Next in Tech | Ep. 250: The Agentic Enterprise

View Full Transcript

Eric Hanselman: Welcome to Next in Tech, an S&P Global podcast where the world of emerging tech lives. I'm your host, Eric Hanselman, Chief Analyst for Industry Research at S&P Global. And today, we'll be discussing the Agentic Enterprise. It also happens to be our 250th episode. So, we're tackling a little larger subject for this go. And joining me to talk about this are analysts of Melissa Incera, Sheryl Kingstone, and Alex Johnston. Welcome back to the podcast to all of you.

Alex Johnston: Thanks very much.

Melissa Incera: Thanks, Eric.

Sheryl Kingstone: Great to be here. Yes. Great topic.

Eric Hanselman: Well, timely, if you look at everything that has been going on and especially as we start looking a little further out in the world at large. We've made this transition from all of the hub of AI. We're now well into the whatever the heck agentic happens to be. And I wanted to look at some of the things that we've been exploring really in some of those further out towards the horizon ideas in terms of what organization should be thinking about, where we should be headed as we start thinking about what that move towards agents and agentic integration and the infrastructure that's needed to support it and really what enterprises are considering in all of this. There's a whole set of different research perspectives that each of you bring to the table. I guess I wanted to start off maybe talking a little bit about the thing that probably most of our listeners have heard more of, which is probably the customer experience piece. This is something where if we think about how agents are being experienced by people, we're already starting to see a lot of that. And Sheryl, that's something that you've been looking at in terms of what's -- what people are putting to work and what some of those expectations are?

Sheryl Kingstone: Yes, Eric, you got to look at it in twofold. What are the consumers wanting to do? And then what are businesses doing to meet those needs within their own customer experience technology stack. And what I mean by that is let's look at the enterprise side first, right? So, the customer experience technology is really things around, for example, customer service and support, right? So how are we using the next-generation platforms to improve using and improving the satisfaction of the AI chatbots, which is what we're calling more of these autonomous assistants or AI agents. We have a lot of research on the difference between them, and that's in the big picture report. You can look at where they're going to optimize marketing, and a lot of the marketing reach out that they're trying to do. There's the whole Agentic commerce angle. So that's more like the inside out. But when we look at the outside in, consumers have been adapting to these kind of technologies for a while, whether it be using more of the chatbots, whether it be using search tools and product discovery, social tools. But what really changed was the generative AI capabilities and the intelligent reasoning. And so now they're really trying to use these technologies. And what's most important is to understand that we're actually tracking the use of these by consumers and what they're mostly interested in, whether it's going to be their traditional use cases and just transforming it into more of an AI agent for them to go out and do research. or more sophisticated shopping experiences. And when you dive down into that shopping experiences, they're still using and preferring some of that human in the loop, right? So, understanding where they're trying to go. But what I'm leading to is a headline that we've been trying to get out is it's still been in this past year, $22 billion of incremental revenue by AI-driven improvements in product discovery and navigation. And so that's why I'm saying it's more of an evolution than a revolution, but we could definitely improve the overall customer experience for both employees and the consumers very strongly down the road.

Eric Hanselman: Well, if you look at the billions that are in play, I'm curious what that shift is. And we start thinking about, of course, AI has loomed large and all this, we've had radical transformation driven by that. But you're talking about evolutionary changes with agents. I'm curious what that general take is in terms of what that -- what the progression is going to be as we start looking forward because it's more than chatbots, right?

Sheryl Kingstone: Yes. Well, it's the evolution of the chatbots, right? So, the evolution, what I mean by that is the underlying technology, and Alex can go into a lot more of that. But it's the shift from being more manually driven roles-based chatbots that have very poor satisfaction rates towards more LLM-driven reasoning-driven interactions for the consumers. So by improving that, and we have already seen between last year and this year, around a 10- to 15-point improvement in the experience where consumers say it's a chatbot, I'm using air quotes like we have our cameras on, -- but where consumers say they're using a chatbot, the satisfaction has improved. And what's improved is behind the scenes iteration of data that's driving it, some of the reasoning engines that's driving it. And so that's why we see this incremental use of these technologies. And then you also have other outside factors like OpenAI partnering with some of the shopping assistants and bringing that more towards the front end of it. And you also have changes in Agentic commerce ecosystems with things like PayPal and Stripe and Visa that's making it a lot more accessible.

Melissa Incera: Eric, I have a view on this from a technological perspective, and you're talking to us at a really great time because I know Alex and I are both armed with very fresh data on agents and AI use cases. So, drawing on some of that, I can share where agents technologically are moving in the coming year. And right now, for all the hype, what we're understanding is that agents are really still thought of as extensions of LLM applications. And at this moment in time, as Sheryl was leading to, they're really being prized for their cognitive abilities. Specifically, memory has emerged as one of the most valued capabilities of an agent. Others are -- would be self-improvement, explainability. But what people want them to be doing right now is essentially to help them think and they want that thinking with context, right? What is less priced at the moment, but a huge focus moving into 2026 is can they execute? Can they automate tasks? Can they coordinate multi-agent flows? Does it have tool calling capabilities to draw on other sources? -- this is coming really quickly, but there are things that have to develop first. We need to get to reliability and trust, which are things we can elaborate on. But that's the direction we're heading. And then beyond that, another element that we hear a lot of talk about is hyper-personalization and persistence and omnichannel interactions. So, these agents that will follow in Sheryl's world, follow a customer across e-mails, applications, web, call centers, carrying through that state and preferences rather than kind of starting fresh each time. And maybe this is a point I can toss to Alex because I know he's done some work on this, but memory and context engineering are one of the most interesting areas right now from an agent development perspective, thinking about designing what resources and information do agents have access to? How do you make sure the right things persist and get handed from one agent to another. So, lots of really interesting things happening as we move into 2026 in terms of the technological sophistication of these systems as they move beyond this thinking of this kind of like LLM orientation into more of a systems and orchestration orientation.

Eric Hanselman: And it seems like your point about memory is really the key to getting to that desired level of hyper personalization, that long continuous relationship. It seems like the regular LLM interactions, chat-based interactions have always been one where they were reasonably isolated maybe for an individual session, there was persistence. But if you want to get to that next level of capability, to Sheryl's point about what they're really looking to be able to integrate for that long-lasting consumer experiences extends to enterprise, you have to have that memory piece and ensure that, to your point, you're actually remembering the right things, remembering them well and setting context for that longer-term relationship.

Sheryl Kingstone: Absolutely. And part of that really comes down to making sure you have strong data management capabilities, right? So, we've been doing the memory and the context management and managing that conversational state in a lot of the customer service and support interactions that were not real time, right? Think about the way we were using chat, but it was human assistant. It was human in the loop driven by a lot of the one-off applications, right? So now if you're really bringing it forward, like Melissa had said, with these agents, we're asking them to do a lot more, right? So, we're asking them to understand how to orchestrate making sure they have the policies and the guardrails.

Alex Johnston: The challenge is quite interesting because obviously, there was a narrative. I guess, part of the excitement around the initial wave of interest in generative AI was based around the idea we've got this model trained on this huge preexisting corpus of data. So, all the things that held us back from patent recognition, which is AI, which is all the sort of data challenges we face there, we find not have to worry about that anymore. And it turns out, now we've got all our data back, we see that these issues of data have risen all over again. When you ask organizations, what are the biggest limitations around your ability to employ generative AI, all the data-related criteria has risen up year-over-year. When you ask organizations, what are the biggest limitations, things like accuracy and quality of responses, which again, they're tying back to data strategy. So, it's quite interesting, the hurdles that held us back in the first wave of AI and now potentially holding us back in the second.

Eric Hanselman: So, we're getting some challenges. I guess we're doing that normal technology cycle of, we got -- we realized what the initial problem was. We thought we solved it. We moved a little further down the road and then realized, in fact, that we really needed to double down on some of those capabilities, especially as I think we've talked about in previous episodes of the podcast, and we certainly mentioned in research, this is something that organizations maturity and understanding what both their data estate looks like, what do they have, what do they have access to? And how do they ensure the quality of that data? How do they actually build effective data pipelines, which is much more challenging, and it's that journey up that maturity ladder that at each new run, they're starting to realize that, in fact, they need more specific capabilities. Now we're looking at building context in ways that I think most organizations hadn't really considered. We built a lot in LLM interactions in terms of prompt engineering. We're now having to really start looking at the data aspect to build sufficient context, getting towards context engineering to be able to actually make that next level of sophistication really deliver.

Alex Johnston: Just to return back to your maturity point, and I guess to draw to Sheryl's evolution versus revolution hypothesis, I find it quite interesting. There's a lot of talk around autonomy, obviously. When you look at the percentage of projects that organizations are trying to sort of point AI to in terms of being fully autonomous with AI executing actions, no human intervention, it's actually a very small minority of projects. It's just under 1/4, so 22% of projects. The priority of AI projects is still in a position where humans are in the driving seat. AI might be recommending actions, but under sort of predefined conditions or rules, it might be to execute some, but the vast majority is human in the loop still. And it's quite interesting when you look at the various controls that organizations have in place around the quality and security of outputs, how much of that is human orientated. So, a lot of that's human in the loop, a lot of that's human-provided feedback should end up in this interesting position where despite a lot of excitement about autonomy and obviously, a proud legacy in areas like robotic process automation things, organizations do have a high degree of autonomy in certain areas. But on the whole, a few organizations feel confident to actually expose their workflows to fully autonomous actions at the moment, which again goes back to that evolution versus revolution hypothesis.

Eric Hanselman: We're moving slowly. We get right back to the challenge we've always had in automation, even pre-generative AI, which is how do you get to the point at which you actually trust full autonomy. And that's still a challenge.

Melissa Incera: There were some interesting takeaways from -- I think this was our last iteration of the use cases survey last year. But interestingly -- so generally, the trend is that trust in AI is decreasing year-over-year. It's been, I think, a 5-year trend. We're at an all-time low, which is very healthy, right? But interestingly, a lot of our data from last year, I don't know if we've done the analysis yet on the newest round, but it shows that the most successful organizations with AI, the ones that are operationalizing it the best and maximizing ROI out of their initiatives, one thing that they share is broad skepticism and mistrust. -- in these systems, thereby -- which thereby prompts them to take seriously the need for security, for governance, for reliability. They're far more likely to have the guardrails in place and be investing in AI governance systems. And to Alex's point, our agent survey showed that similar to his takeaways from his survey, autonomy is something that organizations are healthily skeptical about, but the degree of comfort goes up pretty dramatically, the more control they have over their agentic systems and the more governance they have in place.

Alex Johnston: I find it quite interesting when you'd assume that organizations that are more successful with AI in terms of higher ROI or more likely to be deploying at scale in production, you'd assume that they'd be operating at higher levels of autonomy. But it's interesting when you look at the projects those organizations tend to prioritize, they're actually often those that humans are still in the driving seat for. I think that's really an illustration that some organizations have tried to automate very quickly have found that particularly challenging, at least with the current iteration of technology. And therefore, organizations often going back to, as you say, Melissa, a lot of sort of hum the loop, strictly guard railed, a lot of controls, a lot of rules as these systems still aren't really trusted to operate at a certain level of autonomy.

Eric Hanselman: So, is mistrust a marker of maturity?

Melissa Incera: I think so. I think it is -- I think in order to build trust, you have to start with the foundation of mistrust, right?

Sheryl Kingstone: Really good point, Melissa. Trust actually, when we look at a lot of these applications in my world, trust is one of those foundations, right? So, we have to make sure that we have the ability to trust the data and the data foundation is absolutely critical and the ability to trust the automation and the process. And so that's why I've been talking about you need to get your house in order first before you jump in and understand how you're going to get these fabulous outcomes. And we recently wrote something called an AI agent CX readiness index, right? And a lot of this comes into play and trying to measure where our businesses and how can they start taking advantage of this from scoring around an early fragmented digital experience with some basic AI foundations, merging into some more progress, but inconsistent execution, more advanced, but that's really where the trust and the strong foundations come in. And until you get there, you're not really going to take the advantages of this growing agent potential. And to be agent ready and the ability to deploy and take scalability aspect to gains from that and scale your AI across key -- scale your AI agents across these key journeys, we haven't seen anyone truly ready to do it the way the vision is playing out.

Eric Hanselman: And it sounds like a big part of this is really effectively scoping the use cases, not getting into this expectation that agents are going to somehow magically solve all of your particular problems. or that you're going to get towards full autonomy in some relatively short cycle. I'm curious, some of the recent data also talks about use cases, and Alex, you brought up some of those. Is that something where we're starting to see an evolution in the kinds of use cases that organizations are laying out and attempting to tackle?

Alex Johnston: So, use cases is certainly an interesting topic, Erica. We're speaking about the fact that actually organizations see most success by going through use cases that have human in the loop. And actually, as Melissa says, when we look at the use cases, we're applying agents to, they're actually quite similar to the first set of generative AI applications. A lot of it, as Melissa said, sort of extensions of what they were trying to achieve already with large language models. What I would say is quite interesting, though, if we built up this position where we're saying, still very human in the loop, very few processes are seeing end-to-end automation. There's still a lot of guardrails being implemented often humans as part of that. Maybe there's some sort of post processing going on, but a lot of it's very human orientated. Then you'd expect that those users will be held responsible for the outcomes from those systems, if they're the ones using them, if they're ones validating them, if they're the ones enacting actions from the suggestions that the models delivered to them. But actually, that's not the case. And we find 41% of organizations hold the internal technical team responsible for any unintended outcomes from autonomous systems and a further 21% holds the AI vendor, so solution providers that could be Microsoft or OpenAI. And actually, a tiny percentage of organizations, around 4% hold the individual user responsible. And this is interesting in 2 reasons. I think the first is we sort of understand that from a technical standpoint, the user rarely has any of the data in front of them that allows them to make an informed decision about whether to use an agent or not. they don't tend to have confidence calls in front of them. And perhaps someone has some information to hand around how accurate these systems are. They're probably a technical team rather than someone who's actually sitting and using these systems. So, it's understandable that holding those individuals responsible would be an issue. But the second is when we look at organizations that do partially hold the user responsible. So those that are actually looking at hybrid governance models, they tend to have significantly higher levels of AI ROI. They seem to have more success in terms of delivering capabilities in production. So potentially a model where you treat accountability is something that's distributed not just from technical teams, but something that an individual user needs to be using these tools responsibly, that seems to be a high-value approach to take, even if it's a minority one. It's only around 17% of organizations that have that hybrid governance model.

Eric Hanselman: Well, it seems like that gets back to the idea that this really is a technology that has multiple facets the end user has some responsibility as with any tool to use it well. And yet that core responsibility is that it gets delivered effectively. Interesting to see that from an organizational perspective, there is greater success in those that are using that kind of hybrid model. I wonder whether or not this is something where we'll start to get to maybe regulatory requirements for responsibility and certainly, that arbiter of everything that we get towards further down the road, some case law that actually decide some of this. It seems though that those aspects are still fairly far out. The regulatory efforts are certainly, I think, what we identify as pretty rudimentary today and maybe banning some specific use cases, but not really talking about responsibility, liability, those things still seem to be a ways off.

Melissa Incera: Yes. Regulation is a bit of a grab bag right now. And you can see it in the U.S., right? There's so much tension between wanting to move forward quickly and wanting to move forward safely and protect consumers and protect business users. So yes, regulation, legislation, this is all going to be a big theme moving into 2026. We started to see tendrils of it, specifically related to, unfortunately, tragedies that have happened in the consumer realm with using chatbots in ways that were harmful. But yes, I think this is going to be a thread through the next year.

Eric Hanselman: I think we can probably presume that there's going to be more legislative focus. The challenge, I think, in many cases is that what we've seen is a focus on what you were just pointing out, which is the outcomes of the harmful outcomes for models -- but yet, that doesn't seem to get actually for sort of general-purpose models, great. But as we get to agentic kinds of environments, now we're getting to that next step where I think some of that direct interaction with the model is going to be mediated by the agent. And we now start to get into some interesting new territory in which maybe the direct interaction with the model gets hidden from us. Hopefully, it gets managed well by the agent, but that there may not be as much visibility into where -- what's actually going on in the direct interaction of the model. Now it seems like on the one hand, that's maybe an upside, which, Alex, to your point, if there is some responsibility being placed on the user, maybe the agent actually starts to manage some of that interaction and potentially increase success rates. But yet, it means that the actual implementation now has to spend or has to have a much greater focus on what's actually happening, what the agent is doing outside of the view of the user of that particular application. That starts to get in some interesting territory. It also starts to speak to some of the complexity of the ecosystems that start to support this because we come to the situation today where we've got users that are interacting with models through chat interfaces. And you've got that human-in-e-loop piece of when you get an unreasonable answer or a rational answer, conabulation, what have you, that the human can say, wow, that's really off base. But yet we get to agents, now we've got to have something else that's actually managing that. And I think we've talked about this blending of various models, some that are actually producing the output, some that are checking output. It seems like that's a piece that now starts to become much more important in multi-agent environments and then the real agentic infrastructure we're talking about. It seems like now that means we have to have an ecosystem that's got a lot more models doing a lot of different things, not only just the quality checking, the error checking, there's a lot that has to happen to support that. And that seems to me that now we have a requirement for an overall ecosystem that's got a lot more models and a lot more integration that has to happen to really make this successful. Is that something that when we start seeing the necessary ecosystem to do this, where do we head with that?

Alex Johnston: So, one of the things I think is worth talking about when we talk about all these sorts of models are available and potentially why people might be going towards those models might be interesting to talk about something -- and again, we've got all these various trends of our data. I mentioned data challenges increasing, organizations thinking more about data strategy. That being seen as a key enabler. You run a whole bunch of regression lines to our data and data turns out data strategy really important. One of the other thing that comes up is how important cost has become, and that's a rising concern. And if you -- I'm sure, as everyone does in their day-to-day, the average user reaches out for the highest capability model for most of their challenges. Reasoning mode might be enacted, even there's a need to have multiple iterations of thought generating all sorts of tokens. Organizations tend towards the largest models for almost every challenge they face. And you end up in an interesting scenario where you can imagine agents, I guess, has that abstraction layer also handing off problems to cheaper models potentially, right, or models that are focused on a specific problem. So you might...You might end up in a position where FinOps capabilities are getting integrated into these agents as part of that, you might have attempt to move people to slightly cheaper models because organizations getting very concerned about the cost currently of having all the users across their business and applying the largest models possible to every business case or every use case that they can find.

Sheryl Kingstone: That's a really good point, Alex. And I was listening to what Eric was saying earlier around this whole model governance angle. And it's really what I would call the AI agent control plane that we're looking at what these applications are doing and putting in place for enterprise buyers when they're trying to use these models and AI agents within their enterprise. And so, I'm listening to Alex, and I'm looking at my AI agent control plane around orchestration, policy guardrails, reasoning, right, human in the loop monitoring, cost. He's raising a really, really valid point. I'm going to go change my graphic now because I do think that's the future is to understand how we're going to balance cost with execution and ROI.

Eric Hanselman: Well, what happened with cloud? Isn't this sort of what we keep doing, talking about cycles and technology. We spent a lot of time in cloud and advancing infrastructure and being able to go orchestrate and deploy infrastructure much more rapidly. And what happened? -- suddenly, everybody realized, "Oh, shoot, we needed cost. We have to be able to manage this. And now we're doing a lot with FinOps and being able to go manage your infrastructure costs. seems like we're heading right into FinOps for Agentic.

Alex Johnston: I think one of the reasons you'll likely see it is one of the things we look at is where people are sourcing the AI capabilities from. And what I find fascinating is just how diverse those like sourcing strategy have been the organizations aware that's appearing within the software they already own. They're buying specialist software. They're buying companies they used to partner with but now have better AI capabilities than ones that was a core partner, for example. So, AI is just permeating every facet of the business that's coming from everywhere. So, I think we're going to be in a circumstance soon where organizations have taken this very decentralized staff coming up with ideas, staff identifying where they want to source capabilities from that sprawl. -- at some point, there's going to have to be some form of management. As organizations get their arms around this, I think cost is going to be one of the criteria that organizations use to understand what they're going to prioritize over other initiatives, what they might start scaling down, what they might start prohibiting. So, I do think those sorts of financial metrics might become more important as organizations start to think about, I guess, the sheer scale of this AI landscape within their businesses.

Melissa Incera: Along those same lines and related to the conversation about models, open source is becoming a big thread here in the AI landscape. We saw it growing broadly with Gen AI. And I have to imagine it's only amplifying with agents because of, as Alex was saying, cost concerns, increasing concerns over sovereignty as well. And they're so much more cost effective. It's a great thing to offer optionality apart from those kind of closed source frontier models.

Eric Hanselman: Well, and you've got all those benefits that come along with open source and other environments in that you've got portability, you've got visibility in terms of actually understanding and especially when we're trying to build trust, which conveys a whole bunch of extra benefits on that as well. Well, this is fascinating conversation. There is so much more we could be discussing. We'll have to continue this conversation as we roll through the year. But it has been great to take it as far as we have. We got to get to better trust. We got to get to better optimization, maybe checking a little maturity there, but hopefully, we're getting beyond some of the ponies, rainbows and I guess, the -- I keep resisting saying irrational exuberance. But a lot of the -- hopefully, a lot of that enthusiasm is giving way to a slightly more tempered understanding of where we're headed. But I appreciate all of your insights and thank you all for being back on the podcast. Thanks, Eric.

Sheryl Kingstone: Thanks, Eric. I think it's been a great time.

Eric Hanselman: A good way to kick off the year and our 250th episode.

Sheryl Kingstone: Congratulations.

Eric Hanselman: Well, thank you very much. I also want to thank our production team, including Sophie Carr, Feranmi Adeoshun, and Kyra Smith on the marketing and events teams. And I hope you will join us for our next episode, where we're going to be talking about the earnings season. Melissa Otto is going to be back talking about how we closed out the year from the Visible Alpha team. So, we'll see where that all goes. But we hope you'll join us then because there is always something Next in Tech!