Agentic AI in the Enterprise: What it Means for You

Michael Krigsman: Are AI agents an enterprise savior, workforce apocalypse, or just another tech bubble waiting to burst? Today on CXOTalk episode 876, we cut through the noise with Phil Fersht, CEO of HFS Research and one of the most respected industry analysts in the world.

When we talk about AI agents and agentic AI, what are we referring to? What do we mean by that?

Phil Fersht: It's really the ability to replicate human behavior in software. It's as simple as that. Whether it's mimicking our voices or supporting us in doing our day-to-day work, it's really like the augmentation of humanity in software. And we talk about the blending of humans and technology; this is really where it's at. And it's something that we dreamt about for a long time, but it's only really starting to come into reality, but at an alarming pace. I don't know, we see a lot of fun things flying around on X and LinkedIn and all these types of things; it's just incredible how much development there's been in voice and video in just the last six to nine months. We're going through a complete revolution and agentic is right at the front of that from a technology perspective.

Evolution from RPA to Agentic AI: A Game Changer

Michael Krigsman: Why is agentic AI so important and at this particular time?

Phil Fersht: I'd like to rewind back to the early 2010s when you might've heard of a technology called RPA, robotic process automation, which got very big and very hyped within the technology world. We actually coined the phrase alongside a company called Blue Prism when we launched it in 2012 when we did the first analyst papers on it. And at the time, we were talking about RPA replicating human behavior in software, which would allow us to scale more effectively, threatened elements like offshore outsourcing because companies could technically consider having less offshore resources when you can automate a lot of this stuff. But the problem with RPA was the technology didn't scale well; it was very brittle. But the concept was there, but that was really all about following instructions easily; it was about eliminating manual effort, waste, which was wasted on repetitive tasks.

Then everyone remembers the influx of gen AI nearly... It's gonna be its third year, with ChatGPT really came public nearly three years ago. And that really changed the game in terms of it became a productivity amplifier that accelerates creative and analytical work that really bottlenecks humans. It's the ability to create content. And this is, like, one of the first times we've had nontechnical people have that ability to start to create content, create data, augment their work, create code even. There's a lot of discussions going on around how much code can be eliminated now because of gen AI. And that was all about creating based on prompts.

Now we're into the gen AI phase, which is about understanding goals and figuring out how to achieve them. Agentic AI is a collaborative actor that removes the need for constant human oversight of complex processes. It's self-directing in many respects; it coordinates multiple tasks; it transforms entire workforces; it creates new organizational paradigms. But it's not about the fact that it sounds great. What's exciting about agentic is it really does work, and it's working at an alarming pace that is making, in reality, many people uncomfortable.

Some people are loving it and they're embracing it, and they're realizing, "Wow, I can do my job so much better," and I'm an analyst. I can tell you how agentic and gen AI are helping me do my job. But this is the most, I think, impactful wave in this AI continuum that takes us to the next phase, which we're terming artificial general intelligence, which is much more self-directed intelligence that overcomes human cognitive limitations across all domains. And eventually, artificial superintelligence, which is about computers outperforming humans. We're not there yet obviously, but I watched Terminator 1 with my son the other day; I hadn't seen that in about 30 years, and that brought me back. They actually predicted in 2029 was when computers would become self-aware, that... And if anyone's got nothing better to do this weekend, watch the rerun of Terminator 1. It's uncanny how they got the timings right on this thing.

Practical Applications and Distinction Between AI Agents and LLMs

Michael Krigsman: Phil, I wanna latch onto a particular point that you made. You said, "It really works."

Phil Fersht: Right.

Michael Krigsman: Can you elaborate on that? Because that's kind of the magic point, right? This is not a lab-based theory. It's something that actually works. Tell us about that.

Phil Fersht: Most companies right now have created some standalone single agents. That's an agent that can handle maybe one specific task or function, that could be like an email writer or even a meeting scheduler, things like that. Even like Copilot – you can use Copilot right now to summarize your emails and remind you to do things, or you can use Fireflies, which is a really popular tool, or LinkedIn, not so much LinkedIn as Zoom AI and use it to summarize meetings. Those are single agents, believe it or not, and they're already working. People are already pretty excited about, "Oh, did you turn your Fireflies on we get a good summary of this meeting?"

Where this starts to get really exciting is when we start to build functional multi-agents, which is multiple agents who work together within a single business function. That could be a sales team of agents handling prospecting or qualification and follow-ups, that sort of thing. And eventually, we get to something we're calling horizontal multi-agents, which is where you get different agents collaborating across various business functions and even other supply-chain partners. That could be sales agents working with marketing and customer-service agents. You're actually building out capabilities in business functions beyond one single function.

It works because you just gotta try it. Like, I'd love... There's a demo, it's called Super Film, I think it was, where you could actually put an avatar of me on our research website and ask me questions, and I would literally, in my voice, dig into our research and communicate them back to you using voice. You've just got to see it to see how effective this is. Now, is it perfect? No. Is it as accurate as talking to a human being? Not yet. But in many respects, we're creating agents that are becoming very, very supportive in our jobs. I mean, I'll tell you, for example, I wrote a piece on tariffs the other day, and I put together, "Here's like the big 10, 11 things about tariffs." Some people may have even read it. And for a bit of fun, I pumped it through ChatGPT Pro, and out comes... I said, "Can you replicate this using Phil Fersht's voice just for a bit of fun?" And it came out sounding like the sort of thing that I would have written. And then I asked it to tone it down a bit, that sort of thing.

And then I produced another piece where I said, "Can you produce me a chart that shows life expectancy in the US versus other countries and health issues," and it starts pulling information all over the place. And then you start to train the model. It's like starting to become your own personal agent, and it's getting to know what I need. And then it's like, "Can you produce this in the HFS font and colors?" And you're programming in the colors to use and everything. You're really building out something that can literally become your go-to at work. There's so many different uses. And I don't even know if we're gonna call these agents in another six to 12 months, but this is just how we're partnering with technology now, where we don't have to go to people to get things done all the time now. We can get much done ourselves, and then we, as human beings, become the creators of that content. Like, we're saying, "I've got a big business meeting to go to tomorrow." You're gonna get much content that you need. You set your agenda to the technology, and it gathers you what you need, and it allows you to then curate that to make you effective as a human being.

Michael Krigsman: How is this different from having an interactive chat with ChatGPT? What's unique about agents? But before you answer, I just wanna remind everybody — our regular listeners know this — that you can ask your questions. Right now, if you're watching on Twitter, pop your question using the hashtag into Twitter, using CXOTalk, #CXOTalk. If you're watching on LinkedIn, pop your question into the chat. This is your opportunity to ask one of the top analysts in the world pretty much whatever you want. Take advantage of it. And we have some questions that are coming in now. But first, again, you're describing an interactive process. You go to a meeting, you give it notes, you then say, "Modify this or modify that." Sounds like ChatGPT. How are agents different from LLMs and our usage as we know and love it today?

Phil Fersht: ChatGPT is useful because you can use it for a specific prompt: "I need some information on this or that. Get some information quickly, produce this, do that, do this, do that." An actual agent is the virtual coworker who is completing end-to-end processes for you. It self-directs and coordinates multiple tasks. Once you've... Say I'm using the example of I could use an agent to help me with my research. I would develop, train this as a virtual coworker to be like my research assistant. It would start to, over time, learn what I do, what I need, how I do it. You can start to interact with this like a virtual coworker, like a research assistant, for example. And you can leverage this to create whole new organizational paradigms.

The Rise of AI Agents in the Workforce

Phil Fersht: I'm not joking. In 12 months' time, you can say, "Hey, Phil, I need to have a meeting with you next week to talk about XYZ." I can literally have you talk to my agent, right, who will look at my diary to coordinate what I need and time and maybe ask particular questions. We are training virtual coworkers to do the jobs that we either used to do ourselves or someone else did, and we can start to get into real examples of this.

But the challenge is going to your staff in a company and asking them to almost recreate their jobs into software, which is very different from saying, "Train up another human being and transfer tasks from yourself to another human." We're now expecting people, including ourselves, to transfer human work tasks into software. That technically frees us up to do other things or, let's be honest, could make us redundant, right? We're not needed anymore. We can actually leverage agents to do the jobs of humans.

And we're now seeing enterprises who are really trying to have an AI-first mindset. They're now insisting before you hire any new staff, you need to show that this work can't be done by AI. We've reached that point quite quickly. This wasn't like maybe 20 years ago; people used to say, "Hey, if you hire new staff, can we see if that work can be done offshore in the Philippines or India or something?" Now, C-suite directives are, "Can we not do this with AI?" The whole point of agents is really this ability for companies to grow and scale in a way that you don't need to keep adding more and more people. You do a lot more with the people you have, and I think that's the positive way to think about this.

It's, "I wanna run a marketing campaign. Can I develop a planning agent who can coordinate and break down the campaign requests into specific tasks? Can I create a research agent that gathers market intelligence? Can I create a creative agent that develops creative assets and messaging? Can I develop a strategy agent that optimizes my campaign and engagement across marketing channels?" Right? You can go on and on about... Almost every new staff member you need, you can create an agent for. Like it could be, "Hey, I need somebody to manage social media-"

Michael Krigsman: Right.

Phil Fersht: "-and I need to do automated LinkedIn updates," that sort of thing. Or it could be even a campaign coordination agent to synthesize inputs from all agents into a cohesive campaign. It's creating people into software. We call this thing, you've probably heard of services as software, but this is what's happening in the services industry right now: companies are starting to think about, "How can I replicate the services I'm receiving from an IBM or an Infosys or one of these companies and receive this using agentic software? Versus why do I keep having to add more people all the time?" That's the real nub of agentic, and I think why it's causing excitement and friction at the same time.

Michael Krigsman: Go to cxotalk.com, subscribe to our newsletter, and join us. Join our community, and join our live shows.

Risks and Challenges of AI Adoption

Michael Krigsman: We have a very interesting question coming from Twitter, from Anthony Scriffignano, who is the former chief data scientist of Dun & Bradstreet and has been a guest on CXOTalk a number of times. And he's asking about the unintended risks or unintended harms that can emerge. Can you talk about that aspect of it? I think lots of people are concerned about the impact of AI agents on the workforce. You spoke about the positive aspects, but what about the unintended risks?

Phil Fersht: First thing is, you've got the more general risks of AI. When you read something sent to you now, a lot of people are thinking, "Was this written by AI or a human being?" All right? That's a big problem, and I see that as an opportunity for a research firm like us 'cause people need real more than ever. And do you trust information that's all produced using agents and agentic software? Is it truly trustable? Is it reliable, where the source is coming from, and that sort of thing? And I think a bigger risk right now is can you trust information from people?

The other issue is obviously with interactions and hallucinations and these types of things — all the types of teething problems you'll have with, honestly, any type of technology. I remember when we were getting into more sophisticated accounting applications 20 years ago, and people were worried then about software malfunctioning and producing incorrect calculations and stuff like that. A lot of it is trusting the software, trusting the security of that software as well, and understanding how to navigate your way around this climate, 'cause it's only gonna get more confusing and more worrying through times. We're getting more sophisticated spamming, phishing stuff. We're getting them every day on texts and emails, and they're quite convincing ones sometimes. It's like, I get my own staff coming to me saying, "Hey, Phil, did you send me this text about getting Amazon vouchers?" Things like that. We can talk about this for a very long time — all the different risks, all the different worries, all the different concerns. And I think, I don't think we have a full answer for that just yet.

Michael Krigsman: We have a question from Arsalan Khan who says, "Agentic AI requires the correct data at the right time with the right human and systems integration. Eventually, these agents become autonomous. What happens to humans then?" He's asking about this boundary between human work and autonomous agent work.

Phil Fersht: This isn't just about enterprises trying to cut costs and replace people with bots. This is about us as human beings. We're all threatened by this, and we all have opportunities with this. And if you're in a job where you can be effectively replicated and replaced, you kinda know that, and you need to figure out, "How do I continue to add value in an enterprise?" I think the value comes from collaboration. It comes from people skills. It comes from empathy. And if you can become a great person everybody likes to work with, and you become very thoughtful about what you do, and you start to collaborate beyond your existing area, you become very valuable to your company.

And I can go through many examples of this. I have a guy running my IT systems who actually was a procurement guy just three or four years ago. But he broadened his knowledge into understanding how to manage HubSpot and accounting software. He manages our stack of social and Grammarly and all this sort of stuff. And as part of his job, he started to get to know all his colleagues in different departments in the company, like analysts, finance, HR, and all this sort of stuff. And then you start to develop real value to deliver across your organization. And I think no matter what role you're in — if you're in a sales role, you're in a delivery role, you're in a tech role — you need to become broader and more aligned to your business to add value there. Because if you become just a replicable, solo, task-driven professional, you do run a risk.

You wouldn't believe some of the conversations I have with CIOs right now, who are under immense pressure to wipe out costs because of code. One major organization I spoke to produces half a billion lines of code a year to keep that organization functioning, and they've been tasked with eradicating 90% of the effort 'cause you don't need to have armies of legacy coders anymore. A lot of this code can be rewritten using gen AI and other types of AI software now. We're just all facing the challenge of, how relevant are we?

Michael Krigsman: Be specific on how these agents help you in your job. Tell us the tools you're using, and then we're gonna go back and get some more questions.

Leveraging AI for Personal and Organizational Productivity

Michael Krigsman: Questions are coming in.

Phil Fersht: If you want to develop real value within your own organization, you have to run boot camps with your own colleagues to present to each other how you're using these tools to be more effective at your job. We've even hired an AI expert who's a full-time employee within our company. She's probably listening to this. But who's actually working across our operations people, our analysts. She's working with Amazon and a company called Lyzera, for example, to agentify how we deliver our research to our clients. While yes, I can go on about the personal productivity tools I use, we're using agentic to transform our whole business 'cause we're in the information business, and we have set up a fairly complex system.

We're using an agentic solution called Lyzera, which is a startup, but it's in a pretty mature phase. They're very popular, and they're powered by AWS to produce, at scale, the ability for... we have like 150,000 subscribers to go in and create their own research support agents to help them leverage, get the most out of HFS. That's how we're using it from a corporate standpoint.

From a personal standpoint, right now I use ChatGPT Pro, I paid the extra money. I'm not sure I need the $200 a month package, but I'm loving it right now 'cause it gives me a lot of query time. The computing power is a little challenging. Sometimes it takes a bit of time to produce everything I need, but I'm finding that effective. I'm using Deep Research from Perplexity, which is pretty good as well, and I've also been experimenting with other tools like Claude, which is the Anthropic tool, and I've also looked at some other tools that can be fairly effective, like Gemini — I'm still not completely convinced by, but other people love it. A lot of this is people finding technologies that they think are better than others, and they like the way they're interacting with these tools. But the new suite from ChatGPT Pro is excellent. You've got the image creation. You've got the operations piece. You've got the deep research piece. What I'm seeing right now — this thing is pretty good, and we're gonna get to a stage fairly quickly where we're gonna be whittled down to maybe three or four powerhouses in this space who are gonna be dominating the progression here.

Michael Krigsman: I use many different LLMs. I'm always experimenting to see which one is better.

Michael Krigsman: Here is a question from Wes Andrews, who says, "You jokingly referenced Terminator earlier, but given the struggles that AI and other sectors are having with establishing frameworks, guardrails, standards such as NIST and GDPR, what do you suggest?" Yeah. And I'll just mention also to folks that last week, we had two members of the House of Lords from the UK discussing these issues. If you care about these issues, listen to our last show, and you can get the transcript on our site. But Phil, what about this framework and guardrail set of issues?

Phil Fersht: We looked deeply into this because we cover global services a lot within HFS, and every different region has slightly different attitudes towards AI. Obviously you mentioned GDPR - huge in the UK and Europe. India is a little bit more of a free-for-all right now with how they're accepting AI-based solutions. And the US — well, this could be the second coming with the tech bros driving a lot of policy here. I think we're still waiting to see how a lot of this shapes up. The EU has typically been the most closed from a framework perspective and demanding in terms of compliance. And anyone running a business knows how challenging running GDPR practices has been in recent years to get to the other side.

But I do think that as this continues to evolve, the need for common frameworks is going to become more and more paramount, and the need for cooperation is going to continue to proliferate. I really do. And look what's going on politically across the world right now. In many ways, this is gonna actually bring, I think, a lot of regions closer, the governments and regions closer together, which may actually drive better cooperation with AI. For example, I was hearing today about a strong movement to create the China-less supply chain, right? How can countries start to group together to manufacture goods outside of China to avoid these potential tariffs, right? And in that case, you need to sort of build a supply-chain competency that centers and responds, that manages inventory, that brings cooperation together and these types of things. I think the need to build supply-chain standards, trading standards around AI - I think this is just going to... it's only just beginning, and we're gonna see a lot more of this emerge in the next couple of years.

Enterprise AI Adoption and Challenges

Michael Krigsman: What about enterprise adoption? Where are we today? AI agents are still relatively new. There's lots of promise, but in terms of actual usage and enterprise adoption-

Phil Fersht: I can share the latest and greatest that we've been working with. We spoke to over 1,000 major enterprises looking at the adoption of gen AI and agentic, and 45% of them are either worried about job loss or they're resistant to change, and adoption is fairly diminished. At the other end of the spectrum, only 15% of AI leaders are generally positive about AI adoption, and they have fairly integrated views of where they're going; they have a strong culture of support, and they're embracing this. And then in the middle, you've got about 40% of enterprises where they're still in that sort of pilot purgatory phase. Their culture is becoming more adaptive; they're recognizing the benefits of AI, but they're not there yet. In terms of actual adoption, you've only got about 15%, maybe a little more, who are getting to the point where they have a real clear vision and understanding of where they're going.

One thing that is crystal clear is we're seeing immense pressure coming from board-level people, and also C-suite leaders in organizations, to drive AI adoption a lot faster. There's real pressure coming right from the top to really embrace and become more effective as AI-first cultures. But the reality is, we're still at early days. We've been talking about this for a long time, but the reality is ChatGPT 3.5 only came in not even two and a half, three years ago. We're playing catchup. But what's happening is the technology is staring it in our face. It is really here. We've got big firms really trying to get on top of it. You've got the big software companies like Salesforce, in particular, with their Agent Force rollout, and ServiceNow, some of these businesses really trying to muscle in on agentic 'cause they see that as an opportunity to take market share away from the services firms. And at the other flip side, you've got services companies like Accenture really trying to become more dominant in the services-as-software realm as well. Adoption is low is the real answer to this, but the pressure is there, and it's on like never before.

Michael Krigsman: I just want to invite everybody listening to join the CXOTalk community. Go to cxotalk.com and subscribe to our newsletter we can notify you of upcoming conversations.

AI's Impact on Workforce and Economic Risks

Michael Krigsman: We do this every week, and you guys who are listening — you are the cream of the crop. Join, subscribe to our newsletter, and join these conversations, and add your points of view and your questions. We have an important question from Arsalan Khan, who says, "How do you convince the C-suite that agentic AI is not just a fancy chatbot before they move on to the next shiny object?" What are the challenges and the opportunities associated with this?

Phil Fersht: We're past that point where the C-suite can keep denying that this is just another fancy chatbot, I think. If you're leading a CX function in particular, you—if you're not familiar with how easily replicable call-center agents are with intelligent bots right now, you shouldn't be in a job anymore, to be quite harsh about it. I can tell you, just an example of an organization I spoke to with 50,000 onshore staff responding to healthcare inquiry calls, and the leader basically said, "Look, the bottom line is, there's the same six questions being asked over and over and over again. We've already run the analysis. We can literally replace half these people with intelligent bots," and they call them empathy bots, very, very quickly. "We're not going to do it straightaway, but we know the possibility is there."

And I think this is a typical case across a lot of companies: they're very aware of what they can do, but they're still yet to have that burning-platform trigger to go do it. My concern is if we plunge into a deep recession, you're gonna see some organizations literally come out and say, "We're just gonna start relying a lot more on AI, and we're gonna let people go." My vain hope is we don't fall into recession we can have a more positive view of people and technology. But there is that risk that a negative economy can drive a lot more weaponized AI where companies would say, "Look. Let's just replace these people. We don't need them anymore." I don't think companies I speak to are not aware of this. It's more how advanced they are with embracing this. Are they prepared to do anything?

And my concern is I do talk to a lot of enterprises. We have a lot of summits and roundtables on top of our research where people wanna talk, but when it comes down to what are you actually doing, they're not doing a lot. And I think what I just said about the 15% — that's not a big number. Fifteen percent are kind of on the path. The rest are either still figuring it out or they're not on the path. And this is just going to become more pronounced as we go through the next few months of macroeconomic turbulence.

Michael Krigsman: You just made a wild comment, which is... and I don't wanna put words in your mouth, but it seems like you just said that the technology is becoming good in a sufficient number of use cases that an economic downturn can push many companies to replace many workers because those use cases and the effectiveness are broadly dispersed, even today, or if not today, soon enough.

Phil Fersht: The technology is available. It's there. I think companies are aware of it. I do believe as well most enterprises don't tread lightly on the fact that, hey, let's go replace 5,000 people with 1,000 or 500 augmented consultants who can manage a team of bots. But one of the things that has been looked at in industry is, why do you need 500 people in India, for example, running a bunch of coding or app support things, when you could potentially replace them with a team of maybe 25 people who are local and onshore who are supported and augmented by agentic technology? It's this ability to reduce the scale of people numbers that you have and then augment higher-value people with agentic to support them.

And we put out some research recently around the impact of tariffs, for example, that could have a real impact on what we call... it could drive the whole services-and-software adoption curve, right? 'Cause suddenly, it's like if it becomes really difficult to manage a disparate global workforce manufacturing goods all over the world, you need to bring stuff back home. Suddenly, "Hey, I can actually do what I need in the US with a smaller number of staff." They might be more expensive, but I don't need as many, and they're supported by this technology. We are at a point where companies are starting to make much more radical assumptions on what they can do.

You may have seen a recent announcement from the bank Citibank, who have decided to reduce their 144 service-provider relationships down to 50, and they've actually increased the numbers of staff that they have onshore in the United States and some other regions who are directly within the company, because they... what they want to do is they want to spend less on the legacy and more on the new. I'm not trying to say companies are just gonna fire everybody and replace them with bots. But I think a lot of smart businesses are thinking, "How do we stop spending billions of dollars on maintaining legacy applications and legacy systems when we really want to reinvest that money in modernized thinking, modernized agentic technology?" That sort of thing.

Breaking Free from Legacy Systems and Embracing AI

Phil Fersht: I don't think companies are thinking right now about, "How do we just get rid of people?" They're actually thinking about, "How do we break from the past?" I did a great podcast with Jason Averbook, who's one of the leading minds in HR technology. He's at Mercer these days, and he talks to, like, the CHROs across all the big Global 50 companies. And he actually came out and said these companies have much data, they don't know what to do with it. They can't join it up. They can't make decisions on it. It's got to the point where he's got clients who are literally thinking of just— just get— "Let's just trash this old system and rebuild with the new." And I think this is where some of these conversations are happening with agentic, which is, how do we start to really build out the new and make a break from the legacy that's been holding us back for long?

Michael Krigsman: Anthony Scriffignano makes a comment on Twitter directly addressing this point that you were just discussing. He says, "It's equally likely that the C-suite is being taken to task for not adopting more to drive down cost." He says, "The KPIs need to be more than just cost savings. What new problems are being addressed that were unaddressed before being enabled by this technology?" And it sounds like you're saying the same thing, that cost savings is a part of it, but there's also a whole set of new opportunities.

Phil Fersht: I would agree that the same fundamental issues have remained for a very long time in terms of changing— we call it paying off your debts: your technical debt, your people debt, your process debt, your data debt within companies, and this inertia of companies refusing to change. And there's many managers and leaders within enterprise who, let's be honest, have got away with not having to do much different for the last 20, 30 years. I mean, we still have companies operating with processes that were designed before the Second World War, some even the Industrial Revolution.

What is different this time? I think what's different is the technology is much more pronounced. It's much more ready. It's much more scalable. And there's a final exhaustion, where you talk to CIOs off record; they'll all tell you one thing. They are fed up spending 10% a year on their services firms and then 10% increases every year on their software-license hikes. SaaS is becoming a legacy paradigm, and services— people just don't wanna keep paying more and more and more. You can't keep going up this exponential cost curve. Eventually, the chickens come home to roost. And I think C-suite executives are really being held to task now: Can you deliver an AI-first organization where the culture has to change within the company?

And I think that's the problem we've got with a lot of these businesses: they haven't got the right cultures to shift forward and really embrace. And while I would agree, I don't think the fundamental issues have changed all that much, what is changing is the onus on AI that's coming right from the top. 'Cause when RPA came in in 2012, the reason why one of the reasons it failed was the CIO— it would get dumped on the CIO's docket, and it would eventually get dumbed down two or three layers into what we call the frozen middle within the organization. And that's when technology solutions go to die. That's not happening much with agentic.

I think what is different is that 15% of high performers and, I think, the following 15% behind are organizations where the leadership have realized they can no longer keep paying lip service to not fixing their underlying problems with data, technology, and legacy. And all RPA was really doing was— it was like a Band-Aid tech that stitched together old systems to get them functioning more effectively. It was very useful, but could you use RPA to replace thousands of people, unless it was a very high-throughput process, very repeatable, very predictable? Of course you couldn't. No, this was like a patchwork technology.

If you want to say, "Hey, we have a thousand people answering calls in the Philippines for our consumer products that we're selling," right? That's people en masse at scale, where you need technology that can actually have some empathy with clients, that can replicate CX behavior, that can actually do the job. And I think that's the big difference right now: agentic is much, much closer to doing the job of human beings than RPA was, which it really wasn't. It was a patchwork, back-office, break-fix technology that was great if you wanted to keep your old CICS mainframe working with a COBOL system, working with an SAP system, for example. But now, it's much more — you can see where this is all shifting. And I think there's a real exhaustion with companies having to keep maintaining really creaking old systems in a world where competition is much more cutthroat, and you've gotta be really slick and on the ball if you're gonna be effective in this economy.

Michael Krigsman: Phil, I get the sense that what you're really also saying is that the difference between RPA and agentic AI is that 15% of early adopters of agentic AI have demonstrated that in fact it really does work. It really can bring these kinds of benefits and savings that you've been describing.

Phil Fersht: Yeah, you can actually create a virtual coworker to complete end-to-end processes. It's proven. It works. We've all seen the demos. We've worked with companies who are piloting it. We have done it on ourselves, and a lot of enterprises, more advanced ones in particular, are at least working with single agents and some even to multi-agent models. They're on the path, and it's a different type of technology that removes the need for constant human oversight of complex processes. It's a transformational tool rather than a task-automation tool, which RPA was. That was about tasks. This is about human oversight, support, and real process capability, and the fact that you can build these coworkers.

The Human-Tech Interaction and Democratization of Technology

Phil Fersht: You can engage with these things. You can talk to them, right? I don't even wanna get into some disturbing things about teenage boys building relationships with AI girlfriends and things like that. I don't know if you've been reading about some of these things, but it's— this stuff is real. People are building relationships with their software. You can ask the question: if you ring up customer service today, do you care that you're dealing with a computer or dealing with a human being? When you all go checking in to your airline, do you really want to talk to the gate agent? No, of course you don't. You just wanna use your app and get on the plane. Do what I mean? We're getting to this whole next layer of technology becoming part of our daily lives much more than ever before, to the point where we're actually engaging with technology in a much more humanistic, real way.

Michael Krigsman: It is extraordinary, the level of research support that, for example, these tools can provide. I had a networking issue of my own here in our studio, and doing a little bit of research, I was able to figure out a fairly complex question having to do with routing rather than needing to call an IT person and bring a consultant in. It is amazing.

Challenges and Opportunities with LLMs in AI Automation

Michael Krigsman: But we have... we're almost out of time, and we have a number of questions that are left. I'm gonna ask you, Phil, to answer these questions pretty quickly, pretty concisely. First one is from Prem Kumar Apparanji, and he says, "When LLMs powering the AI agents aren't reliable or predictable, how do we rely on them to automate unpredictable scenarios that need to work?"

Phil Fersht: You have to train the model to work is my answer. If there's something wrong with the LLMs, then you need to really have a look at the underlying technology that you're using and find the right LLMs that can deliver the scenarios you want. I think there's a lot more technology-based conversations we've gotta have to get this really enterprise-grade ready. What I would say is, I know from a lot of friends in the industry that, like, OpenAI, for example, is very, very obsessed with becoming enterprise ready. Like, the leadership within that company are spending all their time with the C-suite within the Fortune 500. They're trying to figure it out. I would say it's a great question, and there are a lot of faults in the system right now, and a lot of this is honing the models and training the models until you get them working. I mean, as I said, we're doing our own model. We're putting our whole business into an agentic solution, and it's taken us three years to get to a point. We still haven't gone live with the new system yet, but you've gotta learn yourself. You've gotta learn your business. You've gotta learn these models. You've gotta try it and try it and try it until what works and what doesn't.

Michael Krigsman: And I remember when, shortly after ChatGPT came out, I remember that your company, HFS Research, was one of the first analyst firms that I was aware of that was making that attempt to put your research online into an LLM. You truly are an early adopter at this.

Concerns and Impacts of AI Adoption in Enterprises

Michael Krigsman: We have a question very quickly now 'cause we're just gonna run out of time, from Elizabeth Shaw, who says, "CEOs and boards are driving the use of AI, agentic and beyond. There are serious implications for worker and social-society impacts. What concerns do CEOs and other senior business leaders have—" with these concerns? And very quickly, please. I know it's a complicated question.

Phil Fersht: Data privacy and cyber are the number-one problems and biggest concerns by a country mile, to be honest with you. And after that it's other areas around transformation and replacing process and compliance. But cyber is, by far and away, I think, the biggest headache as companies look at shifting to these models — maintaining a secure infrastructure.

Michael Krigsman: Arsalan Khan comes back and says, "Who gains the most value from agentic AI: small companies or large companies?"

Phil Fersht: I would say at the moment, small companies. It allows— I hate using my own example, but it allows midsized businesses to really punch above their weight because you can scale fast, you can act nimbly, and you often don't have as much legacy within the business to change. You don't have as many people resisting change. And I think large companies can also benefit, but I just find with a lot of large businesses it's harder for them to shed their legacy. Look at the technical debt they have, the lock-in they've got with legacy software providers—that sort of thing. I think it's harder for large companies to change 'cause there's a huge amount of training and cultural change and shifting that needs to happen, and I think SMEs are a better place to pivot. I see a lot of people I know wanting to go and work for smaller businesses 'cause they are more nimble, and you've gotta be nimble in this market.

Advice for Business Leaders in the Agentic AI Era

Michael Krigsman: What advice do you have for enterprise-technology and business leaders when it comes to how they should be relating to this agentic AI world today? And very quickly, please.

Phil Fersht: Get on top of it, learn it, understand it, experiment with it, do boot camps with it. You've got to educate yourself. The days of BSing around technology are over, and you've got to be much more proficient at knowing what is possible and engaging and building relationships with the whole emerging AI ecosystem around you.

Michael Krigsman: And with that, a huge thank you to Phil Fersht. He's the CEO of HFS Research. Phil, thank you much for being here. I'm just grateful to you.

Phil Fersht: Yeah. A pleasure, a pleasure, Michael. The hour went quickly. Enjoyed it very much and I look forward to more interactions.

Michael Krigsman: And a huge thank you to everybody who was watching today. You guys are an awesome audience. You're intelligent, smart. Go to cxotalk.com, subscribe to our newsletter, and join us. Join our community and join our live shows. We have one next week and the week after that, check it out. Thanks much, everybody, and hope you have a great day, and we'll see you next time.