“Why doesn’t revenue in the sales report match the books?” “Which customer has been paying worse lately?” “Is that invoice posted twice?”
I hear these questions in almost every project. The data is in the ERP. Yet the answer usually comes out of Excel: export, filter, compare, ask around. That takes hours, often days. And in the end only one person knows how the number was reached.
In this white paper I show how Microsoft Dynamics 365 Business Central answers these questions directly. With AI, but without a black box: the ERP does the maths, the language model does the explaining. All examples come from a Business Central test environment with Microsoft’s CRONUS demo company and an invented business story. There is no customer data.
Interactive demo: Open the Management Cockpit for Business Central – Superset dashboard with AI commentary, demo data.
What Microsoft delivers today, and where it stops
Business Central now ships with a range of AI features:
- Copilot features such as list analysis, marketing text, bank reconciliation and summaries. They are included in the licence.
- Agents for sales orders (email in, quote out) and purchase invoices (PDF in, draft invoice out). They are billed by consumption through Copilot credits.
- An MCP server. It lets external AI assistants such as Claude, ChatGPT or Copilot Studio access Business Central directly, with permissions.
That is a good start. In practice, though, three things stand out:
- Microsoft’s agents cannot be extended. They work with what Microsoft has designed.
- They create drafts. Posting, approving and deciding stay with people. Rightly so.
- None of the standard agents answers the questions finance asks: reconciliation, payment behaviour, anomalies.
That is exactly where a custom extension comes in.
The principle: AL calculates, AI explains
A language model is good at explaining, summarising and prioritising. Doing sums over a hundred thousand postings, it is unreliable. So I split the work:
- Business Central calculates. A small AL extension exposes sound accounting figures as APIs: payment behaviour per customer, subledger against general ledger, unusual postings. The rules are transparent and testable.
- The Business Central MCP server automatically turns each of these APIs into a tool an AI assistant is allowed to call. Which tools, and with which rights, is set by the administrator in Business Central.
- The language model explains. It decides which tool it needs, reads the facts and writes the answer in the language of whoever asked: English, German or Italian.
The key point for IT: the configuration is read-only. The assistant cannot post, change or delete anything. It sees only what the signed-in user is allowed to see in Business Central anyway.
The building blocks
- An AL extension with three API pages and two API queries. All figures are calculated on request; the extension writes no data.
- An MCP configuration in Business Central with exactly these tools, no write rights and no dynamic object discovery.
- An app registration in Microsoft Entra ID for sign-in (Business Central does not support dynamic client registration).
- An AI assistant that speaks MCP. In my demo it is Claude; Copilot Studio or other clients work just as well.
Use case 1: “Which customer is paying worse?”
The problem
Every ERP shows open items. What is missing is the trend: is a customer who used to pay on time paying later and later over the past few months? That information sits in closed invoices and payments. But nobody calculates it regularly.
The solution
The extension calculates per customer:
- open and overdue amounts,
- average days to pay over the last 180 days compared with the 180 days before,
- the number of late payments,
- and from that a trend: faster, stable, slower.
The result in the demo
Asked “Which customers have been paying worse lately, and how much cash is tied up?”, the assistant answers roughly like this:
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 Business Central. The prioritisation and wording come from the language model.
Use case 2: “Why don’t the numbers add up?”
The problem
Sales reports revenue from item ledger entries, finance from the general ledger. Mostly the two agree. When they don’t, the search begins: manual postings to revenue accounts, credit memos, payment discounts, prepayments, allowances. I have seen reconciliations that cost a day every month.
The solution
For the last 24 months the extension compares:
- revenue according to item ledger entries (subledger),
- revenue on the general ledger revenue accounts, split by source: from sales documents, from manual general journal entries, from other sources.
A second API returns the individual postings to revenue accounts that do not come from sales documents. In other words, exactly the candidates for a difference. The extension reads which accounts are revenue accounts from the General Posting Setup. Nothing extra needs maintaining.
The result in the demo
Question: “Why does August revenue differ from the sales report?”
In August 2026 revenue in the general ledger is €24,700, in the subledger €15,200. The difference of €9,500 comes entirely from a manual general journal entry on 31 August to revenue account 6110 (“Manual correction project revenue Q3”, document KRBJNL-0001). All other months match. Check whether this correction has supporting documentation, and why the account allows direct posting.
The last sentence is the real gain. The difference is not only found; its cause is named too: a revenue account that accepts direct postings.
Use case 3: “Does anything look odd?”
The problem
Duplicate purchase invoices, lump-sum postings without documents, weekend postings, unusually high amounts: auditors look for these by sampling. Day to day, rarely does anyone look systematically.
The solution
The extension checks the last twelve months with six simple, transparent rules:
- same vendor, same amount within ten days,
- invoices far above that vendor’s usual amount,
- manual postings at the weekend,
- round manual amounts on expense and income accounts,
- manual postings to revenue accounts,
- statistical outliers per G/L account: each posting is compared with the mean and spread of the other postings, so a single extreme value cannot hide itself.
The result in the demo
In the demo data the rules find nine hits, all from the planted story, with no false alarms in the regular CRONUS data. The assistant sums up:
Three points are worth clearing up this week. First: there are two invoices from First Up Consultants for €14,976 each, posted on 15 and 18 September, with almost identical invoice numbers (FUC-2026-0915 and FUC-2026-0915-1). This is very likely a duplicate payment. Second: an invoice from Graphic Design Institute for €48,000 is 36 times that vendor’s average. Third: on Sunday 13 September, €15,000 was posted manually as “Consulting lump sum (no invoice)”. Round amount, no document, weekend.
The assistant does not invent hits. It reads the results of the rules, links them (the duplicate vendor shows up in two rules) and ranks them by urgency.
What you need to get started
- Business Central online. The MCP server and Copilot exist only in the cloud version.
- A sandbox. New extensions are built and tested in a sandbox, never directly in production.
- An Entra app registration with the delegated permission
Financials.ReadWrite.Alland admin consent. The actual rights are still governed by the permission sets in Business Central and the read-only MCP configuration. - A decision on the language model. Microsoft’s own features run on Azure OpenAI within the EU Data Boundary. External assistants such as Claude process the retrieved results in their own environment. That is a question for data protection and IT, not a technical one.
Limits, honestly
- The language model can be wrong. That is why it does not do the maths itself. It receives finished figures and is asked to explain them. Every answer can be traced back to an API and a rule.
- Rules are only as good as the master data. If the General Posting Setup has gaps, so does the list of revenue accounts.
- No decisions. The configuration is read-only. Whether an invoice is reversed or a customer is chased is decided by a person.
- Cost. The API calls themselves cost nothing extra. Microsoft’s agents are billed through Copilot credits, external assistants according to their own pricing.
- On-premises. Business Central in your own data centre has neither the MCP server nor Copilot. There you need a different route.
What comes next
The “AL calculates, AI explains” pattern goes further. The next use cases I am working on:
- E-invoices. Incoming XRechnung, ZUGFeRD and FatturaPA matched automatically to accounts, items and dimensions. The XML provides the structure; the AI only does the matching.
- Dunning in German and Italian. Prioritised open items, reminder drafts in the right tone for each customer.
- Master data upkeep. Duplicates, missing commodity codes, incomplete translations: found, proposed, approved.
- Orders in mixed units. Cubic metres, running metres, bundles: orders by email or Excel, as is common in building materials and timber trading.
Is your team facing a similar problem?
If someone in your company re-does Business Central numbers in Excel every month, that is a good candidate. Tell me which question you would like to ask your ERP: Discuss your challenge. I usually reply within two working days.