Joel Hron

And so I think those have been tremendously important.

We've also built...

as tools for the agent uh so you know ways for it to check its own work so like if it makes a statement like there's a tool for the agent to use to go validate is this statement supported by law and they can go in and like check like okay yes this is supported by law i feel good about and so these like recursive uh checks if you will for the agents are an important way to help

mitigate some of that risk.

Again, I don't think you'll ever like remove it.

But, you know, we try to do everything we can to help mitigate it.

Yeah.

I mean, you know, I guess I would start with saying, like, I don't believe that humans are immune to these mistakes either.

Like, I'm sure many of these things happened even predating generative AI and this happened.

But by all means, I don't think that obviates the need as a software and technology provider to do, you know, 110 percent of our best effort to mitigate as much as we can.

And, you know, the way that we look at that is really probably three dimensions.

The first is around like the user interface elements that I mentioned to you before.

Like, how do we convey some level of like confidence about an answer to a user?

How do we convey, you know, these visual clues that yes, this case does resolve to a real case because it's got a bright blue deep link, you know, to it.

How do we put these flags in place in the answers to allow users to sort of navigate through the trajectory easier?

How do we give them sort of UI clues of how the agent resolved and, you know, what paths of reasoning that it took to get to this answer?

So that's one element.

The second is things like building actual tools to do this validation in real time for the agent itself.

You know, we have an application internally that we also released with deep research.

In the market, we call it litigation document analyzer, which is a mouthful, but internally we called it the BS detector, which is basically like a way for somebody to upload a brief and like,