Copilot

How it works

An administrator can switch the capability off at any time under Business Central's Copilot & agent capabilities — the entry point then disappears for everyone.

Writing prompts that work

• Name a figure. Every analysis needs at least one measure. show net sales per product category works; “show product categories” alone is refused, because there is nothing to calculate.
• Name the attributes to group by, outermost first: net sales by salesperson then country. Say which attribute becomes the columns if you care: years as columns.
• Filters belong in the sentence. net sales per customer, without quotes, only 2026. The full operator set is available — equals and not equals, contains and not contains, starts/ends with, greater/less than, between, blank and not blank.
• Write filter values the way your data does. Copilot never sees your data, so it takes values literally from your question. “Rechnung” finds Rechnung; it cannot know that your model calls it Invoice (see below).
• Refine instead of restarting. After the first proposal, short instructions are enough — add …, remove …, filter by …, sort by …. The prompt guide offers these as starting points.

What Copilot can and cannot do

• It builds layouts, not queries. The model answers with a small layout description — fields, filters, sorting. The DAX query is built and checked by the app, and every figure is evaluated by your semantic model, exactly like a hand-built analysis.
• Everything is checked against your model. Field names are verified against the live field list. A field that does not exist is reported in the review notes — never silently invented. If a question cannot be answered with the model's fields at all, Copilot says so instead of guessing.
• Filter values are verified in your tenant, not by the model. After the proposal arrives, the app checks filter values with ordinary queries against your semantic model: exact matches adopt the model's spelling, close matches are completed (“Adatum” → “Adatum Corporation”), and a value that does not exist is named in the review notes — the analysis never silently shows zero rows. What this deliberately cannot do is translate: “Büromöbel” will not find a member called “Office Furniture”, because that would require sending your data to the model. Pick the right value in the filter pane afterwards — its list never left your tenant.
• Same question, different proposal. Language models are not deterministic; asking again can produce a different (equally valid) layout. The history keeps every proposal, so nothing is lost by trying again.
• Review notes are the reliable half. The explanation is what the model believes it built; the review notes list what actually changed — corrected fields, corrected values, dropped filters. When they disagree, trust the notes and the preview.

Your data

Copilot runs against an Azure OpenAI resource in your own subscription, called server-side from Business Central. The key is stored in Business Central's Isolated Storage and never reaches the browser. No navida service is involved, and AI consumption is billed by Microsoft Azure on your resource.

The analysis itself works like every other analysis in the app: queries run as the signed-in user, row-level security applies, and results travel directly between your browser and your Power BI tenant — see Architecture.

Set-up in short

An administrator creates a chat-completions deployment on an Azure OpenAI resource, enters endpoint, deployment and key on Power BI Analysis Setup, switches Enable Copilot on and runs Test Copilot Connection. The step-by-step guide, licensing notes and troubleshooting are on the Install page. Copilot is included in the app licence — no separate plan.

Contact us

Need support? Use this contact link to send us an email.