We’ve run three real-world matters through Relativity aiR. Here’s what we know so far. 

When we heard about aiR, back in 2023, we couldn’t wait to try it. We got on the Relativity Limited General Availability (LGA) program and so far we’re very impressed. This is one technology that is living up to the hype.

DANNY THANKACHAN
COO, Intrepid Managed Discovery

Is aiR like all those other times people talked about AI but it was just the same repackaged TAR tools? No. This is different.

Yes it is.

No. It’s going to amplify human intelligence, not replace it. Relativity aiR runs on prompts, and a prompt is just telling the AI what to look for and what concepts to consider when it’s doing the looking. We’re not saying the AI isn’t smart. It is. It’s very smart. What we’re saying is that the AI intelligence functions more like a force multiplier. It amplifies whatever you put in. If you tell it to look for the wrong thing, it will. If you don’t give it the correct legal guidance for the matter, it won’t be very useful. On the other hand, if you bring specific legal insight into the matter, aiR will let you scale that insight and apply it instantly to thousands of documents. It takes whatever direction you give it and it scales that direction. This is why, as an input, it actually makes human intelligence and intuition more essential, not less.

We’ve run three real-world matters through Relativity aiR. So far, we’re very impressed. It’s still early to say what the actual aggregate efficiency benchmark is going to be, but based on the matters we’ve used it for, our guess is that it’s going to be about 60-70% more efficient than the top end of TAR efficiency. If you think about it, this would make sense. It’s an evolution. As TAR was to linear, Gen AI will be to TAR.

With TAR, a common opinion was that you needed at least 2,500 documents to make it worthwhile. Because of training, If you had fewer documents than this, it wasn’t worth the cost. This isn’t the case with aiR. Unlike TAR, aiR doesn’t need to be trained, you don’t need to pre-feed it. It works the same on 2,500 documents as it does on 250,000. Just give it a prompt, tell it what to look for, and it goes.

It isn’t going to take over everything. It sorts through data to find the thing you tell it to find. At the moment, this is what it is. It’s still a funnel, it’s just a much, much better version. We’re still going to have to validate results in the same way that we do now, and we’re still going to have all the other steps in the eDiscovery process. But as far as finding and sorting information, we’re looking at a significant improvement.

One way to think about what aiR can do, fundamentally, is that it can scale expertise, consistently. This is particularly relevant when you think about an industry like document review. Currently, size matters. There are certain projects that require a scale of operation that only a handful of providers are able to supply. We call it the brute force advantage. However, as aiR becomes more widespread, it’s going to begin to eat into this advantage and ultimately consume it. Since aiR can scale expertise, it can allow one attorney with specific knowledge to scale that knowledge almost infinitely. It will allow one person to do what it currently takes 10, 20, or 100 people to do. Unlike a review team composed of individuals with differing interprations of the review guidelines, aiR will consistently apply the instructions on all documents.  In the case of review, we’re talking about reviewing and sorting large quantities of documents. Relativity aiR is, essentially, a better funnel. We think it’s very possible that a single attorney with well written prompts will be able to sort 100k documents in 24 hours. Currently those same 100k documents might take a team of 20 one or two weeks. What this means is that soon, smaller teams with more concentrated expertise will be able to compete at the scale that was previously off limits. It’s going to level the playing field. Game on.