabilities with your security model and educate users about required roles.
Data mapping mismatches. Agents rely on standard field names. Custom fields or extensions may not be automatically exposed. You can add custom APIs to MCP, but you need to document them clearly and ensure they align with your AI use case.
Exception handling. Agents may not handle partial successes gracefully. For example, if an order line fails due to insufficient stock, the agent might abandon the entire order. Build post‑processing rules or fallback prompts to capture these exceptions.
Over‑exposure of data. Adding all standard APIs might expose financial or HR data that agents don’t need. Remove unnecessary tools and establish review processes to avoid accidental data leaks.

Extending MCP Through Azure Logic Apps for Cross‑System Workflows

The built‑in MCP server for Business Central exposes only Business Central data. Many organisations, however, need agents to orchestrate processes across multiple systems — for example, creating a sales order in Business Central after a customer completes a purchase on Shopify and simultaneously updating the CRM. Azure Logic Apps (Standard) now supports creating remote MCP servers from your workflows. Here’s how this works in practice:

Create or open a Standard Logic App. In the Azure portal, navigate to your logic app and select the Agents menu. You’ll see a new option to create an MCP server.
Add workflows as tools. Choose existing workflows that handle processes such as order creation, customer onboarding or inventory updates. For each workflow, define metadata that describes what the tool does and what parameters it requires. This metadata helps AI agents discover and call the right workflow.
Configure authentication. Logic Apps supports OAuth and key‑based authentication. For production use, set up Easy Auth and restrict which client applications and identities can call your MCP server. Generate API keys or require tokens from Microsoft Entra ID (formerly Azure AD) to secure access.
Test with an MCP client. In Visual Studio Code or Copilot Studio, register the remote MCP server’s URL and call the tools. You can now ask an agent to execute an entire workflow — for example, “Process this return and issue a refund” — and Logic Apps will orchestrate the calls to your ERP, payment gateway and CRM.
Plan for scalability and governance. When you group multiple MCP servers in a single Logic App, you gain flexibility but also need to manage deployment pipelines, versioning and monitoring. Use tagging and logging to track which agent used which tool and to troubleshoot errors. Establish approval steps in workflows so that high‑risk actions (like issuing refunds) require human confirmation.

Remote MCP servers allow Semantic Lines to build AI‑driven automations that span ERP, CRM and e‑commerce platforms without writing brittle custom integrations. They also let you reuse existing Logic Apps workflows as agent tools, accelerating time to value.

Implementation Challenges and Lessons Learned

Working with clients across manufacturing, distribution and retail, we’ve observed several themes when implementing MCP and AI agents:

Clean master data is non‑negotiable. AI agents amplify existing data problems. Duplicate customer records or inconsistent item codes confuse the agent and produce erroneous suggestions. Investing time in data cleansing and establishing a single source of truth pays dividends.
Keep your scope focused. It’s tempting to expose every table and workflow, but broad scope increases risk. Start with a narrow use case (such as viewing open sales orders) and expand only after user feedback. This also limits the cognitive load on your operations team when something goes wrong.
Design approval flows up front. Autonomous agents can create orders, issue refunds or modify inventory — all high‑impact actions. Implement approval gates within workflows (e.g., require finance approval for orders over a certain amount). This prevents silent errors and builds trust among stakeholders.
Educate your teams. Operations and finance users may assume the AI will replace them; in reality, it augments their work. Train users on how to prompt the agent, what to expect, and when to intervene. Encourage them to provide feedback so you can refine prompts, tools and workflows.
Monitor and iterate. Treat AI agents like any other production application. Use telemetry to track usage, response times and error rates. When an agent behaves unexpectedly, dig into the prompt and tool it chose. Continuous improvement is key to maintaining reliability.

Practical Use Cases for MCP‑Enabled AI Agents

To illustrate the potential of MCP, here are three scenarios we’ve helped clients prototype:

AI Sales Assistant

A sales agent built in Copilot Studio connects to Business Central through MCP, reads customer history, suggests complementary products and creates draft quotes. It hands off the quote to a salesperson for review. Permissions ensure the agent can’t override credit limits or discount thresholds without approval.

Financial Copilot

Accounting teams often juggle late payments and cash flow forecasting. An AI agent can analyse open invoices, flag those past due, draft reminder emails and even schedule collection tasks. When integrated with CRM, it updates customer notes so collections and sales teams stay aligned.

Inventory Optimisation Agent

For companies with both e‑commerce and physical stores, aligning stock levels is a daily headache. An MCP‑enabled agent monitors sales velocity, predicts demand and automatically creates purchase orders when thresholds are met. If the purchase needs approval from the supply chain manager, the agent routes an approval request via email or Teams before posting the order.

These examples demonstrate that MCP is not just a developer curiosity; it’s a practical tool for reducing manual work and empowering business users to harness AI safely.

Conclusion

The introduction of the Model Context Protocol in Dynamics 365 Business Central represents a major leap forward in making AI truly operational. By exposing your ERP’s data and actions through a context‑aware interface, you enable agents to deliver meaningful automation while respecting your governance and security. Preparing your environment, defining clear configurations, and extending MCP through Azure Logic Apps can unlock cross‑system workflows that were previously too complex or expensive to automate.

As with any enterprise technology, success depends on good data, careful scoping and ongoing monitoring. But for organisations ready to embrace AI agents across ERP, CRM and e‑commerce platforms, MCP provides the foundation. If you’re exploring MCP or need help designing AI‑driven workflows, our team at Semantic Lines specialises in ERP and e‑commerce integrations and can guide you from proof‑of‑concept to production.