Context is the opposite half of the story, and the context panorama is actually a multitude. Distributors are working laborious to persuade you that solely their semantic modeling language can prevent, whereas it isn’t fully clear whether or not these are obligatory and even enough. Whether or not you retain your corporation logic in a semantic layer like MetricFlow or Malloy, or simply in plain Markdown, the aim is similar: get that logic right into a type an LLM can use. Context is nearly at all times created by hand, and like all hand-written documentation, it begins drifting the second it will get written down.
This highlights a possibility, particularly that brokers are good at exactly the components of context which are mechanical and unhealthy at exactly the components that aren’t. An agent can infer which tables be part of to which, what values a column tends to carry, what your gross sales areas are, and which tables folks really question. What it will possibly’t infer is the stuff that was by no means actually a knowledge query: the proper approach to calculate income, what counts as a “buyer,” when the fiscal yr begins. These aren’t information hiding within the warehouse ready to be discovered. They’re selections, usually enterprise ones, that an individual has to make. What an agent can do is flag the second certainly one of them quietly stops being true.
Automated agent insights stay a fantasy
The flashier pitch, the place brokers floor insights you by no means requested for, is the one I’d guess on final. It sounds fantastic to have hands-free analytics. An agent will hold watch over your knowledge, discover what issues, and drop a dashboard tailor-made to no matter is going on right now. However the bar is excessive for relevance and false positives could make human customers lose confidence.
