LLMs are getting exceptionally good at interpreting vague asks.
One of the biggest issues I've seen in this kind of environment is that there are 50 dashboards, created by different people, and the data presented on them hasn't been validated to ensure that the data is correct and pulling from the same location. It's not clear that the data can be trusted and it's not clear what the authoritative source of truth is.
IE. They know what they're looking for. They know when the data is right or wrong because they've seen it from other places here and there over time.
When they're not sure and the risk of getting it wrong is costly, they'll find the data analyst to verify.
I've seen it firsthand, multiple time and while anecdotes aren't data, conversations I've had with peers at other companies suggest they have similar experiences.
Furthermore, that persona is also often incapable of self-checking and recognizing errors in their presented data, because they don't know how it was generated in the first place.
In the past, you could look at a web page, see some structure, understand process, etc. and intuitively see and understand what you're suppose to do. When you remove all of that in favor of a text box, you spend a lot of time asking questions and prompting to get information or accomplish tasks that would otherwise have taken a minute or two and when you're done, you still don't understand how the underlying system, process, or policy is and how it works.
I meant that non-developers will also find that staying within their AI agent to do what they want is better than downloading, logging in, learning different app UIs.
In our company, we recently gave Claude subscriptions to all office workers. They all hooked it up to their favorite tools already and they've been raving about it. It was a revelation for them.