Connecting MS Dynamics 365 BC to Claude: how the direct route works, and how to get consolidated figures out of it (2026)

Connecting MS Dynamics 365 BC to Claude: how the direct route works, and how to get consolidated figures out of it

From exporting and pasting to asking questions: what each route delivers, where the limits sit, and how groups with multiple administrations set it up properly.

Can you connect MS Dynamics 365 BC to Claude?

Yes, and the answer comes in three routes. The first is the export route: download a trial balance or report from Business Central and paste it into the chat window. The second is a direct connection: an MCP server that talks to the official Business Central APIs, so Claude can query the administration itself. The third is the data-layer route: connect Business Central to Finstack, which consolidates and structures the figures, and let Claude work on that reporting data through the Finstack MCP. That last route comes with a unique position: Finstack is the first consolidation and reporting tool in the SME segment with its own MCP connection.

‍

Which route fits depends mostly on one question: how many administrations do you have, and do you want Claude to answer at group level? For a one-off analysis of a single administration, the export route is enough. For ongoing questions to one administration, a direct connection is workable, provided someone is willing to do the technical maintenance. For anyone with multiple entities, or with the wish to compare against budget and forecast, the data-layer route is the only one that holds up: the other two hand Claude figures it has to add up itself, and that is exactly the arithmetic where language models make mistakes.

‍

The technology behind routes two and three is the same open standard: MCP (Model Context Protocol), the connection language that lets AI assistants like Claude and ChatGPT query external data sources safely. The full explanation sits in what is MCP for finance; for this article the core suffices: you connect an MCP once, and after that the assistant retrieves the data it needs for a question by itself. The difference between the routes is therefore not the protocol, but the data source behind it: the raw administration, or the consolidated reporting layer. The rest of this article walks through both routes concretely.

How does connecting directly to Claude work?

Fair is fair: the direct route works, and for a technically minded user with one administration it is quite doable. The simplest variant is the export route: download a trial balance or general ledger export from Business Central, upload the file into Claude and ask your questions. No setup, immediate results, but the data goes stale from the moment of exporting and every new question starts with new manual work.

‍

The more serious variant is your own MCP server on the official Business Central APIs. The steps in outline: register an app in Microsoft Entra and complete the OAuth consent, set up an MCP server that translates the API endpoints into tools Claude understands, and add that server through a connection URL. After that, Claude can query the administration live: general ledger entries and balances per period, outstanding receivables and payables, invoices, relations. Do not confuse this with Microsoft's own Copilot features inside Business Central: that is an assistant within the package, not a connection that puts your Claude on the figures.

‍

In practice that works fine for individual questions: what was last month’s revenue, which five items have been outstanding longest, how is a cost category developing over the year. Within one administration, on actuals, with a technical colleague who builds and maintains the connection, the direct route is a real option, and for those who like building even a fun project. Just budget honestly: the hours are not in the first version but in every month after it. And the question is what happens once the questions grow beyond one administration; more on that next.

Why is connecting it yourself more complex than it looks?

Because the demo is built quickly, but the maintenance turns out to be the real work. Four limits that determine in practice whether the route holds.

‍

You become the owner of an integration. Every change to the API or to the MCP protocol means a change on your side: endpoints that move, authorizations that expire, keys that need rotating, a server that has to run and be monitored. The honest mirror: whoever has not built their own API connection to their Excel model until today had good reasons for that, and those reasons do not disappear because there is now AI on the other end. Moreover, the knowledge often sits with one colleague; if they leave, the connection stands still.

‍

Consolidation is not something the AI can do on the side. Adding up across administrations, converting currencies and eliminating intercompany flows is deterministic reconciliation work down to the entry. Especially sub-consolidations through intermediate holdings and eliminations at relation level, such as intercompany receivables that have to net off against intercompany payables per relation, are close to impossible to get reliable with a language model: the result looks right, but double counts and reconciliation differences stay invisible. A group figure from a direct connection is thereby suspect by definition.

‍

No budget or forecast. The API delivers what sits in the package, and that is actuals. Budget and forecast live elsewhere, so the question finance actually wants the AI for, how are we running against expectation, is one this route cannot answer.

‍

No reporting structure and no verification trail. The data arrives in chart-of-accounts layout: Claude has to guess which accounts together form the margin and does not know your definitions. And if an answer looks off, there is no controlled environment to check it in; verifying means going back into the package and reconciling by hand. This pattern is broader than Business Central; see the overview per accounting package.

How do you connect Business Central to Claude through Finstack?

The data-layer route puts a consolidation and reporting layer between Business Central and Claude, and solves all four limits in one move. Three steps, no technical project.

‍

Step 1: connect Business Central to Finstack. Through the official connection, in about 5 minutes per administration. The first synchronization retrieves the full history, not just the current financial year, so trends and comparisons across earlier years work immediately. In groups, Business Central is often the package of the larger operating company or the foreign subsidiary: it simply joins entities on other packages here, in one group picture. Finstack consolidates automatically, eliminates intercompany flows and puts the figures in your own reporting structure.

‍

Step 2: switch on the Connector pack. The add-on that contains the MCP connection and the 2-way Excel and Google Sheets sync. Grant the MCP rights to the colleagues who may see the full consolidated figures; access switches on and off per user.

‍

Step 3: connect Claude. In Claude, the Finstack connection is available in the connector directory: enable it and sign in with your Finstack account. From that moment Claude can retrieve the reporting data by itself: consolidated and per administration, filtered by entity, general ledger account or cost center, with or without eliminations, in every configured reporting structure, and with actuals, forecast and recorded normalization entries side by side. All read-only.

‍

The difference with the direct route is therefore not gradual but fundamental: Claude no longer has to calculate, only retrieve and interpret. "Margin" is the margin line from your reporting structure, the group figure is already eliminated, and the comparison against budget stands ready. The full setup, including rights and working agreements, sits in the main article on connecting AI to your accounting software.

Why is Finstack the only consolidation and reporting tool with a Claude connection?

Accounting packages deliver APIs for software vendors, not a ready-made connection for AI assistants. That connection does exist one layer up, and there the field is thin: Finstack is the first and currently the only consolidation and reporting tool in the SME segment with its own MCP connection. Anyone who wants to query their consolidated figures in Claude can therefore not only do it most easily here, but only here; at the other consolidation and reporting tools in this segment an AI connection is missing, and it stays at exports or custom work.

‍

That head start is practically noticeable in how the connection works. In Claude, Finstack simply sits in the connector directory: enable it, sign in with your Finstack account, done. No server to host, no keys to manage, no maintenance when an API changes; all of that sits with Finstack. The connection is also part of the Connector pack, together with the 2-way Excel and Google Sheets sync, so the same add-on serves the spreadsheet model and the assistant.

‍

The data layer itself widens the gap further. Finstack combines multiple ERP systems into one group picture, so entities on Business Central run next to entities on other packages, with eliminations already processed. And Finstack also stores what the accounting package itself does not know: forecasts and budgets, and the normalization and adjustment entries you record for reporting. The MCP exports those along as well, so Claude sees not the bare bookkeeping but the full reporting picture you steer on.

‍

And it is deliberately one general connection, not an integration per AI tool. ChatGPT connects through a connection URL on the same MCP and has been tested just like Claude; Copilot and Gemini support the standard in their business editions and connect the same way. Whoever switches assistants later simply takes the connection along: that is exactly the point of an open standard, as explained in what is MCP for finance.

‍

How the consolidation tools compare on exactly this criterion, next to pricing, implementation time and functional depth, sits in the comparison of the best consolidation software: there the AI connection (MCP) is included as a fixed variable in the comparison table, with the answer per tool.

What can you ask Claude about your Business Central figures?

Everything you currently look up in exports and pivot tables, but as a question. A sample of what simply works through the data-layer route, phrased literally:

• "Where do my figures deviate from budget this quarter?"
• "How is the margin developing per cost center over the last twelve months?"
• "Which entity is behind on the forecast, and on which lines?"
• "What are the five largest cost increases versus last year?"
• "Draft an analysis of this month's largest movements, per entity."

Every answer comes on the consolidated figures, in the layout of your own report, with the calculation and the provenance included: which structure, entity and period was used, and whether eliminations were on. Important outcomes you verify in Finstack itself, where the same figures sit in the dashboards and every line clicks through to the source entry in Business Central.

‍

It does not stop at individual questions. On this data basis Claude builds complete drafts: a monthly analysis with the most striking deviations per cost center as the starting point for the report, a forecast comparison that shows where realization walks away from expectation, or a dashboard with KPI tiles and charts for your own analysis. How to set up each of those applications properly sits in the deep dives on creating management reports with AI, forecasting with AI and building financial dashboards with AI.

‍

The rule of thumb for the start: begin with questions you already know the answer to, for example a margin development from the latest monthly report, and expand as the trust grows. That way the connection proves itself on familiar ground before it gets new questions.

What about security and access rights?

With the same sobriety as any other data access, and with a few clear agreements.

‍

The connection is read-only. Claude can retrieve figures, but change nothing: not in the administration in Business Central, not in the mappings or the reporting setup in Finstack. The entries stay where they belong.

‍

The access has its own rights. Owner users in Finstack have MCP access by default; for other users, that access is determined separately. Whoever has access can query the full reporting data; in this version the access cannot be sliced further per entity. The practical rule is therefore simple: grant the AI access only to those who may see the consolidated figures anyway, typically the finance team itself. For other readers, the Finstack dashboards remain the sharing channel, with rights per user.

‍

What happens at the AI vendor is up to you. Retrieved figures are processed on the AI vendor's servers to generate the answer; that is inherent to any AI assistant, including the export route. Whether they are stored or used for training is yours to check in the terms and settings of your account. Vendors do typically state that they do not train on your data, provided you use a paid or business account; verify that, and explicitly switch off training on company data where possible. The MCP helps here too, because the assistant retrieves only the reporting figures the question needs, not complete exports.

‍

Compare that with the direct route, where an API connection opens up the entire administration down to transaction level and management sits with whoever configured it. The data-layer route gives you a smaller data surface, explicit rights and one switch per user. The full framework, including the question of what you do and do not share, sits in the main article on connecting AI to your accounting software.

What does connecting MS Dynamics 365 BC to Claude cost, and how fast is it?

What differs per route is mostly what you pay in time and maintenance, not just in money.

‍

The export route is free, but costs manual work per question and delivers stale, unconsolidated data.

‍

‍The direct route requires your own MCP server on the Business Central APIs: a technical setup and maintenance effort, plus a Claude subscription. Manageable for one administration and a technical user; for a group, the consolidation and structure layer still has to be added on top.

‍

The data-layer route is a subscription without a project. Finstack starts from EUR 39 per month for the first entity, with no implementation fees; the AI access sits in the Connector pack, the add-on that also contains the 2-way Excel and Sheets sync. On top of that you need a business Claude license. On speed: the first Business Central administration connects in about 5 minutes, the full setup with reporting structure and dashboards typically stands within a day, and connecting Claude is a matter of enabling and signing in. If the Finstack environment is already running, AI access is literally a matter of minutes: switch on the pack, grant the rights, connect.

‍

The payback sits not only in the exports that disappear, but above all in the questions that currently go unasked: every analysis that until today cost too much digging to even do, per cost center, per entity, against the forecast, becomes a one-minute question. And unlike a custom integration or BI project, everything here sits in one fixed monthly amount, with no maintenance on your side. Try it yourself: start the 14-day free trial, connect your Business Central administrations and ask Claude the first question the same day. Prefer to watch first? Book a demo and see the connection live on a sample environment.

The 3 most common mistakes when connecting MS Dynamics 365 BC to Claude

Three patterns we keep seeing at teams connecting MS Dynamics 365 BC to Claude. Each costs trust in the answers; each is preventable with the right setup.

Sources and provenance

• Microsoft Learn: Business Central API documentation: learn.microsoft.com/dynamics365/business-central
• Anthropic: connectors and MCP in Claude: support.claude.com
• Model Context Protocol: open standard for AI data connections: modelcontextprotocol.io
• Finstack: Connector pack (2-way spreadsheet sync + MCP): finstack.io/solutions/connector-pack
• Finstack: integrations and pricing: finstack.io/solutions/integrations, finstack.io/pricing-plan

Last reviewed: 9 September 2026 · Next review: December 2026