Validating Your BC MCP Setup with Real-World Queries
Introduction
Your BcMCPProxy is
built, Claude Desktop is configured, and the BC MCP Server is live. Now let’s
put it all to the test.
In this final post, we
go beyond simple data lookups. The real power of connecting an AI agent to
Business Central is not just reading records. It is asking business questions
in plain English and getting visual, actionable answers. We will walk through two
scenarios that demonstrate exactly that, followed by best practices for working
with AI agents.
Scenario 1: Project Budget vs. Actual Analysis
This is where things
get interesting. Instead of opening BC, navigating to project ledger entries,
exporting to Excel, and building a chart manually, you just ask.
Prompt
"I want to see budget vs actual for all Projects
with a nice graphic
view"
What
Happens Behind the Scenes
Claude does not just
make one API call. It thinks through the problem and executes multiple steps:
1. Claude identifies that
it needs project data with budget and actual figures. It calls the BC MCP
Server to read the Project API page.
2. It retrieves budget
amounts and actual costs for all projects. If the data spans multiple API pages
(project tasks, project planning lines, project ledger entries), Claude calls
each one as needed.
3. Claude then generates
an interactive chart, typically a grouped bar chart, comparing budget vs.
actual for each project side by side.
4. The chart is rendered
as a visual artifact directly inside Claude Desktop. No Excel export. No
copy-paste. Just ask and see.
|
✅ Setup Note For this
scenario, I created a custom API page in Business Central that exposes the
Job Planning Lines table with all the relevant budget and actual fields. The
standard BC APIs do not include project planning data out of the box, so a
custom API page was needed to give the agent access to this information. This
is a common pattern: if the data you want is not available through standard
APIs, build a custom API page, add it to your MCP configuration, and the
agent can use it immediately. |
Why This
Matters
Think about how long
this takes the traditional way. Open BC, navigate to the right page, apply
filters, export to Excel, format the data, build a pivot chart, adjust colors
and labels. That is 15 to 30 minutes of work, depending on how many projects
you have.
With MCP, you type one
sentence and get a visual answer in seconds. The agent handles all the API
calls, data shaping, and visualization for you.
Expected result: A bar chart showing
budget vs. actual cost for every project in your BC environment along with
total budget, actual cost, remaining budget and budget utilization.
|
✅ Pro Tip You can
refine the prompt to focus on specific projects or add more detail. For
example: "Show me budget vs actual for projects J00010 and J00030 with a
breakdown by project task." The more specific your prompt, the more
targeted the output. |
|
✅ Prompt Refinement If the
initial chart does not look quite right, just tell Claude what to change.
"Make it a horizontal bar chart instead" or "Add percentage
variance labels" or "Only show projects that are over budget."
The agent iterates on the visualization without you touching any code. |
Scenario 2: Available Inventory Overview
Inventory visibility is
one of those things that everyone needs but nobody wants to build reports for.
Let’s skip the report.
Prompt
"Show me available inventory in a nice chart"
What
Happens Behind the Scenes
1. Claude calls the BC MCP
Server to read the Item API page and retrieves all items with their inventory
quantities.
2. It filters for items
that actually have inventory on hand (skipping zero-stock items to keep the
chart clean).
3. Claude generates a
visual chart. Depending on the number of items, it might use a bar chart for a
smaller catalog or a treemap/bubble chart for a larger one.
4. The chart is rendered
as an interactive artifact in Claude Desktop, with item numbers, descriptions,
and quantities.
Why This
Matters
This is a question that
gets asked in every warehouse meeting, every sales call, and every planning
session. Instead of logging into BC, navigating to the item list, sorting by
inventory, and maybe exporting it for a visual, you just ask.
And because it is
conversational, you can immediately follow up: "Which items are below
their reorder point?" or "Show me just the items in the FINISHED
category" or "What is the total inventory value?" The agent
keeps the context and builds on the previous answer.
Expected result: A visual chart showing
available inventory quantities across your items. Items with the highest stock
levels are immediately visible. Zero-stock items are excluded for clarity.
|
✅ Taking It Further Try chaining
requests: "Show me available inventory, then highlight items that have
not had any sales in the last 90 days." Claude can cross-reference
inventory data with sales data across multiple API calls to surface
slow-moving stock. This kind of ad-hoc analysis used to require a dedicated
report. Now it is a conversation. |
The Bigger Picture
These two scenarios are
just the beginning. The pattern is always the same: ask a business question in
plain English, and the AI agent figures out which BC API pages to call, fetches
the data, and presents it visually.
Here are a few more
ideas to try once your setup is working:
•
"Show me a trend of sales order totals by month for the
last 12 months."
•
"Which customers have overdue balances? Show me a breakdown
by aging bucket."
•
"Compare actual vs. budget for GL accounts in the 6000
range this quarter."
•
"What are our top 10 selling items by revenue this year?
Show me a chart."
Each of these would
traditionally require navigation, filtering, exporting, and charting. With MCP,
they are one-line prompts.
|
|
Best Practices
Security
•
Least privilege: Start with a read-only MCP configuration. Only
enable write operations after thorough testing.
•
No secrets in config: No passwords or secrets are stored. The
proxy uses Azure’s delegated auth flow.
•
Separate configurations: Create purpose-specific configurations
(e.g., Finance-ReadOnly, Sales-ReadWrite).
•
Audit access: Review who has the MCP - ADMIN permission set
in BC.
Working with Agents
•
Be precise: Agents work best with specific, clear requests.
“Show me customers” is good; “Show me the top 10 customers sorted by balance”
is better.
•
Request complete data: When you need all records, say so
explicitly and instruct the agent not to use the top parameter.
•
Understand tool calls: Each MCP tool call is a separate API
request. Complex queries may take a few seconds.
•
Review before approving: For sensitive operations, the agent will
ask for confirmation. This is by design.
Maintenance
•
Stay current: The BC MCP Server is in public preview and may
change. Watch the Microsoft Learn documentation for updates.
•
Expand over time: As BC evolves, new API pages may become
available. Add them to your MCP configuration to expand agent capabilities.
•
Test after updates: If something breaks after a BC update, re-check
Feature Management and MCP configuration.
Conclusion
Over these four posts,
we have gone from understanding what AI agents and MCP are, through enabling
the BC MCP Server and configuring it, to building a proxy executable and
testing it with real business questions.
The key takeaway is
this: the value is not in reading individual records. It is in asking business
questions and getting visual, actionable answers in seconds. Budget vs. actual
comparisons, inventory overviews, trend analysis. All of this is now a conversation,
not a report-building exercise.
The BcMCPProxy is a
temporary bridge. Microsoft will likely provide direct MCP client support in
future releases. But it works today and demonstrates the enormous potential of
agentic AI in the Business Central ecosystem.
The future is agentic.
Let’s build it.
Blog Series Navigation
Blog 1: Understanding AI Agents and MCP
Blog 2: The Business Central MCP Server
Blog 3: Building BcMCPProxy.exe and Connecting to Claude Desktop
▶ Blog 4: Testing Scenarios and Best Practices (You are here)
