I recently received an email from a popular analytics provider telling me I was one of the most active users of the product, letting me know that I could save my queries into a dashboard, and that they could show me how to do that in a quick 10 minute call.
I never responded.
I don't want to spend 10 minutes on a call talking about your dashboard for the same reason I don't save my queries. I do not care enough about the prescribed way to hold the product, instead I am more than happy to burn tokens and regenerate the view each time I come in. Why? Well because even though the prompt might look very similar to the previous question I asked, it's probably a little bit different, the questions I need answered today aren't necessarily a 1:1 to questions I asked in a previous session. Sure there's probably a lot of
overlap, but the parameters change, the time frame, the goal is different. If I end up having to tweak a previous session's output I might as well start over from scratch.
As long as the token supply is practically endless, I think this same mentality is going to carry over to many other workflows. For instance, I recently built a dashboard to track my finances. I literally never use it. It was great, pretty, I spent a bunch of time polishing it up, but I haven't opened it back up since the day I created it. Instead it's easier for me to feed my transactions into this month's agent of choice and have it re-review and flag any discrepancies or outliers.
Right now we treat these interfaces like they are our pets. We maintain them, check on them daily, and nurse them back to health when they break. I think that's going to change, and dashboards will end up closer to cattle: something to consume today and grab a fresh steak the next time I'm hungry.
The obvious pushback here is going to be consistency. How can we make sure the user is consuming the interface in the appropriate manner? I think there are a number of solutions to this without overprescribing for the problem.
First the data layer is always consistent, it's the source of truth and, barring any sort of hallucinations from the model, it will always ground the interfaces that are being generated.
Another idea is giving out approved components. What if your API/CLI/etc… shipped with skills that generated the right point-in-time component? Whether that is a table, chart, etc… you could package those up as skills or blocks that live alongside your data giving your users the best possible way to experience your data. If at that point there's inconsistencies in the column ordering in a table, or showing a line instead of a bar chart it's really not a huge deal.
There are going to be exceptions of course, I think for a long time there will be a need for a traditional dashboard, a place where you can go when something is mission critical, or if you need a fully repeatable way to produce a result. However for personal inspection or even getting answers to quick questions, generating it yourself oftentimes feels a lot faster than navigating someone else's user experience.
So sales rep from big analytics co, no, I will not be saving any dashboards. I've already got my question answered, and for tomorrow's questions? I'll just prompt again.