AI
10 questions to ask your own business data – today
Most people don't know what to ask their AI when it comes to their own business. Here are ten concrete questions – about customers, margins, inventory and production – that you can put directly to your data, plus what it takes to make the answers trustworthy.
In short
- Ten ready-made questions on sales, purchasing, inventory, production and finance that you can put directly to your own business data.
- The value only appears when the assistant reads the company's actual figures instead of guessing.
- The answers only hold up if there is a semantic model, saved history and controlled access underneath.
- Three things immediately improve accuracy: state the period, ask for the breakdown and follow up.
Good answers start with a good question
The most common objection we hear isn't that AI feels difficult. It's that people don't know what to ask about. As long as the assistant is used to summarise emails and polish text, the value is clear but small. Only when the questions concern your own business – customers, margins, inventory, production – does it start to make a difference day to day.
So we have gathered ten questions our customers put directly to their own data. All of them can be asked today, in Claude or ChatGPT, once the assistant is connected to a semantic model via Vellox MCP. Model Context Protocol (MCP) is the open standard for connecting AI assistants to external systems – read more about how to prompt your business data with AI. Copy the questions as they are, or swap the period and dimension for something that suits your business better.
Sales and customers
1. Which ten customers did we lose the most revenue from compared with the same period last year – and which item groups account for the drop?
The classic decline rarely shows in the total figure. A customer halving its volumes is hidden behind two new ones that are growing. Breaking it down by item group also shows whether the whole relationship is slipping or just a single product range.
2. Which customers used to order regularly but haven't placed an order in the past 90 days?
A list of forgotten customers is usually the cheapest sales activity there is. Ask the assistant to sort by historical order value, and the sales rep knows where to start calling.
3. How has gross margin per sales rep and customer segment developed over the past twelve months?
Volume is easy to measure and easy to steer towards. Margin is harder – and that is where the difference between a good and a bad deal actually lies. The data comes from the sales module, the same source the management report uses.
Purchasing and suppliers
4. How much of our suppliers' price increases over the past year have we passed on to customers, by item group?
The question that usually provokes the strongest reaction in the management team. Purchase prices are adjusted continuously, sales prices sit in a price list updated once a year. The gap between them is pure margin loss.
5. Which suppliers have the worst delivery precision, and which items are hit hardest?
Material for your next supplier meeting. Confirmed delivery date against actual receipt date, by supplier and item – instead of a gut feeling about who tends to be late.
Inventory and tied-up capital
6. Which items tie up the most capital without having moved in the past six months?
Sorted by inventory value, the list becomes concrete immediately. Often a handful of items account for most of the dead capital, and they are rarely the ones you would guess.
7. Which items risk running out within 30 days given current consumption rates and scheduled receipts?
The forward-looking version. The same data, but the question is about the future rather than the past – and the answer can be acted on before the customer gets in touch.
Production
8. Which items show the largest deviation between calculated and actual manufacturing cost?
The costing is set once and then lives a life of its own. When material, setup time and outcome are compared against the calculation, it quickly becomes clear which items are sold at a margin that only exists on paper.
9. How have scrap and rework developed per production group over the past year?
A trend that rarely has a report of its own, but which often explains why a seemingly stable product has lost profitability.
Finance and management
10. What explains last month's deviation from budget – break it down by volume, price and mix?
The hardest question in the monthly review, and the one that usually takes longest to answer manually. With the definitions in place in the model, the breakdown takes seconds instead of an afternoon.
What makes the answers worth trusting
The questions above are easy to ask. What determines whether the answers hold up is what sits underneath them – something we covered more broadly in AI and data that create real impact.
-
A semantic model at the core. Margin, revenue and inventory value are defined once, in one place. The assistant doesn't reinvent the concepts on its own – it reads the same measures as the management report.
-
History that is actually retained. ERP systems often overwrite instead of saving. Without stored history, there is no way to answer how something has developed.
-
Read-only connection and controlled access. No write operations are possible, and you decide which parts of the model are exposed and to whom.
-
Traceability. Every answer can be traced back to measures and dimensions. That is the difference between decision support and a guess that sounds convincing.
Three tips that improve the quality of the answers
Always state the period. "Last quarter" and "year to date" give different answers. Be explicit and you avoid misinterpretation.
Ask for the breakdown right away. Add "per customer group", "per item group" or "per month" to the question. An aggregate is rarely enough to base a decision on.
Ask the follow-up question. The biggest gain isn't in the first answer but in being able to drill deeper without commissioning a new analysis. "Why did the margin fall precisely there?" is usually the question that leads somewhere.
What makes an answer useful isn't that it arrives quickly, but that it is based on the same figures as the management report.
Sources
-
Model Context Protocol: Introduction
-
Microsoft Learn: Semantic models in the Power BI service
Frequently asked questions about querying your own business data
Which ERP systems does it work with?
Vellox retrieves data directly from Business Central, Dynamics NAV, Jeeves, Monitor, Pyramid and Visma Business. The model comes pre-configured for your system from day one, so the questions above can be asked without anyone having to build a data model first.
Do I need Power BI to ask these questions?
No. The same semantic model powers both the reports in Power BI and the AI assistant – Microsoft describes the concept in its documentation on semantic models. You can use one, the other or both – the figures are the same.
Can the AI change anything in the ERP system?
No. The connection is read-only, no write operations are possible. You also decide which parts of the model are exposed and to whom.
How do I know the answer is correct?
Every answer can be traced back to the measures and dimensions it is based on. Because the definitions live in one place, the assistant doesn't come up with its own versions of margin or revenue.
What does it take to get started?
Data is retrieved from the ERP system and structured in a semantic model. For the most common systems it's a matter of days, not months – see Get started.