Building PicoAsk: where 'just send your data to an LLM' breaks

https://www.producthunt.com/products/picoask/launches/picoask

We are building picoask. You ask a question about your data in plain English, and you get an answer backed by evidence and transparency. Not just another chart to figure out yourself.

While building, we kept hitting the same idea: "AI analytics" just means pointing an LLM at your database and asking a question. Basically, text2sql. Looks great in a demo, breaks in real use. Here's why

It answers what you asked, not what you mean. "Why did the signups drop last week?" turns into a query that spits out a few numbers. The numbers are often right and useless. It doesn't tell you if the drop is real or just noise.

It ignores sample size. LLM will cheerfully tell you A beats B based on 30 data points vs 4000. It has no sense of when there is enough data to trust the answer.

It can't dig in. Real analysis is a loop: form a hypotheses, test it, get surprised, and ask why. A prompt-to-answer skips all that. You get a believable story, not the real cause.

It sounds confident. LLMs are confidently wrong, doesn't matter if the answer is right or made up. You can't check how it got there.

So we flipped it. Instead of "fetch and phrase", picoask comes up with many hypotheses, tests which one holds, and keeps digging like an analyst would. Every claim backed by the tests behind it. And we openly showcase the work done.

I never believed a plain text2sql adds much value. Most of the analysis time is spent in problem understanding, forming intelligent hypotheses, testing the data for significance and then coming up with actionables. Plainly, things one has to do before and after the sql.

Curious what people think. Would love the counter-argument.