Natural language analytics. Letting people ask the data questions, without a new platform.

Asking your data a question in plain English sounds like a feature for the future, but it is already in the tools many mid-market businesses are paying for today. The harder question is whether the answer you get back is one you can trust.

The demo is impressive, and that is the problem. Someone opens the Microsoft Fabric data agent or Power BI Copilot, types which region grew fastest last quarter in plain English, and an answer comes back in seconds, no report builder, no waiting on the data team. It is easy to watch that and conclude the hard part of analytics has been solved, and that the thing to buy is the chat box. The chat box is the easy part. The value, and the risk, is entirely in what sits behind it.

The demo is the easy part.

Natural language analytics means letting people ask questions of their data in ordinary words and get an answer back, rather than building or reading a report. The Fabric data agent and Power BI Copilot both do it, and they do it well. The reason the demo always lands is that the language step, understanding the question and writing the query, is the part the technology has cracked. It is not where your effort goes, and it is not where your money should go first.

Point it at a mess and you get a confident wrong answer.

Here is what the demo never shows. Point one of these tools at an ungoverned estate, scattered reports with their own quiet versions of revenue, no agreed definition of an active customer, no single model, and it will still answer. Fluently. Confidently. With a number that is wrong, and no warning that it is. A person reading a clumsy report might sense something is off and check. A clean sentence from a confident assistant invites no such doubt. The tool does not reduce the risk of a bad number. It removes the friction that used to catch one, and serves it in language that makes people trust it more.

The chat box is not the product. The model is.

What makes natural language analytics safe and useful is the thing underneath it: a governed semantic model. One place where each measure is defined once, where revenue means revenue, where an active customer is agreed, where the assistant has a single trusted source to reason over rather than a choice of contradictory ones. The assistant is only ever as good as that model. Give it one clean, defined, secured model and it earns its keep. Give it a mess and you have automated the production of confident mistakes. The platform you may or may not need is a separate question, and a Microsoft Fabric estate is not always the answer; the foundation is the same either way.

What to fix before you switch Copilot on.

Four things decide whether turning it on is an asset or a liability, and they are the same four that make ordinary reporting trustworthy. Agreed definitions, written down, so the assistant is not guessing what you meant. One semantic model, so there is a single truth to answer from rather than several. Lineage you can show, so when someone senior asks where a number came from, you can say. And security tied to your existing identities, so the friendly assistant only ever surfaces what each person is allowed to see, rather than becoming the quickest route to a leak in the building. Get those right and Copilot is worth switching on. Skip them and it is a risk with a pleasant manner.

So this week, before anyone buys anything, run one test. Ask Copilot or the data agent a question you already know the hard answer to, and see whether it agrees. If it does, your foundation is sound and you should press on. If it does not, you have just learned where the work is, and it is in the model, not the chat box. Getting that governed foundation right, the definitions, the single model, the lineage and the security, is the heart of our analytics strategy work and of what our Do You Actually Need Fabric guide sets out. The chat box will take care of itself. What it speaks for is the part worth your attention.