David Cushman — HFS Research[00:21]
Hello. We’re here at Inflection AI with the COO, Ted Shelton, and we’ve recently published a paper together — a market vision paper in which we’re laying out the reasons for enterprises to start getting serious about AI and to start thinking about scaling AI into the way they work. And that’s been a really big theme for Inflection AI. So, Ted, tell me a little bit about where you are on the journey right now. But also, we’re here in the Valley, and the talk around here — and I think spreading across the world — is all about agentic. How’s that impacting the whole journey into the enterprise?
Ted Shelton — COO, Inflection AI[00:54]
Sure. Well, I think we’ve all been on this journey for several years. Inflection, of course, was founded in January 2022 by an amazing research team out of Google, led by Mustafa Suleyman. And actually the vision that they had right away was what we are now all using the term for: agentic. Basically, recognizing that if you’re going to build an assistant, that assistant has to be able to take actions on your behalf. And so the Pi consumer product, which Inflection launched in March of 2023, always had as an architecture the ability for that assistant to start doing things like order lunch for you, book appointments for you, get a car service. So, of course, Pi was very much a personal assistant, and we pivoted in March of 2024 to be focused on the enterprise. But that same philosophy was, from the beginning, a part of what we set out to do in March of 2024. We called it action quotient back then, because we said it was IQ, the intelligence, the models; EQ, the emotional intelligence; and then AQ, the action quotient. Now, of course, we have to go with the flow and we’re all calling it agentic now, but absolutely, it’s core — it’s core to the future of the way AI is going to deliver value.
David Cushman — HFS Research[02:29]
So it’s an idea whose time has very much come, and it seems to me that it’s an idea that is driving a lot of interest from the enterprise, because they have to think about it beyond point solutions and POCs. This is about actually delivering end-to-end value now.
Ted Shelton — COO, Inflection AI[02:52]
Well, let’s just say, by the way, there’s still a ton of value in the fact that these models can read and write. But yes, actually, when they can take actions on our behalf, that value multiplies, for sure.
David Cushman — HFS Research[02:52]
And so we were talking earlier, and it might be worth sharing with our audience this idea that there’s going to be a kind of multi-speed approach to agentic coming into the enterprise. What do you think are the controls, and where do you see agentic making its biggest impact fastest?
Ted Shelton — COO, Inflection AI[03:21]
Well, I think the first and foremost is consequence. I think that’s one of the speed controls, as you put it. If I am putting a person, or even just data, at risk, then I’m going to be a little more cautious about how I roll it out. So I think the first really fulsome agentic approaches, like our own Inflection Insights product, is a data analysis tool. Because what we’re doing is saying, look, let’s actually allow a person to ask a natural-language question that requires a complex analysis process — something I would normally give to a business analyst. The business analyst would say, gee, how do I break this down? What data do I need to answer it? What analytical tools do I need to use to look at that data? How do I put it all back together? How do I generate a graph or a table or whatever is necessary? And so that is an agentic process: to be able to say, I’m going to understand what the outcome is that the user wants, and be able to put together that whole plan and execute it. But what are the consequences? The consequences are, oh, you didn’t answer the question I wanted you to answer. Now, if I take that to the next step and say you’re actually going to modify an information system that I have — well, now my consequences are a lot higher, and I’m going to be more risk averse. And if I go to the next step and say I’m going to actually modify the operating parameters of a physical piece of equipment, and something could blow up and kill people — OK, wait, let’s slow down here. So I do think that’s one speed control. The other speed control — and this is one of the philosophical beliefs that we have in approaching this whole topic — is that we need to make sure AI is aligned with human outcomes. And so the notion of unleashing completely autonomous AI to go do whatever — I don’t think we should feel comfortable with that. I think we should make sure that there is both observability of what these agents are doing on our behalf, as well as what I’ll call reversibility. I want the human in the loop. I want to be able to say, oh, actually, that’s not the action I wanted to have taken, I want to back that out. And so until we have the architecture for these systems to really make sure the human is in the loop — again, scaled with consequence — we’re going to be cautious.
David Cushman — HFS Research[05:42]
So you mentioned architecture, and that maybe brings us on to our third point, around what you described as cognitive surplus. We’re clearly in a place where we’re having to rethink how work gets done. That’s kind of a layer-one piece of work. And then there’s the business architecture to support that, which I think is probably the area that has least been dug into just yet. And then there’s the technology to deliver on that business architecture. So tell me about how you’re thinking of this idea of cognitive surplus changing the way we work.
Ted Shelton — COO, Inflection AI[06:34]
So, cognitive surplus — by the way, I should give credit to Clay Shirky for coining that phrase 15 years ago. His original framing of cognitive surplus was that all of these labor-saving devices we are inventing, all technology, are freeing up people to spend more time thinking. And of course, the challenge of spending more time thinking is: what are we going to do with those thoughts? A lot of the cognitive surplus that was unleashed in the 20th century was then spent on media consumption or other diversions. But the interesting thing he highlighted was that there were these tools starting to emerge around the 2010 timeframe which coordinated the work that people could do together. And so, AlphaFold — you know, the ability to understand protein folding — somebody created a game, and people could play the game and try all these different ways of folding the proteins. And if you have 100,000 people all trying to fold, all building on each other’s previous work, you can actually get to the right outcome much faster.
David Cushman — HFS Research[07:56]
The whole excitement around crowdsourcing.
Ted Shelton — COO, Inflection AI[07:56]
Yeah, exactly — crowdsourcing, perfect. But now what we have is a different kind of cognitive surplus. So I’ll add to Shirky’s idea and say, well, now I have a machine that can fold. It acts as the crowd — it is the crowd. And, by the way, it doesn’t want to go look at TikTok, so I also don’t have to worry about my cognitive surplus being leaked out into diversions. But I do think companies now have a real challenge on their hands, because the task of the leadership of the company is now going to be: how do I manage the cognitive surplus that I have? Both the human cognitive surplus — because machines are going to take over so many more of those rote, repetitive tasks, which then unleashes the ability for human creativity to be applied in other places — but then also the cognitive surplus of the machines, where I can suddenly say, wow. Well, I’ll take the Insights application I was just talking about: for the cost of roughly two business analysts, I can give you a thousand business analysts.
David Cushman — HFS Research[08:57]
That’s quite scary for a business analyst, right?
Ted Shelton — COO, Inflection AI[08:57]
Well, OK, a business analyst should think about doing something else as a career. But what’s also interesting, though, is this: if the CEO of the company says, I want to understand something, then a bunch of business analysts scurry off and try to work on that problem. If the person who manages a warehouse says, I want to understand something, nobody is going to help him. But now, if I have this cognitive surplus, I can suddenly give business analysts to my warehouse manager. So now there’s thinking and capability for everyone in the organization, and fundamentally work changes, because that warehouse manager is now responsible not just for making sure the same process is followed every day, but for trying to understand how to improve the process. And he has the tools to do so.
David Cushman — HFS Research[10:08]
So, taking ownership of an interim process and becoming the CEO of that little fiefdom, with a whole team of agents to help them do it.
Ted Shelton — COO, Inflection AI[10:08]
Yeah, yeah. People and agents working together are going to accomplish more than agents alone or people alone. That’s our philosophy — it’s really about how you create the right structure for a co-work relationship between people and AI.
David Cushman — HFS Research[10:38]
Well, Ted, thank you so much, and thank you for hosting me here in Palo Alto. It’s been a real pleasure speaking with you.
Ted Shelton — COO, Inflection AI[10:38]
A pleasure speaking to you, David.
