How to Actually Integrate AI Into Business Central and LS Central

AI in ERP is discussed in broad terms by almost everyone and implemented in practical terms by very few. Here is a grounded look at what AI integration actually means for BC and LS Central — what is possible, what is useful, and how it is done.

The Two Approaches to AI Integration

There are two fundamentally different ways to add AI to BC or LS Central. The first is to connect an external AI service via API — sending data out, getting a result back, and surfacing that result in the ERP interface. The second is to embed AI logic inside the system's native processes using BC's AL extension framework and Copilot SDK. The first is faster to build. The second is more reliable in production.

AI for Business Central — Practical Use Cases

Copilot for BC (Microsoft's built-in AI)

BC now ships with a Copilot layer that can generate sales line suggestions, assist with bank reconciliation matching, summarise overdue accounts, and draft marketing text for item descriptions. These are available to any BC cloud customer today. They use your own BC data as context, which makes them more accurate than generic AI tools for these specific tasks.

Custom Copilot capabilities via AL

BC's Copilot Extension Framework lets developers register custom business functions as Copilot actions. In practice, this means you can build a Copilot capability that understands your specific chart of accounts, your approval hierarchies, and your pricing rules — and makes those accessible through natural language. A finance user can ask "why did our gross margin drop in Q3?" and get an answer grounded in your actual BC data, not a generic financial explanation.

Posting anomaly detection

Using BC's OnBeforePostJournalLine event, you can intercept every journal entry before it commits and run it through a classification model. The model checks the entry against your company's historical posting patterns and flags anything statistically unlikely — a freight expense posted to a revenue account, a VAT code that does not match the vendor's country, an intercompany entry hitting the wrong dimension. The flag appears as a confirmation dialog in BC's UI before the user clicks Post.

Intelligent financial close

After a period closes, an LLM can analyse the variance between actual and budget, identify the entries driving significant movements, and draft the management commentary in your standard format. The finance controller reviews and edits — but starts from a 90% complete draft instead of a blank page. Month-end commentary that used to take half a day takes 20 minutes.

Natural language reporting via Aurora BI

Our own product, built in partnership with EBT, sits on top of BC's data model and lets business users ask plain-English questions — "show me our top 10 vendors by purchase value this year, excluding intercompany" — and get structured, accurate answers. The system knows BC's schema specifically, which is why it produces correct results where generic tools struggle with ERP-specific joins.

AI for LS Central — Practical Use Cases

Demand forecasting for replenishment

LS Central's standard replenishment uses min/max levels. A machine learning model trained on your transaction history can generate daily demand forecasts per item per store over a 14–21 day horizon, accounting for seasonality, promotional effects, and local events. The model's output feeds into LS Central's replenishment suggestion workflow as a more accurate alternative to static min/max levels. Businesses that switch from min/max to ML-driven replenishment typically see stockout rates fall by 30–50% and overstock carrying costs reduce by 20–35%.

Real-time basket analysis at the POS

Every LS Central transaction is a data point. A real-time inference pipeline — sitting between the POS transaction stream and the promotion engine — can predict, mid-transaction, which items the customer is most likely to add next and which promotional offer is most likely to influence their decision. The offer appears on the cashier screen before the transaction closes. This is not theory; it requires a stream processing layer and an ML endpoint, both of which can be built on Azure and connected to LS Central's Data Director output.

Shrinkage and loss detection

A pattern analysis model running on LS Central's POS entry data can detect suspicious cashier behaviour — unusually high void rates, systematic scanning errors, no-sale-open-drawer patterns — that is statistically anomalous compared to the cashier's own baseline and to peers in the same store. The model generates an exception report daily for the loss prevention team, shifting shrinkage detection from reactive (discovered at audit) to proactive (flagged in real time).

Chatbot for store operations

A WhatsApp or web-based chatbot connected to BC and LS Central's APIs lets store managers and warehouse staff query stock levels, check purchase order status, raise transfer requests, and approve replenishment suggestions — from their phone, without opening the ERP. The chatbot is wired to BC's business logic so that actions it takes (posting a transfer order, approving a requisition) go through the correct approval chain, not around it.

What to Think About Before Starting an AI Project

AI integration in BC and LS Central delivers the most value when three conditions are in place: the underlying ERP data is clean and consistently posted, the business process being automated is well-defined and not changing rapidly, and there is a clear way to measure whether the AI output is better than the current approach. If any of these are missing, it is better to fix those first than to build an AI layer on top of unreliable foundations.