Hello and welcome to People in Profit. I'm Shal Pel, and this week, an in-depth conversation about Cisco. The San Jose, California company has been a mainstay of the computer age, providing its clients with the equipment, services, and technology necessary for our computers, phones, and other devices to communicate safely and effectively. When it was founded in 1984, the revolution of the personal computer had started. And now, 40 years later, a new revolution is underway, that of artificial intelligence. And Cisco's president and chief product officer, G2 Patel, is one of the people making sure it doesn't miss the boat. G2 Patel, thank you for joining us on People in Profit.
Thank you for having me, Charles.
So you've said it before that we're entering what is referred to as a second phase of artificial intelligence, where it's no longer about prompting artificial intelligence to respond to specific questions like ChatGPT does, for instance, but it's about artificial intelligence being able to have agency in running designated tasks. What is referred to as agentic AI exactly? Tell us more about that. How does it work?
Yeah, that's a great point, which is, you know, you described it pretty well, which is we used to have for the past two and a half years, we've been used to these intelligent chatbots that answer questions that we actually posed, and it's largely been used for individual productivity. We're now moving to this world of agentic AI, where agents will be able to conduct tasks and jobs almost fully autonomously to essentially automate workflows rather than just enhance individual productivity. And so it fundamentally changes the kind of value that you can derive from these systems. Probably one of the first few agents that we all might be familiar with is deep research, where the duration of autonomy is about 20 minutes. You actually can ask it to go out and do some research for you. Go get a cup of coffee, come back, and then it'll actually have a report printed out for you, you know, kind of ready and composed for you. And those kind of things will just, as the duration of autonomy goes from, you know, 20 minutes to two hours to two days to two weeks to two months to two quarters, you'll start to see that the value compounding will be pretty material as you move forward.
And so exciting times ahead in that field. And there are constraints, there are hurdles to developing that kind of agent AI. What are those hurdles that need to be cleared?
Yeah. So if you think about the key impediments that you might have in these technologies, we think of them as three key impediments. The first one is what we call the infrastructure constraint. You know, we don't have enough power in the world, nor do we have enough compute capacity, nor enough network bandwidth to satiate the needs of AI. So that's going to be the first problem that we need to solve. The second big problem is what we call a trust deficit, which is if you don't trust these systems, you're not going to use them. And safety and security tend to be pretty paramount to going out and getting this right. And so we need to make sure that safety and security are kind of catalytic to the adoption of AI, which is the first time that this has ever happened. Usually you would either be safe and secure, or you would actually argue for productivity. This is the first time that AI will not actually take off if people don't feel safe and secure with it. So that's the second big area that needs to be solved for. And the third one is the data gap, where, you know, most organizations think about their data as their moat, their data as their differentiator, but how you use your data and how you organize your data to use it and train the models actually has a lot to do with how much kind of power you can harness from AI.
I want to get back to that idea of the trust deficit, which I find really interesting. It's difficult getting people to feel safe in using AI processes in their business operations or even in their daily lives. This is an extremely emotional issue for people. So, how safe is Agentic AI? Actually, it's extremely worrying for some people to think that there are these systems that could operate autonomously, and how do you build that trust afterwards?
So, if you think about how these systems are designed from the get-go, is they're built on top of models. And these models, by definition, are what they call non-deterministic. Now, what does non-deterministic mean? It means that they're unpredictable. If you ask a chatbot a question three times, you'll get a slightly different answer each time. And that's why it's kind of unpredictable. And so what we want to make sure is that we can validate the model, that is, the model behaving the way that you want it to behave, and when it doesn't, can you put the appropriate guardrails so that every single time someone is building an application on top of a model and someone is using these models, they don't feel like they're going to be exposed because a model might do something that they don't want it to do. And so that's an extremely important kind of dynamic that we have to keep in mind. Cisco has built a technology called AI Defense that actually helps solve exactly that problem, where we can not only have full visibility on what data is going into the model, how do you validate the model so it is behaving the way you expect it to behave, and when it doesn't, can we put guardrails so that when you build these applications, they have boundary conditions that they don't kind of go.
