2026-09-05

Natural-language BI vs Power BI: when you do not need a semantic model first

Microsoft Power BI is excellent at published reports on a curated semantic model. That is also the cost: someone must own Power Query, relationships, measures, and usually DAX before a business user can click a slicer.

Natural-language BI (text-to-SQL) starts at the other end. You connect a live MySQL, PostgreSQL, Oracle, or SQL Server database and ask: sales this quarter by store. The engine writes a SELECT. There is no cube to refresh first.

What Power BI is actually good at

• Certified datasets shared across a Microsoft 365 tenant.
• Pixel-perfect paginated reports and Excel muscle memory.
• Row-level security already modelled in the dataset.
• Teams that already pay for Fabric / Premium and have a BI developer.

If that stack is in place, do not rip it out. Power BI Copilot still answers from the model you built, not from arbitrary production tables.

Where text-to-SQL is faster

Most mid-size companies have the facts in OLTP tables and a backlog of “can you pull this?”. Hiring a Power BI developer to model every new question is slower than letting an analyst ask in natural language against a read-only connection.

You still need a glossary (what “revenue” means in your schema) and a read-only database user. You do not need to migrate to a Microsoft warehouse first.

BI Jedi is that layer: chat, visual Builder, templates, or SQL Pro on the same connection. Data is not used to train models. If sources are scattered, the built-in datawarehouse can land CSV/FTP on a schedule.

A practical split, not a religion

• Board pack every Monday from a certified model → Power BI.
• Ad-hoc “why did pending orders spike last night?” on MySQL → natural-language BI.
• You already live in Excel + Microsoft 365 → Power BI.
• You live in Postgres and Slack, not Fabric → natural-language BI.

Many teams keep Power BI for the official pack and use NLQ for exploration. That is cheaper than forcing every question through a semantic-model ticket.

Try a bounded question on a real database: start a free BI Jedi workspace.