The most underrated AI feature in Business Central isn’t a Copilot button.

It’s the MCP server.

IT teams asking where to start with AI in Business Central usually expect an answer about Copilot or Microsoft’s agents. Those are useful. But they do what Microsoft designed, and they don’t answer the questions finance actually asks: why the numbers don’t match, which customer is paying worse, what looks odd. For that you need your own logic. The MCP server is how that logic reaches an AI assistant.

What the MCP server does

MCP, the Model Context Protocol, is an open standard that lets an AI assistant call tools. Business Central now ships its own MCP server. Every API page or API query you write in AL can become a tool that an assistant (Claude, ChatGPT, Copilot Studio or any other MCP client) is allowed to call.

You decide which tools, and with which rights, on one configuration page in Business Central.

How I build with it

This changes the order in which I build:

  1. Write the business logic in AL. Deterministic, testable, explainable. The same input gives the same figure every time.
  2. Expose the result as a read-only API. An API page or an API query, nothing more.
  3. Add it to an MCP configuration. Pick the tools, switch off write access.
  4. Let the AI do what it’s good at. Choosing the right tool, reading the result, ranking, explaining in the user’s language.

I call this AL calculates, the AI explains. A language model is good at summarising and prioritising. Summing a hundred thousand postings, it is unreliable. So it never does the maths. It gets finished figures and puts them into words.

What I configured for the demo

For the “Hack Business Central” series I built one AL extension with three API pages and two API queries:

  • payment behaviour per customer (open and overdue amounts, days to pay now against before, trend)
  • monthly reconciliation of the revenue subledger against the general ledger
  • revenue postings that don’t come from sales documents
  • the results of six anomaly rules

All figures are calculated on request. The extension writes no data.

The MCP configuration contains exactly these tools. No write rights. No dynamic object discovery, so the assistant cannot browse other tables on its own. And it only sees what the signed-in user is allowed to see in Business Central anyway.

The result in the demo

The demo runs on Microsoft’s CRONUS demo company with an invented story. No client data.

I asked: “Which customers have been paying worse lately, and how much cash is tied up?” I did not name a tool. The assistant chose the payment behaviour API by itself and answered:

Two customers are paying markedly slower. Alpine Ski House took 68 days on average to pay over the last six months, against 42 before; €25,124 is overdue. Adatum Corporation has slipped from 32 to 49 days, with €14,409 overdue. School of Fine Art, by contrast, now pays faster than before (49 → 26 days). I would start with Alpine Ski House: highest amount, sharpest decline.

The numbers come from AL. The ranking and the wording come from the model. Ask the same question in Italian or German and you get the same figures in another language.

What you need

  • Business Central online. The MCP server only exists in the cloud version.
  • A sandbox. New extensions are built and tested there, never directly in production.
  • One Entra app registration. Business Central does not support dynamic client registration, so the assistant signs in through an app you register, with the delegated permission Financials.ReadWrite.All and admin consent. The real rights still come from the Business Central permission sets and the read-only MCP configuration.
  • A decision about which assistant may see which data. Microsoft’s own features run on Azure OpenAI within the EU Data Boundary. External assistants process the retrieved results in their own environment. That is a data-protection question, not a technical one. Make it before the first test, not after.

What it doesn’t do

  • It doesn’t post, approve or decide. It shouldn’t. The configuration is read-only on purpose.
  • It doesn’t fix bad logic. If the AL calculation is wrong, the assistant explains a wrong figure fluently. The test effort belongs in AL.
  • It doesn’t replace permissions. The assistant sees what the user sees. If a user should not see salaries, the permission set has to say so.
  • The language model can still be wrong in the wording. That is why every answer must trace back to a tool result.
  • It doesn’t exist on-premises. Business Central in your own data centre has no MCP server. There you need a different route.

Try it

You can see the tools and the AI commentary in the interactive demo: Open the Management Cockpit for Business Central. Demo data only.

Architecture, building blocks and the full list of limits are in the white paper: Ask Business Central: three AI use cases that work today.

Is your team facing a similar problem?

If you have a question your ERP should answer and an IT team that wants it read-only and traceable, this is a good place to start. Discuss your challenge. I usually reply within two working days.