Intelligent Diligence: is AI a game-changer for due diligence?

Undoubtedly, AI has the potential to speed up the DD process: rather than the hours which can be spent by advisor teams reading through documents individually and extracting relevant information, AI can be used to identify and compile relevant information from voluminous sources more efficiently for subsequent human analysis.

Imagine an acquiror is concerned with understanding the longevity of key customer relationships within a target business. An AI tool could be used to quickly identify a list of top 10 customers by revenue, then identify whether the contracts for those customers contain change of control provisions that could be triggered on a sale of the business. This could reduce a couple of hours of work to minutes or even seconds, reporting back in a more efficient and detailed manner than a typical “Ctrl + F” search.

A well-organised and suitably populated virtual data room (VDR) has always been key to an efficient DD process. A suitable AI tool has the potential to remove much of the pain of compiling these, by organising information as it comes in and producing easily navigable VDRs, sorted by theme, value or any other category which helps make sense of the data. This removes the need for a person to assess, file and index each document manually or via more “traditional” sorting tools.

Bespoke software already exists to ease the process of compiling and organising a VDR, but the right AI tool could conceivably manage this just as smoothly (or more so) with a lower barrier to entry.

Financial advisors, corporate finance professionals and those in the VC and PE industry will already have a variety of tools and templates available to them for efficiently calculating financial metrics and KPIs. However, an AI tool could supercharge this further by aggregating all available financial information to very quickly produce EBITDA reports, burn rates, return on assets, earnings per share and any other desired assessments.

AI tools can be very useful for distilling information into accessible and easily digestible bullet points and summaries (some, like Google’s Notebook LM, can even produce engaging podcasts from selected source material).  Taking the example of the financial information mentioned above, this could result in not just efficient delivery of metrics but succinct and approachable explanations of why these are important, meaning professional advisors spending less time explaining common technical concepts and being able to focus more on their application to the transaction at hand.

Despite what you might think from the inevitable “I love Excel” mug you’ll find in most office cupboards, not everyone is a fan of spreadsheets. AI could create professional-grade cap tables using requests inputted from a natural-language chatbot and could be set to monitor and continually update these based on legal documents or public filings made accessible to it. Again, there are already software tools out there which can automate much of cap table management, but an acquiror or investor could use this to quickly and easily run its own assessment of the “correct” shareholding structure, and the benefits would be particularly enhanced when combined with some of the other points listed in this article.

With the ability to scan through a VDR and locate relevant information without the need for manual searches, AI could be used to remove human error from the DD process. However, as we note below, there appears to be some work still to be done on an AI’s ability to gather and interpret information reliably to answer specific questions.

Taking legal advice as our obvious starting point, one of the major benefits of obtaining professional legal advice is that good lawyers know which questions to ask and how to contextualise the answers to produce commercially sensible and relevant advice. Much of the value in good legal advice comes not only from a solid understanding of the law (which should be a given!), but from applying that understanding in a commercial manner, often with elements of give and take, in order to arrive at a solution which is acceptable for all parties.

Even if AI could be trained to understand commerciality and market norms, some parties to transactions can act based on their emotions, rather than pure commercial drivers. There isn’t a definitive rulebook on human emotion and empathy, so AI likely does not have the source material or capacity to account for this.

The key underlying point here is that, despite the impressive and approachable responses it is capable of, AI is not sentient and is not capable of truly understanding what it is generating.

AI doesn’t always produce an accurate or comprehensive answer. Many people are now aware of the possibility of “hallucinations” appearing in responses by LLMs (so prevalent that “hallucinate” in this context was Cambridge Dictionary’s word of the year in 2023, ‘Hallucinate’ is Cambridge Dictionary’s Word of the Year 2023 | Cambridge University Press & Assessment), often arising as a result of the AI model predicting the most suitable next word in its output in a probabilistic manner by correlation with the previous words used (we’ve all tried forming messages on our smart phones using the predictive suggestions, which often end up in a nonsensical string of words which almost, but not quite, sound like a coherent sentence). A key concern with AI is that it tends to present its responses with complete confidence and little to no acknowledgement of its scope to be wrong.

It could be difficult to discern where the AI has gone wrong if the evidence appears sound and you don’t already know what answer you should expect.

Looking at the Shakespearian case study from our colleague Imogen, set out below, the answer that ChatGPT gave us is evidentially insufficient; it does not go into any detail about the nature of Shakespeare’s authorship and does not appear to have even considered conflicting views. Indeed, Imogen was very keen to tell us that although the chatbot confidently stated Shakespeare wrote 39 plays, in fact only 37 are agreed to have been written by him and opinion varies on his contribution to the remaining two.

Evidently AI is currently unable to replicate the level of detail and analysis required for certain tasks when it attempts to produce a slick and easily digestible answer to a multifaceted problem. This could be particularly problematic in the context of legal advice, where accuracy and the quality of advice is paramount.

The output of an AI tool is often only as good as the information fed into it. If a business has been poor at keeping written records and ensuring its information is up to date, it would be difficult for AI to produce accurate and actionable insights based on that information. A human advisor should be able to spot and question any material weaknesses in the source material (e.g. inaccurate share capital records, or failure to follow generally accepted accounting principles), allowing this to be addressed before analysis and output is completed.

Examples of text-to-image generative AI have proven to be subject to the influence of prolific biases based on those that occur in its training data (Leonardo Nicoletti and Dina Bass have produced a really interesting article on this for Bloomberg UK). Could this mean that AI tools used for DD could produce results based on conclusions drawn from what it expects the answer to be, propagating bias within its training materials? If so, how do we police the bias?

There are clear risks associated with asking a third-party entity (i.e. an AI tool) to process information. A typical VDR will include a wealth of sensitive information pertaining to a business and its clients / customers, and a data breach or leak could have material implications. Insurance products can help to manage the financial impact of these risks, but no business is going to be fully comforted by insurance if their most commercially sensitive information is leaked to competitors.