Companies Will Take Legal Responsibility for AI Agents | Ken Wong

Ken Wong (0:00)

Today, when we talk about these autonomous agents, oftentimes, especially in the commercial consumer side, these agents are acting as extensions of you. And so, it is your authority, you're accountable for the actions of these agents.

But as we think about the enterprise world, I think there is gonna be a near future where these agents don't act on your behalf, but as agents themselves. And the organization itself takes on the responsibility of what it does rather than the individual. Actors that the agents are impersonating.

SPEAKER_2 (0:34)

Hello, everyone, and welcome back to MTS. Today, I am joined by Ken Wong, who is the Senior Director of Product Management at Databricks. And this week, Databricks had some big launches. They launched their Always On Genie. This is an AI agent. You hand a goal, and that does the work in the background inside all of the company's own governed data. So Ken, welcome to MTS.

Ken Wong (0:58)

Hey, great to be here.

SPEAKER_2 (0:59)

Great to have you. Let's talk about the big announcement. What did you guys launch?

Ken Wong (1:04)

Well, there were really three big things, actually. So you already talked about one of them, which is the new version of Genie.

Genie, for those who don't know, is our AI that's really specialized to work with your company's data. Obviously, that's critical for every organization out there. But what we've launched this week is a new autonomous mode for Genie. It was really kind of a big step forward. So instead of you having to ask Genie to kind of tackle a task for you, like build me this dashboard and understand how my sales pipeline is looking, you can actually just tell it to say, hey, I would like to make sure I have enough pipeline to hit my next quarter's goals. And then Genie will actually stay on that job on an ongoing basis, monitoring data and then acting on it. So that's actually just one of three announcements that we're announcing this week. The second one is the GA of Genie Ontology.

And that's our horizontal context layer that makes sure Genie actually has the context it needs to answer these questions and perform these tasks that involve data with a high degree of accuracy and efficiency. We launched some new benchmarks working with an external consulting agency called Turing to demonstrate the impact of this approach. And the third is we really stepped up the capabilities in terms of co-work in Genie as well. So now Genie can generate beautiful interactive artifacts like slides with interactive visuals that stay up to date with all of your data and all available just by talking to Genie.

SPEAKER_2 (2:40)

All right. You got the Always On Genie, the Genie Ontology, all of the updates with co-working. So I have to ask, what is happening at a macro level? What are you noticing amongst how enterprises are using AI and agents in the workplace that led to these big products?

Ken Wong (2:57)

Yeah, I think the macro thing is like most enterprises are obviously starting to roll out these tools much more broadly, but they are seeing that the challenge is not the intelligence of these tools that they're rolling out. Obviously, the latest models from all the frontier labs are incredible. But in terms of actually being able to automate work, or take on complex knowledge work, there's a really big gap between what these models are capable of, and what they are able to do in production in the real world today, and that comes from context. If you think about data work inside enterprise, or any type of knowledge work inside enterprise, it's very different from coding, where all the context the agent needs nicely packaged in a repo.

In enterprise knowledge work, your data is all over the place, all the understanding of how to interpret this data is disparate, at least spread out in dashboards, in different systems, and in people's heads. Giving agent that horizontal context across all these things is a tremendous challenge, and I think more and more organizations are recognizing that and looking for a solution. That's really why we really focus a lot of energy on solving that problem and packaging it up in what we call the gene ontology.

SPEAKER_2 (4:20)

Yeah. You say giving agents context amongst all of this horizontal knowledge is a huge challenge. Can you help me understand why? What is the hardest part about this?

Ken Wong (4:30)

Yeah. Ultimately, if you look at the whole problem space of agents translating your questions expressed in natural language to data, it needs this understanding of what those things mean. And if you look at even the research that's out in the world today, like say the famous BIRD benchmark for measuring conversational analytics efficiency. Basically, if you look at the leadership board, it's like the 50 top entries have this concept of oracle knowledge. And what is oracle knowledge? It's like this explicitly spelled out understanding of like, hey, this means this, this means that.