We recently had this discussion with an Australian importer and distributor that had moved from MYOB to Business Central. Business Central was working well enough as the ERP, but the standard reporting was too generic for how management wanted to run the business. They wanted better executive reporting, sales analysis, customer reporting and budget tracking, so Power BI was the obvious next step.
Then the conversation shifted.
With MCP developing quickly, and AI tools becoming more capable of connecting directly into systems like Business Central, the business asked whether it still made sense to invest in Power BI at all. If AI can query Business Central directly, why not just ask questions as needed and build reporting conversationally?
The answer comes down to the difference between accessing data and defining business logic.
MCP is useful because it gives AI applications a structured way to interact with systems like Business Central. Depending on what has been exposed, that can include API pages, API queries and actions. For operational use cases, this can be very effective. A user may want to know which orders are still open, what a customer bought last year, which purchase orders are overdue, or what stock is currently available.
Those questions are close to the underlying transactional data, and direct AI interaction can work well.
Where it becomes more complicated is management reporting.
Business Central stores transactional data according to how the ERP works. That is not necessarily the same thing as how management wants to measure the business.
Take revenue as a simple example. A reliable reporting model still needs to determine which transactions count as revenue, how credit notes are treated, which posting date is used, whether freight recoveries are included, how intercompany activity is handled, and whether open orders belong in the same measure or a separate one.
Gross profit has similar issues. If the business uses FIFO costing, the model needs to reflect the actual costing treatment in Business Central rather than applying a generic calculation.
Budgets create another example. If the business maintains a quarterly sales budget by item but reports performance monthly, something still needs to define how that quarterly budget is allocated across the three months.
None of that logic exists automatically just because an AI model can access the ERP.
That is what the semantic model is there to solve.
In a typical Business Central reporting environment, we would extract the required data through standard APIs, custom API pages or queries, or OData where appropriate. That data can then be landed into Azure or Microsoft Fabric and transformed into a reporting structure that is much easier to work with than the raw ERP schema.
The simplified architecture looks something like this:
Business Central → APIs / OData → Fabric or Azure data layer → Semantic Model → Power BI
Once that model exists, other tools can use it as well:
Semantic Model → Power BI / Copilot / Fabric Data Agents
MCP can then sit alongside that architecture as a more direct route into Business Central:
AI Client → MCP → Business Central
That distinction matters.
The semantic model contains the agreed logic of the business. It defines relationships between tables, calculated measures, financial periods, product and customer groupings, budget logic and reporting security.
For example, a salesperson may only need to see customers assigned to them, while a regional manager may need access across multiple reps. Finance may need full access, while executive reporting may include measures that should not be exposed to all operational users. That reporting security does not always match the permissions configured inside Business Central itself.
This is one of the reasons we would not normally use direct AI access to the ERP as the main management reporting layer.
There is also a structural issue with querying transactional systems directly. ERP schemas are designed to support transactions, not necessarily analysis. Data can be spread across headers, lines, ledger entries, dimensions, item records and posting tables. The same business concept may need to be reconstructed across several of those objects before it becomes useful for reporting.
A semantic model gives the business one controlled version of that logic.
That becomes even more important as AI is introduced.
If five users ask an AI tool broadly similar questions about sales performance, you do not want the result changing because the model has interpreted the underlying ERP structure differently each time. Ideally, the AI should be working from the same measures management already uses in formal reporting.
That is where tools like Fabric Data Agents and Copilot become more useful. Rather than querying raw Business Central structures directly, they can operate over a governed semantic model where the meaning of revenue, GP, budget, region and other measures has already been defined.
Power BI then remains useful for a different reason.
Management reporting is repetitive. The executive team wants to see the same core measures each week or month, calculated consistently and available without anyone having to formulate the right prompt.
A dashboard can also surface something the user did not know to ask about. A customer may have stopped ordering, a region may have fallen behind budget, a product category may have declined, or gross margin may have moved outside an expected range.
Direct AI interaction is much better suited to follow-up questions.
Why did sales fall in that region?
Which customers contributed most to the decline?
Which products changed?
What happened to margin?
That is where the combination starts to make sense.
The part of these projects I think businesses should focus on is no longer simply the dashboard.
A client may come to us asking for an executive dashboard and a sales dashboard, but the more valuable output is usually the underlying data and semantic model.
Once that exists, Power BI can use it for recurring reporting. Copilot or Fabric Data Agents can use it for conversational analysis. MCP can provide direct operational access to Business Central where that is more appropriate.
The interface can change over time without rebuilding the underlying logic.
That is why I do not think AI removes the case for Power BI. It changes the architecture around it.
The more useful question is not whether the business should choose Power BI or AI.
It is whether the business has a governed data model that both can rely on
