Once, there was an answer
It was a good answer. Someone had asked how revenue moved month by month, and a few seconds later there it was: a clean line, two years long, with a dip in the spring that nobody had noticed before.
The previous chapter was about how hard it used to be to get an answer like that. This one is about what happened to it next. The answer existed. It had nowhere to live.
Every day, it was lost or boxed in
There were two homes on offer. The first was a chat thread, the way AI assistants work. A thread is a good shape for a conversation, which runs in one line from top to bottom. An analysis does not run in a line. It branches.
So the person asked the next question: break it down by region. A second chart arrived and pushed the first one up. They noticed one region and asked about it. A third chart arrived, and now the first was off the screen. The more they learnt, the harder it was to see what they had learnt. When they closed the window, the work became a transcript.
The second home was a dashboard grid, the way BI tools work. A grid keeps everything in view, but someone has to design it before the questions are known, and analysis rarely arrives in a tidy three-by-two layout. Our answer could have a box there, if somebody built the box first, published it and promised to maintain it.
One home forgot the answer. The other made it wait for furniture.
Until one day, we asked a different question
Most of the effort in this field goes into making answers: better models, faster queries, nicer charts. The question that shaped this product was a quieter one. Once the answer exists, where does it live?
It needed somewhere that kept it in view, as a grid does, without anyone designing the room in advance, as a thread does not need. It needed space that grows with the question. So we gave it an infinite canvas.
Because of that, an answer became a thing you can pick up
On a canvas, every chart, figure, table and written insight is an object. You can move it, resize it, duplicate it or delete it. Nothing is locked into a layout someone else designed.
The follow-up sits beside the chart that prompted it, so the first answer is still there when the third arrives. You can select a chart and ask to compare it with last year, without restating the measure, the filter or the time range, because the question knows what you are pointing at. And the charts that answer one question can be grouped, so the canvas stays readable as it grows.
What you asked, what you found and what you asked next is the shape of an investigation. It is worth keeping, and a canvas is the only one of the three that keeps it.
The features page shows the canvas in more detail, and the comparison of conversational analytics and traditional BI sets the approaches side by side.
Because of that, the numbers had to hold up
An answer that stays in view gets looked at again, and by more people. That raised the stakes. A free-form surface is only useful if the numbers on it are right.
We did not want to ask anyone to trust a new, closed way of calculating figures. So charts are built on Apache Superset, an open-source analytics project that many data teams already run and know.
It also settled what the AI is for. A language model is good at understanding a question and poor at arithmetic you can audit. So the model's job is to work out what you are asking, and the figure itself is calculated from your data, the same way each time. Every chart has a How it's calculated view that lists the steps in plain words. The guide to how a question becomes a chart explains it without the jargon.
Because of that, we met the canvas's limit
Our answer finally had a home, and for a while the canvas was the only view. A dashboard was the thing we had set out not to build.
That was too strict. A canvas arranged for your own thinking is the right place to work something out. It is the wrong thing to open on a phone, and not what you want to send to someone who only needs the headline figures. A board laid out by one person does not read well to another seeing it cold.
Until finally, the answer had a home and a way to travel
So canvases now open as a tidy dashboard, with a switch to the free-form canvas when you want it, and phones always show the dashboard. The canvas is where the work happens. The dashboard is how it travels. What's new has the release.
Our line chart, the one with the dip in the spring, is still there. It sits beside the regional breakdown it led to, in a group with the chart that explained the dip, and the person who needed only the headline saw it as a dashboard on their phone.
And ever since
Would we make the same choice again? Yes. Almost everything that sets the product apart comes back to the canvas: answers that stay put, follow-ups that keep their context, work a colleague can open and question. The lesson was narrower than the decision. A good place to think still needs a good way to present.
It was not our only wrong turn. The next chapter covers what else we got wrong. Or judge the canvas for yourself in the playground, which needs no account.