AI in eDiscovery ‘in action’ - which AI features are being used and three real-life examples

eDiscovery, in a nutshell, ensures that legal reviews and other types of document review are as efficient as possible (in terms of speed and cost), by reducing the ‘review’ population down to a sensible and proportionate size, through the use of technology.

Various forms of AI are now available but are still not being used enough.

AI is extremely important in the context of litigation and investigation and should be thought of as ‘Augmented Intelligence’ rather than ‘Artificial Intelligence’.

AI should be utilised in one form or another on all cases using an eDiscovery platform because it is already ‘ready to go’, and in many cases free of charge.

Using AI in litigation and investigation

We are now using AI on a large percentage of our clients’ cases, although not every case due to:

• Perceived large costs
• Perceived inaccuracy – the ‘hallucinations’ (mistakes) reported against ChatGPT, Gemini, Meta AI and Grok fuel this fear.

We need to view the AI tools within eDiscovery review platforms as simply something to ‘augment’ a review process, not to replace anything (or anybody), and merely consider them as an additional ‘pair of hands’.

In the past, many proponents of AI have either:

• over complicated the explanation, or
• over hyped its abilities leading to low take up and low trust. 

This has meant that legal advisors have found it hard to convince their clients of the benefits and therefore been limited to more standard searching and filtering.

Using AI in litigation/investigation is now far more user-friendly and usually built into the systems at no cost.

Three examples of AI that we use are listed below, explained by real life case studies:

TAR/Machine Learning

• 900,000 client docs requiring assessment
• ‘Standard’ feature used – Keywords/Dates/Filters were applied to reduce volume down to 80,000
• ‘Machine Learning’ feature used – a lawyer then ‘trained’ the system by tagging documents for relevance
• A ‘likely relevance ranking’ score was then applied to all 900,000 documents
• A proportionate % was agreed upon i.e. >65% likely relevance
• 45,000 documents were excluded from the 80,000 documents
• 35,000 documents were then considered for review

Therefore, only two days of assessment using a combination of ‘standard’ searching and filtering, plus ‘machine learning’, resulted in a reduction from 900,000 to 35,000.

It is worth noting that if the other 45,000 had required 1st Pass review, at an average rate of 100 docs per hour, this would have required an additional 450 hours, meaning a saving of £27,000.

Sentiment Analysis

• 380,000 documents from four family members mailboxes
• System analysed the sentiment/emotion of every document
• We then selected ‘High Negativity’ AND ‘Emails’ = “find me the negative emails”
• 1500 documents immediately identified for immediate ‘priority’ review 

A better understanding of the family relationships and arguments was immediately possible, leading to a more targeted review strategy for the rest of the database (which was of course then searched with more traditional filters).

The point here is that due to the claim value, it would not have been proportionate to attempt to search/review 380,000 docs without the use of more sophisticated features. A workflow using AI meant that all parties could agree to a reasonable search.

‘ASK AI’ (Generative AI)

• 180,000 emails from three Directors, plus Sharepoint
• Pertinent questions were asked in a search box, such as:
  • “which directors were present at discussions about investing in X?”
  • “were any concerns raised about debt?”
  • “list the investments made by the company between 2019 and 2021”
• ‘Generative AI’ then provides:
  • A ‘natural language’ answer to the question
  • bullet points to better present the answers
  • Lists and values depending on the question
• Importantly it also provides:
  • The reference documents that it used to create its ‘answer’
  • The reference sentences or parts of the document it focused on 

The quality of the answers depends on the quality of the questions, meaning that, as ever, legal input and knowledge of the points in dispute are essential, both to ask the right questions and to be able to analyse and interpret the answers.

In essence, what we are saying is that this feature is not a replacement for legal professionals and there should be no fears of ‘Rise of the Robo Lawyer’… it is merely to augment legal workflows.

In summary

The AI ‘tools’ mentioned are merely an evolution in how ever-increasing volumes of data, from disparate sources, are searched.

Using ‘Augmented Intelligence’ in litigation and investigation is steadily becoming the norm, not as a replacement for work, but as a way to capitalise on advances (and increases) in data and provide time and cost benefits for both the end-client and the law firms using it.

It’s worth mentioning that as easily as these tools can be turned on, they can be turned off. Not all cases will necessarily benefit from them, but it’s worth testing them out at the earliest opportunity in the case to find out.