BI and AI-powered analysis: querying your data in plain language
"How did the margin in western Switzerland develop against the previous quarter?" Typing that question into an input box and receiving a sound answer with a chart has been possible since 2024. I reviewed the available demonstrations and set out what holds up today, where the limits lie, and which preparation decides the outcome.
What natural-language querying means
Natural-language querying means you ask a question in everyday language and the system translates it into a database query. Microsoft built this capability into Power BI under the name Copilot. The May 2023 announcement describes the ambition of weaving large language models into Power BI at every layer.
For the finance function this shifts a responsibility. Until now management asked a question, controlling built an analysis, and the answer arrived days later. If the question can be asked directly, that loop shortens to minutes.
What works today
The reviewed demonstrations produce a clear picture of the dependable applications:
| Application | Classification |
|---|---|
| Answering questions about existing reports | Dependable |
| Generating a report or page from a data model | Dependable, rework needed |
| Summarising and commenting on variances | Dependable, review needed |
| Proposing formulas and metrics | Partly |
| Forecasts without an underlying model | With reservations |
The last row deserves attention. A language model produces a plausibly worded projection, and with forecasts plausibility and correctness part company. Dependable forward views require an underlying model whose assumptions are named and testable.
What the quality of the answer depends on
The most important finding across all demonstrations: the quality of the answer depends on the data model, rarely on the language model. An October 2025 contribution goes so far as to put data preparation in its title and covers it before report creation. Four points decide:
- Meaningful names. A column called CUSTNO_2 stays mute. Call it Customer number and the query finds it.
- Defined metrics. As long as gross margin is missing from the model, the system assembles it itself – and returns a figure different from your accounts.
- Clean relationships. The links between tables determine which analyses are possible at all.
- Permissions. Anyone who can query all data can also see payroll. Access rights belong before the rollout.
These four points are the same ones that make any good data model. In that sense Copilot rewards groundwork that pays off independently of it.
The Swiss context
The AXA study of October 2025 shows data analysis gaining ground as an application: 32 per cent of Swiss SMEs use AI for it, up from 22 per cent the previous year. Communication tasks remain more common, with translation at 52 per cent and correspondence at 47 per cent. Among SMEs using AI, 57 per cent report time savings, up from 46 per cent.
In practice this means your staff are probably already working with language models, namely in correspondence. The step to your own figures is smaller than it appears – and it calls for a rule on which data may enter which tool.
The author's assessment
Copilot changes access to analysis more than it changes analysis itself. The gain arises where a question lands in controlling today, sits for three days, and ends in an analysis containing a single figure. That loop can be closed, and it is worth a great deal.
My reservation concerns accountability. An answer from a language model carries the same confident tone whether it is right or wide of the mark. For figures that go outside – the bank, the auditor, the board – review remains with a human. I therefore recommend a simple separation: Copilot for exploration and preparation, the reviewed report for decisions and external use.
And the uncomfortable truth to close on: turn Copilot loose on a neglected data model and you get fast wrong answers instead of slow right ones. The order stays the same as in any BI initiative – the model first, the convenience after.
The short version
- Natural-language queries have been production-ready since 2024 and shorten the loop from days to minutes.
- Dependable: questions about existing reports, report generation, variance commentary. Forecasts need an underlying model.
- Answer quality depends on the data model: meaningful names, defined metrics, clean relationships, managed permissions.
- 32% of Swiss SMEs use AI for data analysis; 57% of AI users report time savings (AXA, October 2025).
- A practical separation: Copilot for exploration, the reviewed report for decisions and external use.
Read on
Sources
This article draws on the metadata and full descriptions of the YouTube contributions listed below, together with the chapter markers and quotation highlights the channels published themselves. YouTube no longer releases verbatim transcripts, so statements are attributed to their originator and marked as that person's position.
- Microsoft Power BI: Copilot in Power BI Demo, 23 May 2023, — min.
- Power BI Community: AI-Powered Data Analysis & Reporting with Copilot for Power BI Demo, 5 Feb 2025, — min.
- Power BI Community: Copilot in Power BI Unplugged: Part 1 – Ask Data Questions, 20 Feb 2025, — min.
- Power BI Community: How to Use Power BI Copilot: Prep Data for AI, Create Reports & Ask Questions, 1 Oct 2025, — min.
- AXA Switzerland: SME Labour Market Study 2025 – Artificial intelligence is conquering Swiss SMEs (in German), media release of 8 October 2025.
Is Copilot worth it for your figures?
The answer depends on your data model, not on the licence. We check together whether the preconditions are in place.
Book an initial consultation