AI for Due Diligence Use Cases

Either wait or you have to work very hard to get it right.

One of the top expectations of many professionals we work with is that they want to examine large piles of documents for due diligence or review using AI.

Unlike in other contexts, say - dashboards, expectations in a review or due diligence usually are - accuracy and reliability.

The technical approach most popular to try to solve for this includes - (a) converting documents to text, (b) running a search and generation using AI, also know as RAG.

RAG is short for - using traditional methods to search and then feeding the results with limited context size (word count) in a language model to produce a statement.

Let's keep aside the complexity of accurately converting different formats of documents and images into legible text.

Over and above errors that creep in there, we have found that inaccuracy compounds because (a) traditional/semantic searches on large corpuses of text have high error rates; and (b) LLMs have their own error rates, and inherent weaknesses (example, context size and yes/no questions).

error % x error% makes depending on AI unreliable for DD use cases because error percentages so high mean the whole effort has to manually repeated for reliability.

TLDR - There are ways to power your DD process using AI, but don't expect it to cover the whole workflow.

Here is a good post by the great

Timothy B. Lee

that explains this from a technical standpoint.

Which is why despite all the great advancements in language models recently, I would still wait before investing in document review and diligence technology that claims to be powered by AI.