The Fifth Elephant 2024 Annual Conference (12th &13th July)

LLMs have transformed data analysis. With their ability to generate code to analyse data given appropriate prompts and instructions, LLMs are forming the bedrock of a new suite of data analysis tools.

However, thanks to their ability to hallucinate, and that they are occasionally very poor logically, the role of LLMs in data analysis so far has been restricted to being copilots, who human data analysts can “pair program” with.

This talk is about how LLMs can go beyond being just copilots - how we can effectively pair LLMs with more conventional statistical methods to produce true “AI Data analysts”, that can generate insight and tell stories from data without any human prompts.

Outline

• Data analysis capabilities of LLMs
• Where LLMs fail at data analysis
• How smart human analysts analyse data
• Can LLMs plan? If they can’t, how can they be effective data analysts?
• Putting together LLMs and statistical methods to generate data insight
• Iterative reasoning and insight using a multi-agent framework
• Using LLMs to tell effective data stories

Impact

All of this is core to the product we are building at Babbage Insight, with the objective of delivering Exceptional Insights to business decision-makers.

This approach will enable our product to anticipate questions that business decision-makers might have, and present them with a personalised news feed of key stories about their business.