The recommendations don't look very insightful, and seem like a rephrasing/summary of the alert above it. For example the first student who has account holds, bad grades, etc. the recommendation is just to schedule a meeting. I don't think the LLM will be able to provide super insightful recommendations. Even in the improvement plan generated by the agent, the steps seemed pretty generic(as expected from LLMs).

I do think you have value in pulling in the disparate data sources and using LLMs to present the data in a clean way to the advisor/user.

That is great feedback, and agreed that LLMs definitely have generic outputs, especially if missing the right context. To combat this, we're actively working on playing with which data we can pull, how to cleanly give it to an LLM and which models to use to improve the inference (while staying within the compliance boundaries).

We've found the "chat" functionality to be especially useful for advisors since we've been able to surface insights to them without them having to log onto many different systems and just present it in a clean output, as you pointed out.