Most founders and operators running lean businesses face the same quiet bottleneck.

The data exists. Revenue, inventory, outstanding invoices, vendor balances — it is all sitting inside an ERP system, updated in real time, technically accessible. But getting a useful answer out of it requires logging in, navigating menus, knowing which report to run, exporting a file, and then interpreting it.

That process takes 20 minutes if you know the system well. It takes longer if you do not.

Multiply that by how many times a week your team needs to answer a customer question, prepare a proposal, or check a cash position. The ERP becomes a data warehouse that nobody visits until they absolutely have to.

The result is decision-making based on memory and gut feel instead of live numbers.

This issue walks through a practical workflow for changing that — using Model Context Protocol (MCP) as the connection layer between AI agents and your ERP data. No rebuilding your stack. No full-time developer required.

The common assumption is that integrating AI with an ERP system is an IT project. Something that requires a budget, a vendor, a six-month timeline, and a dedicated integration team.

That assumption was accurate two years ago.

The problem was real. Every system spoke a different language. Connecting an AI model to a database required custom API development — writing code to translate between the model's input format and the database's output format. For each system. For each AI tool. Each integration cost between $3,000 and $15,000 to build, according to ERP implementation benchmarks from Panorama Consulting, and the result was brittle: if either side changed its API, the connection broke.

Small and mid-sized businesses could not absorb that cost or that maintenance burden. So they did not integrate. They kept logging in manually.

MCP changes the economics of this problem.

Model Context Protocol is an open standard developed by Anthropic and released in late 2024. Its job is to define a single, consistent way for AI models to communicate with external tools, databases, and business systems.

The analogy that works: MCP is to AI integration what USB-C is to hardware charging. Before USB-C, every device had its own cable. After USB-C, one standard port connects to everything.

Before MCP, connecting an AI model to your CRM, your ERP, and your accounting system meant writing three separate custom integrations. After MCP, any AI tool that supports the protocol can connect to any system that exposes an MCP server — through a standard interface, with no custom glue code.

By early 2025, the protocol had been adopted by OpenAI, Microsoft, Google DeepMind, Salesforce, and ServiceNow. It is no longer experimental infrastructure. It is becoming the standard connection layer for production AI workflows.

What this means in practice:

Microsoft launched an ERP MCP server for Dynamics 365 at Microsoft Build in May 2025. The initial release exposed 13 tools for Finance and Supply Chain Management. By late 2025, a dynamic version was in public preview that unlocked hundreds of thousands of ERP functions. The Business Central MCP server, covering customers, inventory, sales orders, and vendor balances, shipped as part of the 2025 Wave 2 release and became generally available in the 2026 Wave 1.

SAP Business One, Xero, and other platforms have followed with their own MCP servers. For platforms without a native MCP server, third-party connectors and open-source implementations have emerged.

The infrastructure now exists. The question is how to use it.

Here is a practical workflow for connecting an AI agent to your ERP data layer and making it usable for operators who are not data analysts.

This is not a technical implementation guide. It is a workflow design template — the structure you need to think through before you build anything.

Before connecting AI to your ERP, define the categories of questions you actually need answered on a recurring basis.

For most lean businesses, these fall into four buckets:

Cash and receivables: Outstanding invoices, overdue accounts, current cash position, payment timing.

Inventory and orders: Current stock levels, pending purchase orders, delivery timelines, low-stock alerts.

Customer and vendor status: Account history, payment terms, contact records, recent transactions.

Operational performance: Sales by period, margin by product, expense trends, comparison to prior periods.

Write these down. They become the agent's job description — the set of questions it needs to be able to answer reliably.

Once connected to your ERP via an MCP server, the agent's core function is natural language to data retrieval. The user asks a question in plain language. The agent queries the ERP through the MCP connection and returns a structured answer.

A practical example: a finance team member types, "Show me all overdue invoices over 30 days and the total outstanding balance." Without MCP, that is a manual report. With an MCP-connected agent, the query goes directly to the ERP and returns the data in seconds.

A more advanced example: an operations manager asks, "Compare operating expenses this quarter versus last quarter — which categories increased most?" The agent pulls both periods, runs the comparison, and returns a structured breakdown.

The intelligence layer here comes from the language model's ability to interpret imprecise inputs and map them to the correct data fields. This is where what practitioners call entity resolution becomes relevant — the ability to handle inputs that are close but not exact. If a user queries "outstanding balance for Nguyen Trading" but the ERP has the account recorded as "Nguyen Trading Co. Ltd.", a well-designed agent can resolve that match through semantic similarity rather than requiring an exact string.

This capability does not come from MCP itself. MCP handles the connection. The language model handles interpretation. The combination is what makes the workflow practical for non-technical operators.

This is the most important part of the workflow that most guides skip.

Not every output from an MCP-connected agent should be acted on without review. Define explicitly which outputs require human confirmation before any action is taken.

Review required before action:

Can proceed without individual review:

The principle: AI handles the data retrieval and initial interpretation. A human handles the judgment call and the action.

This distinction keeps the system practical without creating risk. An agent that can tell you your receivables position in 30 seconds is valuable. An agent that can modify your ERP records without approval is a liability.

Define the format before you build the workflow. An agent that returns unstructured text is less useful than one that returns a consistent format you can act on.

Practical output formats for ERP-connected agents:

The output format should match how the person receiving it will use it. A CFO reviewing a cash position summary needs different formatting than a sales manager checking on a specific account.

Define success before deployment, not after.

Useful metrics for ERP agent workflows:

An agent that nobody uses is a failed automation regardless of its technical sophistication. Track adoption alongside accuracy.

Here is how this workflow runs in practice for a small business managing customer accounts.

Before the workflow:

Every Monday, the finance lead logs into the ERP, runs the aging report, exports to Excel, filters for overdue accounts, and prepares a summary for the leadership meeting. That process takes 45 to 60 minutes. It often gets skipped when the week is busy.

After the workflow:

At 8 AM Monday, an automated query runs through the MCP-connected agent: "Retrieve all accounts with outstanding invoices over 30 days. Include invoice number, due date, amount, and last payment date. Flag any accounts over 60 days."

The agent queries the ERP, retrieves the data, and delivers a formatted exception report to the finance lead's inbox or internal dashboard. The finance lead reviews, confirms accuracy, and the summary is ready for the meeting in under five minutes.

The finance lead's time shifts from data extraction to data interpretation and decision-making.

Where human judgment still applies:

Before any follow-up action — a call to a specific customer, a payment plan discussion, a credit decision — a person reviews the account history and makes the call. The agent surfaces the data. The operator makes the decision.

The honest answer: it depends significantly on your ERP platform and your technical starting point.

For organizations already on Microsoft Dynamics 365 Business Central or Dynamics 365 Finance, the MCP server is available as a native feature in the 2026 Wave 1 release. The connection layer is already built. The primary effort is configuring the agent, defining the queries, and building the output templates.

For organizations on Xero, dedicated MCP servers from third-party providers are available and operational. Similar pattern: the connection layer exists, configuration is the primary work.

For organizations on SAP Business One, third-party MCP connectors have emerged. B1 Bridge, for example, has documented that the traditional cost of building custom API connectivity — often cited at $20,000 or more in consulting fees — is displaced by an MCP-based connection that takes under 15 minutes to configure.

For organizations on less common ERP platforms, the picture is more variable. Some will have native MCP servers. Others will require a middleware layer or custom MCP server development, which brings back a portion of the integration cost.

The practical takeaway: if your ERP is on a major platform, the infrastructure already exists. The implementation cost is configuration, workflow design, and testing — not custom development. If your ERP is on a niche platform, verify MCP server availability before assuming the same economics apply.

The shift MCP enables is not primarily technical. It is operational.

ERP systems have always contained the data operators need to run businesses well. The bottleneck was never the data — it was the friction between the question and the answer.

A sales manager who needs to know a customer's outstanding balance before a call should not have to log into a system, navigate to the right screen, and wait for a report to load. That friction slows decisions, creates inconsistency between what people know and what the system says, and pushes operators toward gut-feel judgment even when accurate data is available.

An MCP-connected agent removes that friction. The data is still in the ERP. The MCP server makes it accessible to any AI tool that speaks the protocol. The language model handles the interpretation. The operator gets an answer in seconds instead of minutes.

That is the workflow value. Not replacing the ERP. Not replacing the operator. Removing the friction between the two.

Before building anything, do this:

Write down the five questions your team asks most frequently that require ERP data to answer accurately. Not strategic questions — operational ones. The kind that come up in weekly meetings, in customer calls, in the middle of a workday.

Then ask: how long does it currently take to answer each one? Who has to do it? How often does it get skipped or approximated?

That list is your workflow backlog. The questions at the top — highest frequency, highest friction, most often approximated — are where an MCP-connected agent delivers the clearest return.

Start with one. Design the input, the AI task, the human review point, the output, and the success metric. Build that workflow before expanding to the rest.

The architecture gets complex over time. The first workflow should be simple enough to run in a week.

If you found this useful, forward it to one founder or operator who is still running ERP queries manually.