EVERYONE is Missing the Most Important Part of Contract AI

Transcript

Hello and welcome, I am Josh from Nomio, and I am the only person in the world who has spent the last seven years, as their full-time job, and their passion, building and maintaining contract repositories, and doing nothing else.

Today, I want to talk about how the industry is misleading everyone, and is being really sloppy with how it’s talking about contract AI.

So, I’d like to set a little bit of that straight today. We talk about AI for contracts as if it’s just one big thing, but it’s so untrue, and, even in the world of asking questions about your contract repository, or your existing set of signed contracts, there are two major ways that AI plays a role, and yet our sloppy language has collapsed it all down to one, and it’s resulting in lots of people making some really bad mistakes, We are using tools like Claude, ChatGPT, Gemini, Copilot, to ask questions about our contracts.

And this is exactly what we should be using these tools for. They are excellent for this. So questions like, which active supplier contracts come with early termination penalties?

We ask our AI the question and then shortly after it gives us an answer. The big problem is that this is just half of the picture.

And the other half is more important, and what gets me annoyed is that no one is talking about it, no one seems to be thinking about it, and yet it’s so important, and that is, what are you feeding into Claude so that it can answer your question?

Most people are very naively just doing this. They’re just smashing a big pile of documents in a messy file system into this AI, and they’re expecting good results.

Now, I’ve already done a bunch of videos, so go and check those out, on why this is a terrible, horrible idea.

And you should be doing something like this instead. You should be preparing the question surface. So that you get much better answers, so that you get them much faster, and so that your AI is, literally ten to a hundred times cheaper than if you’re doing this.

And one thing I like to ask people is, do you really think you’re going to get identical results? Just querying a big pile of mess versus querying a high-level highly-structured, already-validated-to-be-correct database.

It’s a completely different ballgame. And yet, we are being sold this lie that you simply can chuck Claude on top of a pile of documents, and your job is done.

Now, let’s just revisit something. Remember that when you, you just naively chuck your contracts into Claude, what you are doing is you are asking Claude to go through this entire big white box every single time you have a question.

When actually, the only information that matters is a very small proportion of that entire contractual estate. So, you’re going through so much excess material, you’re burning through tokens, it’s taking a long time, and because you’re going through so much information and it’s all raw data, you’ve not calculated any of the implications inside your contracts, you’re ending up paying a huge amount, it’s taking a long time, and your answers are crummy.

And what that means is that every single time you get an answer back, the onus is on you, a person, to do manual work to check that that answer was correct.

You’re massively undermining the very reason you got a system like this in the first place. And here’s why it’s so tempting, it’s because you can set up a system like this really easily and quickly.

And you can defer all of the pain to when you actually need to use it. The only time it’s going to look cheap to you is because no one’s using it because it’s so rubbish.

There has to be a better way, and there is a better way, which is to build a layer in between.

So, we take our documents, we have some magic layer, that organises them, and make sure that whenever we use AI to ask a question, we’re feeding in just that small organised sliver of information that is required to answer the question.

And this is where the other use of AI comes in. We can use AI to take our pile of documents and do the work once, to produce an outstanding, organised database that is a much, much better source of information to feed into AI when we ask a question.

And just to make very clear that this is AI with a different role, I’ve used a diamond instead of a circle.

And we have a little person here because AI is still not reliable enough to build this entire thing by itself, you need some sort of human supervision to make sure that the database that you’ve built is correct.

The good news, though, is that once you’ve done this job, you’ve banked it. You can reuse this again and again and again.

Every single time you have a question, no repeated work. And it’s much, much cheaper. So, in the end, what we see is that we have two places, not one, where AI is used, and two very distinct roles.

One of them is to turn your raw data into really clean, structured data, and then the other one is to query that structured data.

Thank you very much, have a wonderful day.