AI and machine learning in cloud Business Central

Business Central has had machine learning built in long before Copilot became a sales word. Inventory forecasts and prediction of late payments have been there for years. What is new is generative AI on top: reconciliation, matching of invoice lines and agents that read the inbox.

Here is what works today, what it requires of your data, and where we still want a person to look over it.

The AI and ML advantage in cloud Business Central

It runs in Azure — not on your server

The models behind forecasts and Copilot run in Microsoft's Azure services. That is why the features only exist in Business Central online. If you run on-premises, you do not get them.

Built into the standard product

Sales and Inventory Forecast, Late Payment Prediction and the cash flow forecast are Microsoft extensions that can be switched on without development. The Copilot features are switched on from the page Copilot & agent capabilities. Some are still in preview and work best in English, so test them on your own data before you promise anything internally.

Your data must be in order first

No model can rescue bad master data. An inventory forecast on items that have been created three times with different item numbers is worthless. The same goes for payment prediction on customers where half of the payments have been applied manually to the wrong entries. Cleaning up is always the first step.

Organisations that implement AI/ML with their cloud Business Central see on average an 18% improvement in forecast accuracy and a 23% reduction in manual data processing tasks in the first year.

Practical uses of AI

Predictive inventory management

• Sales and Inventory Forecast uses sales history to predict sales and expected stock-outs per item
• The longer and cleaner the sales history, the better — new items and seasonal items with a short history give uncertain forecasts
• Can form the basis for purchase suggestions, but does not replace reorder points and safety stock in planning
• The Power BI app for inventory has a forecast report if you would rather see it visually

Intelligent cash flow forecasting

• Late Payment Prediction scores open invoices and flags the ones likely to be paid late
• The cash flow forecast can use Azure AI to include expected receipts and payments
• Both get better the more consistently you apply payments to the right entries
• Use it to prioritise reminders — not to block customers automatically

Automated document processing

• OIOUBL and Peppol invoices via NemHandel (the Danish public e-invoicing network) need no OCR — the data comes in structured
• PDF invoices from email are read by Payables Agent or Continia Document Capture
• Copilot can match e-invoice lines against open purchase orders and suggest G/L accounts for lines without a match
• An approval flow should still come before posting — AI suggests, a person approves

Our experience: Continia Document Capture is still the most mature choice for Danish companies with many vendors and approval rules. Microsoft's own agents are moving fast, so it is worth testing them on your own invoices before you choose.

Anomaly detection in financial transactions

• Copilot in bank reconciliation suggests matches between bank statement lines and entries, and suggests accounts for the lines that do not match
• Field monitoring sends an email when someone changes, for example, a vendor's bank account — it catches more fraud than any model
• A change log on critical fields gives the auditor a trail to follow
• Duplicate invoices are best caught by the document number check in Document Capture or the standard check of the external document number

Ways to implement

Start with what is already in the standard product: switch on Sales and Inventory Forecast, Late Payment Prediction and Copilot bank reconciliation in a sandbox, and run them on your own data for a month. It costs almost nothing and quickly shows whether your data can support it.

The next step is AppSource apps for, for example, document processing. Custom development only makes sense when you have a specific problem the standard product does not solve — for example a price calculation or a forecast that has to account for something specific to your industry. Then we build it in AL, possibly with an Azure Function behind it.