[code_biologist]

If you're targeting use within software and engineering teams, that thesis may be right. If you're targeting adoption across whole businesses, I think the thesis is pretty wrong and will end up hampering adoption. To broadly bucket BI challenges, there's first the challenge of getting people to use the thing, then the challenges that come when everyone is using the thing. Tech types seem to underrate the challenge of getting people to even use a BI tool in the first place.

I've found self serve to be a really effective tool in getting engagement with BI. My onboarding for new non-tech BI users was always to have them build a basic dashboard for the business process they were most focused on. Maybe set an alert or create a scheduled report delivery. By the end of a 15 or 30 minute onboarding session you'd see the click as they realized what they could do with it.

That mess of never ending dashboards has another name: BI engagement. Though a product can help, having core dashboards and KPIs is a social and analytics leadership problem and not a technical one.

Though I have issues with Looker (their dev experience is crappy), their approach to this is effective: make it difficult for self-serve users to get incorrect or nonsense answers, and make it easy for analytics admins to designate core dashboards and jockey a few hundred custom dashboards and reports as the underlying data models change. Every business unit got pretty attached to what they'd built for themselves.

[louisjoejordan]

You're spot on that BI adoption is largely a social challenge. Our thesis is that by defining the entire journey from source to viz as code, we create a structured foundation that LLMs can build upon, democratizing access to the transformation layer for non-engineers in a way that point-and-click BI tools can't.

[ringobingo]

Can you please elaborate on how you see LLMs could build upon this model/journey?

[beardedwizard]

Llms would generate the code/definitions underlying these dashboards, presumably a model could be trained for the task. I'll argue it trades one version of the sprawl problem for another. Unless this generated code is easy to debugs and comprehends other generated code, it will still be a spaghetti mess at scale.