LLMs can already write correct SQL from a plain-language question, faster than an analyst can type it. But real business data has traps built in: an order that's cancelled versus returned, a transaction that's debit or credit. Ask an AI agent the wrong way and it will pick an answer, confidently, without knowing it should ask.
We tested this ourselves, on a realistic simulated data warehouse: accuracy on real business questions started at 87–93%. It reached 100%, reproducibly, once we wrote down what a good analyst already knows and gave the agent access to it. That's the difference between a demo and something your business can actually trust.
Three habits, not one big project.
If you already have pipelines, models, and a well-layered platform, with personal data handled properly, you're well positioned to start. If you're starting from scratch, do the foundation work first. The conversational layer comes after, not instead.
Choose one business goal that matters. Work it the way you always would: source, pipeline, model, dashboard. Then add the conversational layer on top, and build a small set of test questions for it. Iterate until it's right, then move to the next domain.
You cannot just feed all your data to an AI. Decide what it may see, ideally only aggregated, anonymous data. Confirm it isn't stored or used to train a vendor's model. Build this in line with GDPR from day one.
Fewer questions stuck in the queue. Quick lookups get an instant answer instead of a wait on Slack.
Faster access to the deep data in the warehouse, for the harder questions a dashboard was never built to answer. A shortcut, not a replacement.
The platform used to give the same answer to the same question every time. Now many more people are asking, in their own words, and someone has to make sure "sales" still means the same thing across the business.