AI + legal: use cases, data quality, gold nuggets and answering questions

A picture of some gold nuggets

Speed read

• Getting success from AI relies on your picking and identifying a specific use case. Don’t just stop at “use AI”, but be really specific on what the purpose of using AI actually is
• If your use case relies on your own specific data sources, work out where that data is and understand how you are going to separate good data from bad data
• Don’t shy away from non-technological or non-AI methods of achieving the above goals
• AI can produce excellent first drafts of documents, but work out whether this is actually the best thing for your specific situation
• It’s long been a dream of legal technologists that AI can replace things lawyers don’t want to do, like sharing knowledge. Don’t get too sucked into this too much in the short term

This question posted on LinkedIn by Colin Bennett last week caught my eye this morning:

Serious post: if you have a legal AI solution that will 1) take every single document in my very messy and unstructured file system (contracts, maybe emails), 2) can ingest/train in semi-real time on an ongoing basis, and 3) answer queries regarding my files with citations/references (including reference to broader pre-trained legal knowledge), then let’s talk (note, no offense but only looking for extant solution here, not “this is on our roadmap”).

With excitement still very high around AI, this kind of question is being asked by legal leaders fairly regularly. At its core, it asks whether an organization can leverage its own work product, files and knowledge to supercharge the services it offers. The use case, while not 100% clear from this post, is usually something along the following lines:

• Contract management — scanning executed contracts and extracting core provisions, putting them into a database so that obligations can be tracked easily. This solves the pain point of organizations missing contract deadlines, and people having to spend time repeatedly CTRL+Fing both in their DMS (to the extent they have one) to find the right contract, and then doing exactly the same within the contract to find the correct part. This can also happen on a large scale, and in that sense can sometimes overlap with due diligence or mass amendments to contracts
• Reusing work product or precedent — for a legal team, this might mean finding all memoranda and advice sent to them by external counsel on given legal points so that they can answer legal queries without having to pay for advice they have already received, but cannot retrieve. For a law firm, it might mean finding example contracts that can be repurposed for a similar niche situation, it might mean finding prior cases/adjudications that are similar to the current situation and can be used for precedential value, or it might more broadly mean answering a question that has already been answered by somebody else.

The reference to “including reference to broader pre-trained legal knowledge” makes me think this use case is probably more the second than the first. It is of course possible that the use case is entirely different. As always, this is the very first question that should be addressed in any project of this nature.

One of the first things that intrigued me about the post was the number of vendors jumping to say they can achieve what Colin has in mind, despite the use case not being abundantly clear — but I’ll leave that to one side for now. Here’s a few other things that interest me about this question

Data quality

Whenever you’re thinking about AI, you should first think about data quality — specifically, “what is the data the AI is using to generate the output, and how does this impact my use case?”.

In the first two parts of the question, Colin is looking to include “every single document” in a messy and unstructured file system — both historic, and present. He’s also looking to include “broader pre-trained legal knowledge”. The second can be accomplished by using data from approved authoritative sources, such as legislation, case law and curated third party sources (e.g. Practical Law).

The first is more challenging. It has long been a dream of legal technologists (and presumably this dream exists in many other industries as well) to be able to extract the “golden nuggets” from a document management system without any effort at all from lawyers (who are notorious for not wanting to do anything that they deem unnecessary in the context of the job they have at hand). The question is always “what are the golden nuggets I need to extract for the use case I have in mind”.

I’d encourage Colin to think about whether it is the case that his use case relies on every single document. For example, if the repository is a general document management system that hosts a bunch of drafts, prospectuses, redlines and emails, this presents challenges for both the use cases I mentioned. In experiments I have been part of around AI, I’ve seen situations where a user wants to ask a question about their own organization’s anti-bribery policy, only for the response to be generated by an email footer they received from a third party 5 years ago. Data quality matters, and if you ignore it, the output you get from these kinds of systems suffers.

So the real question is, having settled on your use case, what are the “golden nuggets” required to drive that use case? Where is that data stored? Does the location store things other than the golden nuggets? If so, how are you going to separate the golden nuggets from everything else?

Panning for gold

There are a few approaches to “panning for gold” in data repositories:

• Direct extraction — when you ask a rules-based or machine learning algorithm to cherry pick specific documents based on fixed criteria or a pre-determined dataset
• Indicators — work out the kinds of characteristics that might indicate a golden nugget (e.g. does it have the word “executed” in its name, how has the document been classified etc.). This creates a smaller sub-set of data less “noisy” than the rest. It might be this is sufficient to resolve data quality issues, or it might allow human curation to be done more easily
• Human curation — deploy processes that allow humans to easily earmark documents that are “golden nuggets”, and label these as such. For example, insist that lawyers submit all finalised/executed documents to a specific place

The method you choose again depends on the use case. In a contract management use case, it might be enough to use a combination of (1) and (2). You could rely on things like naming conventions, whether or not something has an e-signature certificate attached to it, contract classification etc.

In a knowledge management use case, it’s far from clear that AI is able to determine what is a “good” document and a “bad” document as this often depends on information not contained in the document itself. In this example, (2) and (3) are probably a better example, although you could look to do (1) as well depending on the use case. For example, somebody trying to spot market trends for a given document might be able to use (1) and (2) but somebody looking to create a set of contracts that created the best outcomes in negotiations would find (2) and (3) more fruitful.

Whenever we start talking about involving humans here, the answer is nearly always “but attorneys don’t have time, and we can’t make them”. Then the conversation stops, and usually so does the project. But clever organizations try two things to enable human curation, and unlock some of these use cases:

• Culture and incentives — they don’t just admit defeat and say “it won’t work because our attorneys refuse to do it”. They put in cultural measures (e.g. leading from the top, setting examples) and incentives (e.g. monetary and reputation) to encourage the behaviour they have in mind. They don’t try to change the world, just one thing at a time and accept that culture moves slowly, but that it can move
• Process definition — they try to leverage human curation as part of processes users do anyway. For example, can you create specific places in your document management system for “executed contracts” — which not only help attorneys find things quicker and organise themselves better, but also enable you to add some structure to your data.

It may well be the case that we end up with AI in a sufficiently developed state that it can extract any kind of golden nugget we want. I don’t know how far off this has been, but I don’t think waiting for this to happen is a good idea. I say that for two reasons. First, there are things you can do in that interim period to move the dial. Second, we don’t know if it will ever come. Third, it seems a reasonably good assumption that any form of AI might rely on sufficiently-good data quality being in place.

So the strategy should be to work out how you “pan for gold” in the relevant data repositories, and start deploying that method — whether it involves technology in whole, part or not at all. Unfortunately, it is rarely the case that a single tech vendor can say they’ve solved this for you — it will undoubtedly require efforts on your part to understand what data you need and how to get it.

Answering questions

The final thing that intrigued me about Colin’s post was the reference to “answering questions”. I’ve seen at least three ways in which legal technologists talk about the application of AI on your own data:

• Finding things more easily — no questions answered, just find me the document (and probably the relevant parts of it). I’ll consider it, read it, and write up the answer myself
• Answer my question — do (1), but also draft me a response, preferably with your sources so that I can consider it, read it and change it myself
• Do a first draft of the overarching process — I want to ask a simple question, I want AI to work out what questions to ask, answer them, then produce a first draft of the response. I’ll then check the response and amend it accordingly

The first of these is what would probably appeal most to a traditional lawyer, because it’s a better way of doing what they do now. It just cuts out (hopefully) the repeated searches they have to do to find what they need.

The second and third can be seen to encroach on what lawyers have done historically in their work, because it starts to get involved in the drafting itself. And not even from a “check what I’ve written” perspective — the ask here is for AI to do the first draft.

I’ve written about AI first draft use cases before, and I’ve historically not been a fan, for three connected reasons:

• Hallucinations — even where you are connecting AI to a trusted data set, you can’t completely eliminate hallucinations. Don’t get lulled into the marketing hype of AI companies — these things at their core predict words, not produce factually correct output. The difficulty is that word prediction does often produce factually correct output, but not always. Where a trusted data set is involved, the risk is mitigated, but then again, you don’t know what has originated from the trusted dataset and what has originated from the training data of the foundational model
• Review difficulty — if a junior lawyer presented me with a document with visible typos, I’d check it extremely thoroughly because it gave off the appearance of a carelessly-written document. But AI tends to produce confidently and well-written documents. As a result, reviewing these things for mistakes and inaccurate citations can be challenging. Furthermore, you’re only reviewing what you see, not what you don’t see
• Cognitive processing — for some complex legal questions, the value in answering them is really in you, the lawyer, understanding things, rather than the words you end up writing on the page. The more time you spend on the subject-matter, the more opportunity you have to understand them. If you haven’t actually read any of the source material, you probably won’t understand the subject-matter as well as somebody who has

I’ve mellowed my views on this somewhat, and I think the suitability of AI-written output depends on the kind of organization you work in, your business model, your own knowledge, the process in question and the clients you serve.

For example, if you work in the complex commercial litigation practice at a huge law firm with clients with deep pockets, your clients are probably wanting more than just a bunch of legal memos and one-word answers. They’ll expect you to strategize with them, know their business and the legal context like the back of your hand and to be their true partner. If you are answering knotty legal questions, I don’t think you’d come across well if your response every time is “let me just ask my AI”.

However, there may be situations where clients value timeliness of response over you gaining a better understanding of things. In less high-stakes matters, or with more straightforward questions, it might warrant you using AI to produce a first draft, checking it and sending it out. Perhaps the project does not fully merit you spending more time really getting into the weeds of the issue.

And then again, we have situations where you already know the law very well indeed, and you can spot things that don’t look right from a mile away. Perhaps you have done hundreds of cases on directors’ duties in bankruptcy scenarios, and you are being asked for a summary of what might happen in a given situation. Maybe the thing that needs automating here is the synthesis of knowledge already in your head, and perhaps you can craft a sufficiently detailed prompt pointing to specific cases that AI can then produce a first draft from.

Yesterday I read with intrigue a post by Kyle Bahr entitled, “ChatGPT has Moved Past “The ChatGPT Lawyer” — the Legal Profession Should Follow”, which explores some of the new capabilities in ChatGPT in producing legal advice. This is useful material, and we should apply it correctly depending on our scenario. Sometimes, checking your citations is all you need to do. In other situations, maybe it isn’t such a good idea to put AI in the driving seat.

For the purpose of Colin’s question, it seems like he has more pointed questions in mind. Here, I think a combination of techniques (1) and (2) is sufficient, but I would encourage him to find a solution that produces a response with the relevant sources, and to verify those sources each time.

I hear questions like this an awful lot, and ultimately answering them comes down to evaluation of three things: (1) the specific use case, (2) the data you need and your ability to obtain it, and (3) the process and technology you can deploy to achieve (1) and (2).