Detailed Analysis
A small family-owned nursery business operator is seeking to build an AI-powered analytical layer on top of Microsoft Business Central (BC) ERP, representing a practical, ground-level example of how non-technical business owners are now attempting to deploy AI agents against enterprise operational data. The business uses BC as its core system of record — covering sales, inventory, accounts receivable and payable, shipping, and project management — and supplements it with Power BI for reporting. The owner's stated goal is to enable natural language querying of live business data, such as pulling aged AR reports for specific customers, forecasting inventory movements, or generating shareable sales summaries, all without requiring manual report construction each time.
The operator has correctly identified the two primary architectural paths available for this kind of integration. Power Automate represents the lower-code, Microsoft-native route, capable of triggering data pulls and sending structured outputs, but it operates largely on scheduled or event-driven logic rather than on-demand conversational intelligence. The Model Context Protocol (MCP) approach, by contrast, involves standing up a server that exposes ERP data as callable tools, which an AI model like Claude can then query dynamically in response to natural language prompts. The owner's intuition that MCP is more powerful but more complex is accurate — MCP-based architectures allow the AI to reason about what data it needs, call the appropriate endpoints, synthesize results, and respond in a contextually intelligent way, rather than simply retrieving a pre-defined dataset.
The technical pathway for a Claude-based implementation against Microsoft BC would most likely involve either the BC OData or API v2.0 endpoints, which expose standard business objects like customers, vendors, items, sales orders, and ledger entries in a queryable format. An MCP server written in Node.js or Python could wrap these endpoints as discrete tools — for example, a `get_ar_aged_report` tool or a `get_inventory_movements` tool — that Claude could invoke when a user poses a natural language question. Authentication would be handled via Azure Active Directory service principal credentials, and the resulting MCP server could be connected to Claude either through Anthropic's API or through a Claude.ai Pro integration. The complexity the owner senses is real: configuring OAuth flows, mapping BC API schemas, and handling pagination and filtering for large datasets requires meaningful technical investment even with AI-assisted coding.
This case reflects a broader trend of mid-market and small business operators attempting to compress what has traditionally been an enterprise BI and data engineering function into a single AI-powered interface. Historically, the kind of on-demand analytical capability described — live AR aging, rolling inventory forecasting, ad hoc sales summaries — would require either a dedicated analyst, a heavily customized reporting infrastructure, or expensive middleware like Jet Reports or Continia. AI agents capable of querying APIs and synthesizing business logic are beginning to make these capabilities accessible to operators who understand their data but lack formal engineering backgrounds. The "vibe-coding" methodology the owner references — iteratively building with Claude as a coding collaborator — is itself emblematic of how AI is lowering the activation energy for this class of technical project, even as the underlying infrastructure complexity remains substantial.
The nursery business use case also highlights a nuanced challenge that will become increasingly common as more SMBs pursue similar AI integrations: the gap between what feels achievable through conversational AI interfaces and what is actually reliable enough for operational decision-making. ERP data carries real business consequences — inventory decisions, customer credit terms, cash flow visibility — and an AI agent that occasionally misinterprets a date filter or miscalculates an aging bucket creates meaningful business risk. The maturation of this space will likely depend on the development of better-validated, business-domain-specific MCP tool libraries for common ERP platforms, alongside clearer frameworks for testing and auditing AI agent outputs against known-good financial data before deploying them in production environments.
Read original article →