Aaron Levie and OpenAI on the Rise of Enterprise AI Agents

Thank you so much for having me.

I'd say there's a couple things that are happening. In November there was this seismic shift in what agents are capable of, and I think that was the result of two things. One, the models themselves and the underlying intelligence and reasoning getting really good. And then two, the agent harnesses getting significantly better. So, agents suddenly have access to a lot more tools. They know how to use a computer really well. If you think about what type of knowledge worker task can't be substantiated by really good coding, really good use of a computer, I think that there's not a huge amount of tasks that are left that are not going to be touched or changed by agents.

And as a result of that, we're seeing kind of two categories of agents emerge. The first I would just call the sort of like AI super assistant.

No, yes. This is absolutely a technical term. Sometime we not to be confused with the super duper assistants, and those are we see those as the frontier, clearly.

Yeah, that's AGI, exactly. And so, where these agents are really representative of is they're almost like an extension of my identity. A lot of them, the agent runtime runs on my computer, it has access to all of my same tools, it has access to all of these different things and applications that I leverage. These things are magical, right? Because I can suddenly ask an agent, 'Hey, look at my calendar. Where are there opportunities to streamline my day?' I can go in and say, 'Hey, go in and reach out to these different members of my team that are managing a project and consolidate all the updates into a summary for me.' I can go in and say, 'Hey, I need to go and schedule a meeting with this person. Go in and do the coordination back and forth in order to arrive on a time.' I think it's going to be a period of time where in the future we're going to think it's crazy that we used to email back and forth multiple turns to try and find 30-minute slots where we could speak. But, that's one agent category and it's booming and there's a lot of really interesting developments that are going on. Now, there's a lot of security and governance challenges that people are also saying, you know, like how do I think about identity in this world where it's not clear if it's the super assistant acting on my behalf or if it's me directly? And there's a lot of questions and a lot of really interesting decisions that are going to come about.

The other category that I'd say is really more of these sort of workspace agents and you could almost say the metaphor there is the AI co-worker. And it's kind of like, 'Hey, we have this process in legal where there's someone who's hired that goes in and every time we review a new agreement, they have to go in and see if there's any non-standard terms and they go through the process of pulling those all together and then do a review of them.' And there's that whole triage process that happens as a result of it. Those types of things can be very much automated now with the new capabilities of these agents with the new capabilities of these harnesses. Now, it doesn't mean that we're still not going to have people that are human in the loop to go in and judge exceptions, but for a lot of these processes where maybe 75% of it was just evaluating are the terms standard or are they non-standard, that should just be offloaded to the agent. And we're going to start working with those more and more as co-workers, and that means that they might not just be locked in these chat interfaces, but they're headless and it's just another person on Slack that's got a funny emoji and responds a lot faster than some of my other co-workers. But I think that that's where we're starting to see these agents now as they're capable of more and more knowledge work really enter the workforce.