When we started building Logit, the first prototype did exactly what people expect from an AI analyst. You typed a question, it thought for a moment, and it gave you a clean answer with a chart. The demo went well. Then people asked the question every analyst hears in their first week: where did this number come from?

We didn’t have a good answer. The model had written a query, run it, and summarized the result, but none of that was on the screen. The answer looked finished. It wasn’t checkable.

The shortcut we didn’t take

The easy fix was to make the answer look more confident. Add a source label, round the numbers, and keep the interface clean. It works until the first time an answer is wrong in a meeting.

We took the other path. When an answer comes from a query, you get the SQL, the tables it read, and a way to open the rows underneath. It makes the screen busier. It also means you can check the answer without asking us.

What showing the work means in practice

When you ask Logit a question, the answer has four layers. You see them only when you want them.

The first is the answer itself: a sentence or two and a chart. Often, that’s all you need.

The second is the query. It’s SQL for your warehouse, with comments on the filters, joins, and date logic. You can copy it, edit it, and run it yourself.

The third is the data. Click a number, and Logit opens a sample of the rows behind it. If a total looks high, you can see which orders make up that total.

The fourth is the list of checks and assumptions. If Logit had to decide what “last quarter” means, or whether to count refunds, it says so. If a join could multiply rows, it tells you how it checked.

Why this is slower, and why that’s fine

Showing the work costs time. Logit checks its own query before it answers, which makes some answers slower. We decided that’s cheaper than a wrong number in a board deck.

It also changes how teams use the answers. When an answer can be checked, people forward it. When it can’t, they ask an analyst to redo it, and nobody saved any time.

What we’re still getting wrong

Not every question has a clean query behind it. Some depend on definitions your team never wrote down, like what counts as an active customer. Today, Logit asks you when it isn’t sure. We’re working on letting teams save those definitions once, so every answer uses the same ones.

We also know the extra layers can feel like a lot the first time. We’re testing ways to show less by default without hiding anything.

If an answer ever surprises you, open the query. If the query surprises you too, email us at hello@logithq.com.