The bottleneck was never the data. It was translation.

• Non-technical people queued behind analysts for one-sentence answers.
• Static reports couldn't answer the obvious follow-up, so every question restarted the tour.

The obvious brief was "build a better dashboard." Twenty-three interviews with finance, ops, and compliance killed it.

• One pattern under every complaint: the tool makes me translate my question into its language.
• Nobody wanted a better dashboard. They wanted to stop needing one.

One line from our CEO: "we should be able to chat with documents." Turning it into a product was mine. I scoped v1 with the CEO and PM:

• What a query could touch.
• How documents came in.
• What we'd show on screen.

The safe version was a chatbot bolted onto the dashboards, an assistant in the corner of the maze. We rejected it early.

• A bot that answers questions about a maze is still a maze.
• If natural language was how people thought, it had to be the whole interface.

"Excluding returns and taxes" carries real business logic. The system should translate intent, not make users do it. But a translation you can't see is just a black box, and finance won't sign off on numbers they can't audit.

• Every answer shows its work: what it read, what it filtered, how it totaled.
• Visible reasoning turns "a number" into "a number I'll put in front of my CFO."

Every tool we studied put visualization somewhere else: a tab, a builder, an export. But the point is the follow-up. "Now break that down by outlet" only feels natural if you never left the thread.

• Charts render inline, and the conversation keeps going.
• A separate charts tab would have been the maze sneaking back in.

Power users needed their own data on the table: invoices, POS exports, P&L, outlet masters, ready to be questioned. The lazy version is a setup wizard; the clever-looking version sprays settings through the chat. We did neither.

• A side panel holds everything you've connected, summoned when needed, invisible otherwise.
• Clean for the person who just wants an answer; depth one click away for the person who doesn't.

A chat interface has a dirty secret: an empty input box is scarier than a bad dashboard. New users don't know what to ask, so they ask nothing and leave.

• Suggestion chips propose the next question from what's connected and what was just asked.
• Not a convenience feature; it's how the product teaches its own ceiling.
• The rejected alternative was onboarding tours and docs, for a product whose whole promise was no learning curve.

Every dataset lives in its own space: invoices in one, reviews in another. You land on the workspace you've built, and adding more is one click away.

Add opens a picker of files and tables: everything a new space should see, chosen before the first question is asked.

Sources land in the Data panel. Nothing has been asked yet, but the conversation can start.

The real question, typed the way you'd say it out loud, with no syntax or field names. It answers with a reasoning drawer, the prose, and a table where every figure traces back to its rows.

A follow-up becomes a real message, answered by a chart that renders inline. You never leave the conversation.

Custom business rules, like "treat anything below three stars as negative," quietly applied to every answer.

Time-to-insight went from 15 minutes to 2, not because people got faster, but because the tour of five dashboards stopped existing.

• Within a quarter, 67% of queries came from non-technical people.
• The analyst queue dissolved, and the founding promise reached the people it was made for.

Chat removes the dashboard, but you're still typing. The next step I designed was voice.

• Ask out loud, and drill in by just asking.
• Text was the right v1: precise, auditable, shareable. Voice is the v2 that matches how people already think out loud.