Neeraj Verma of NiCE on What Separates AI Pilots From Production

Neeraj Verma of NiCE on why AI pilots stall, building agents as small reusable units, and the scale problem personal agents are about to create.

Full show notes
This one was recorded live at NiCE World 2026 in Orlando, and Neeraj Verma, Head of AI at NiCE, didn't dodge the hard parts. We started where most enterprise AI conversations should start and rarely do: the outcome. If you can't name what the technology is supposed to pay you back for, you're experimenting for experimentation's sake, and that's the pattern I see stalling pilots everywhere.

From there we got into the fast-moving stuff — what an agent actually is, where skills fit, and why the smart move is building small, reusable units of work rather than monolithic agents. Neeraj made a point I keep thinking about: agents aren't humans, they're context engines, and they need small context to execute well. We also dug into guardrails and observability, and the real tension there — you have to have them, but not in a way that doubles your cost or ruins the experience.

If you're a technology leader trying to move from pilot to production, or a CX leader watching personal agents start to change what "scale" even means, this is worth your time. The honest through-line: this wave will move faster than any before it, and it still takes real infrastructure and a continuous-innovation mindset to get right.

What we cover

• Why an outcome has to come before the AI build
• What an AI agent is, and how skills package tools and processes
• Building small, reusable, modular units of work for agents
• Context as compressed enterprise data — and why compression is the hard problem
• Harness engineering, and "agentic whack-a-mole"
• Compounding intelligence and what self-learning really means today
• Personal agents, agent-to-agent communication, and the coming scale problem
• Guardrails, observability, and keeping shadow AI in check

STAY CONNECTED

• Subscribe to the AI with Maribel Lopez audio podcast: https://www.buzzsprout.com/1947446
• Subscribe to my LinkedIn newsletter — AI Decoded with Maribel Lopez: https://www.linkedin.com/newsletters/ai-decoded-with-maribel-lopez-7312533413582827520/
• Lopez Research blog: https://www.lopezresearch.com/research/
• Follow me on LinkedIn: https://www.linkedin.com/in/maribellopez/
• Follow me on X: https://x.com/MaribelLopez

Transcript

Copy

Maribel Lopez  0:00  
Hello and welcome back to the AI with Maribel Lopez podcast. I am, in fact, Maribel Lopez. I am joined here today with Niraj Verma, who is the head of AI at Nice. We are at Nice World in lovely Orlando. We're having a good time talking about AI. We're always talking about AI, and this is no exception. And to be talking to one of the really intriguing leaders in AI, I'm very excited about this. Welcome to the program.

Neeraj Verma  0:25  
Great to be here, and you know, I'd say it's it's hot outside, but it's cold in here.

Maribel Lopez  0:30  
It's freezing. I'm wearing a

Neeraj Verma  0:30  
sweater, and I'm I'm still shivering.

Maribel Lopez  0:32  
Yeah, yeah, yeah, yeah. All right, but you know, we got hot stuff to talk about, so let let's do it. I've been out talking to a lot of enterprise buyers. They're all very excited about AI, right? Everybody's doing AI. You guys talked about that on stage. Now we're at the point where how do we move it out of pilot and into production? So I wanted to talk to you about you know what are you seeing when you talk to organizations that are really operationalizing AI and making it happen within their organization? Is there something that they're doing differently? Is it just that we now have the power of the platform and they hit a go button? Like, how do you think about that?

Neeraj Verma  1:07  
So, you know, I'm going to say in more general terms, even outside of Nice, I think I think AI experimentation is still very heavy in the world because people have FOMO, and I mean, it's true, right? People have FOMO, and I think, regardless of what technology you're implementing, whether it's AI or something else, you have to have an outcome in mind. You cannot have experimentation for experimentation's sake. And by the way, we see this on the platform all the time. Like, well, I need to get into the AI game. What can you do for me?

Maribel Lopez  1:34  
Yeah. So it's not really about I've got a business goal that says I want to try to create this type of experience, how can AI help you? Yeah, just like oh, I want to

Neeraj Verma  1:43  
use ChatGPT. Can you make it look like Chat?

Maribel Lopez  1:45  
My boss said I needed to do some AI. So I mean, sure, we could. Yeah, we could

Neeraj Verma  1:48  
do it. But what are you going to get from this? Yeah, and that's. But really, that's what differentiates success from failure. Is like you got to have the outcome in mind that's going to pay for this thing.

Maribel Lopez  1:56  
Yeah. the The other thing I've been thinking a lot about is so you've got what are you trying to do? Then you've got how do you measure it? You've got what the executive support is around it, and I think those are things I've been talking about for a while. One of the things that I'm a bit intrigued by, and you know we we didn't rehearse any of this ahead of time, so I'm just going to ask him whatever the heck I want. So in the category of whatever the heck I want, there's been a discussion about hey, you start with POCs, and then some people are like, well, you you start with AI skills, but I think skills require you to really know a lot about AI, like what you're trying to create, how to break it down. So I talk to Yum Brands, and Yum Brands is like, well, we just do you know reusable AI skills. Think about it, like we create a pitching agent, and that's like all they do is pitch. It doesn't matter what they pitch or what they're trying to do. Is that a type of strategy that you think can work now, or do you think that people have to start in a different place?

Neeraj Verma  2:52  
I think so. Before we can get started, what is an AI agent, right? Yeah, an AI agent back in the day. When I say back in the day, six months ago, was primarily just system prompts with a bunch of tools and a tool, you know, just a description of a tool. You can do X, Y, Z using A, B, C. And Anthropic introduces concept of skills, which essentially packages up descriptions of tools and processes, right? Using a bunch of tools, right? So a skills is think about one level higher than a tool itself. Now, what people are doing today is they're creating AI agents, and different teams are structuring essentially compostable skills for those AI agents, right? Yeah, and that is the right path. But the problem is that this market is changing so fast that I'm afraid the skill concept will be superseded by something like a subagent pretty soon, but but the whole goal of when you we've got

Maribel Lopez  3:43  
subagents, some people have super agents. There's you know and and different classifications. But is it is it

Neeraj Verma  3:48  
that you do want to create the smallest modular piece of work execution for an AI agent to understand? Why? Because they're not humans. They're context engines. They need small context to execute. So you want to have these small, reusable, describable piece of execution that they need to

Maribel Lopez  4:07  
do. Okay, you mentioned the word context, and this is a fascinating word, right? Because we've had it for a while, and I think within the past six months, you know, you actually mentioned how how quickly agents and skills are changing. One of the things we've seen at all these conferences, as everybody's talking about business context or context now, how is context different from data?

Neeraj Verma  4:27  
I'd argue that context is accumulation and compression of enterprise data that drives AI agents. So, what what is an LLM? Right, an LLM is just a compressed representation of a shit ton of data. I mean, it's true, right? That's what an LLM is, and context is data that that that that LLM can operate on top of. That's all it is, right? But you want to make sure that it's the compressed representation of all the things you want the LLM to know. And by the way, this is why context compression is such a hot topic in today's world, right? Because even. If you have lots and lots of context for your agents, they can still get lost within the context. So, how do you reliably compress that context into something that's easy for an agent to understand?

Maribel Lopez  5:09  
So, another thing I want to talk to you about, since you spent a lot of time thinking about this, and since things are moving so fast, I've been following a lot of the Harrison Chase stuff in harness engineering, and that's something that I think is a term we didn't even really talk about six months ago. Really, truly, when you think about that, what does harness engineering mean to you?

Neeraj Verma  5:29  
Okay, man, you know harnesses themselves. I hate the word, but harnesses are essentially the the thing that runs a model or many models for a given task or purpose. And there's 1000s of harnesses in the world. Clot Code is arguably the best coding harness, right? But there's lots of other ones. Codex is another one. Open Code is another one. Cognigy, as an example, is a harness that that uses many LLMs orchestrated in order to resolve customer experience, right? So harnesses are essentially just execution engines that combine logic and LLMs to perform a task and harnesses matter, right? I mean, you could see that Claude Code's harness is fricking amazing, even though the models underneath it are very similar to what OpenAI is offered in Codex.

Maribel Lopez  6:15  
Yeah, I think this is fascinating how quickly we're creating terms, and it's like maybe they're not exactly the right term, but you sort of get the concept of okay, harness scaffolding. It's like something that kind of goes around it, right? It holds it together, and we we just keep finding missing pieces as we roll out something,

Neeraj Verma  6:32  
right?

Maribel Lopez  6:32  
So you know, like first we're talking about an agent. It's like okay, now we can build an agent, and you're like, great, okay, we have to have agent identity now. We have to have the orchestration of the agents. We have to have governance around that, right? So it's like we keep finding things. We have to have harnesses around all these things, right? Yeah, agentic

Neeraj Verma  6:48  
whack-a-mole. Agent whack-a-mole. I love it. That's a

Maribel Lopez  6:51  
great, great phrase. Agentic whack-a-mole. You and I were talking briefly a few minutes ago about compounding intelligence. Right. What does that mean to you, and and how does it help businesses?

Neeraj Verma  7:02  
Okay, so another history lesson because I love history lessons. Everybody talks about self-improving a learning AI, and we've been talking about it for I don't know two decades. Yeah, it's never been true.

Maribel Lopez  7:11  
Never been true. Never been true. No, I'm serious. I know

Neeraj Verma  7:13  
you-based systems, especially even conversational AI and other things, were never really self-learning. Never really. And I think even in today's world of LLMs, when we talk about self-learning, it's essentially contextual hints and improvements to agents' context. Right? You talk about context.

Maribel Lopez  7:28  
Right.

Neeraj Verma  7:28  
Right. When as you use cloud code, it understands who you are, what you're doing. Right. Starts keeping notes.

Maribel Lopez  7:35  
Yeah.

Neeraj Verma  7:35  
And essentially modifies its system prompt to be more appropriate to your use cases, right? That's all that self-learning. More your AI, right? It's you know in the world of conversational AI, it's things like dynamic personality warping. Well, I know Maribel; she typically likes to have conversations that are short, so I'm going to modify my system prompt to be, "Don't reply with anything over three words,

Maribel Lopez  7:58  
as

Neeraj Verma  7:59  
an example.

Maribel Lopez  8:00  
Okay, that makes sense. I think you know there's so many different ways this could roll out right now, but I do believe that we're trying to get to a point where we talk a lot about the AI working for you, and you know what that means, and and how do you create that? And I think we're still working on the tooling to actually make that easier for you know Maribel to create the AI that works for her. That would be specific. When you think about all the things that are going on with Nice and Cognigy and with AI, you know if you were going to put on on your crystal ball, what are the types of things that you think have to happen over the next year for AI to really be successful in customer experience. Do we have to build something? Is this about deploying what we already have? How are you thinking? It's

Neeraj Verma  8:48  
interesting. So I'm going to zoom out just from the nice lens because I think it's important to look at the greater context. I think what Apple just announced with the super powered Siri, and I'm sure you heard Phil's talk about

Maribel Lopez  9:00  
it. Yeah,

Neeraj Verma  9:00  
that is the sort of spark that's needed in the customer experience world to have personalized agents that do work for you. And I think all of customer experience, including Cognigy, Nice, we're going to be answering to that need very quickly. Right? Imagine that you've got a personal agent on your phone that makes calls for you, that contacts brands to do business. How do you handle that load? How do you handle agent-to-agent communication in the most reliable way? Right. I mean, it's it's a it's a really big change that's coming within the next six months to a year.

Maribel Lopez  9:34  
Well, and I think that's interesting from the perspective of scale. How we define scale has changed, right? Because scale used to be human scale, let's say. But the the world that you were just describing, I could have hundreds and hundreds of agents, right? And those agents, maybe 50 of them, you know, want to call Citi and ask Citi to. Things, right? So it's not just like one Maribel trying to do a compounded whatever. It's a bunch of little Maribel agents off trying to do a whole bunch of things. And I, as an organization, have to figure out how to scale that, how to route that differently, right? When it requires human engagement, when it requires you know the agent I've created talking to your agent kind of thing. So it's it's a different, you know. We used to talk about nice interactions back in the day, right? The conference was called interactions, and now I think interactions are even so different than you know we used to have. Well, it was it was omni-channel, and I could send you a text message, and it's like now we're so far beyond this concept of you know its voice is it is it all the like it's just a different different way of thinking about the problem and if the problem is going to be broken down into pieces or not or how complex that problem but

Neeraj Verma  10:51  
is it like do you even need more humans right so it's like this this concept of unless we figure out this this personal agent thing very quickly you don't have enough humans on the planet to deal with the problems, right?

Maribel Lopez  11:04  
Right.

Neeraj Verma  11:04  
I mean, you have 50 maribels calling in. Who handles those conversations, right? And

Maribel Lopez  11:09  
by the way, we never had enough humans to deal with the problem, right? There was always a queue in the contact center. There was always a huge churn rate in the contact center. So we were constantly in this battle of like, how do we get people up and running fast enough to service our problems. So, I I don't worry as much about the human per se as some some people do. I stress

Neeraj Verma  11:30  
about it all the time because I I know, you know, part of what I do here is I I'm a builder. I'm not you know I just I don't lead the teams. I also build with them because I I think it's really hard to lead in today's world without knowing what the technology is even capable of.

Maribel Lopez  11:44  
Absolutely, and I

Neeraj Verma  11:46  
build these things with my team, and it scares me how easy it is to build some of these things and to kind of overwhelm systems. Right, that's the big thing, and it's it's also the interesting sort of environmental cost. Right, you've got 50 maribels running around. What's the what is the environmental cost, and is it worth the environmental cost? I don't know. It's interesting.

Maribel Lopez  12:07  
Yeah, you you actually bring up a good point because one of the things that we talk about when we talk about the word operationalization, which I apparently can't even say very well at this point in the day, if you want to make this stuff work in your organization. You know, one of the things we've been talking about is how do you coordinate them, right? So there is that discussion that we've had orchestration, which is a very difficult term to to deal with as well. But it does start to get into the question of just because you can build something, should you build it? Right. Are you inherently better off as a result of having built said thing. Is it the right thing? And just as though we've had shadow IT in the past, you know, we've got shadow AI going on now. People are trying to figure out. You know, is there, are we burning a lot of money? Are we burning a lot of tokens for no apparent reason? Are people doing cutesy cat videos with our AI tools, or are they actually doing something meaningful, so we we have a lot of challenge. There's not there's not an implied question there. It's more of a statement about part of it. It circles right back to the first question: how you answered the first question, which is, what are we trying to do? And you know, we used to spend so much time talking about the models, and I think we talked about the models because we got enamored with the concept of tools.

Neeraj Verma  13:25  
Right?

Maribel Lopez  13:26  
You know, what tools do we need? And it's sort of like trying to build a house, not knowing if you're putting in a window or a door or drywall. You know, what do you need? Do you need a nail gun? Do you need a hammer? Do you need a screwdriver? So we spend all this time doing that, but it's like, well, if you don't know what you're building, then you don't know what tools you need. And I feel it's the same when we are talking about everyone's going to have agents, and it's like, okay, then there's a whole who creates the agents and do they have the right permissioning and all the other stuff. So when I when I look at your world and when I look at customer experience, I see both the opportunity and the risk, right? I feel that there's a lot of opportunity, but there's also a lot of ability to get it wrong,

Neeraj Verma  14:07  
right?

Maribel Lopez  14:08  
You know, the agents could go sideways on you. So, when when you think about from a platform perspective, are we looking at sort of inherent observability that's just always checking what's going on to make sure these things are operating the appropriate way, or how do we think? You've got to

Neeraj Verma  14:24  
have the right guard. I mean, look at the reality is you got to have the right guardrails in place. I think guardrails have been around for quite a while, but you also have to be careful in sort of how you structure a guardrail so the experience isn't ruined, right? Do you have guardrails that run all the time and cost double the interaction, or do you have guardrails that add latency? But you have to have the right guardrails and observability layers. I can't say that word either. On top to kind of ensure conversations are going the right way, and it's we talk about this compounding intelligence.

Maribel Lopez  14:51  
Yeah, this is

Neeraj Verma  14:51  
where compounding intelligence is really important. This observability layer that allows your agents to self learn that if if I'm going to get red teamed, here's how I. Handle that red teaming,

Maribel Lopez  15:01  
right?

Neeraj Verma  15:01  
But I think you know, just just kind of going back to your previous point, I think ideas are still valuable. Code has always been cheap. You can always hire people to build code, and I think ideas are still. It's always been the same problem. Yeah, you have to have the right idea at the right time in order to make something good, so that's the limiting factor, not code generation.

Maribel Lopez  15:27  
But I think this gets back to the whole discussion about the human when we when we talk about like human versus AI, which I don't really. I'm I'm hoping it's not that kind of discussion moving forward, but this gets back to if I have the right ideas, I can push the business forward. And so, one of the things I'm really excited about with AI that we have not spent a lot of time digging into, because I think we're still so early in getting this part of it to really work smoothly, is well, could we do something different? You know, how are people going to actually reinvent the customer experience? We've been talking about customer experience reinvention for a very long time now, and it still feels like we are stuck in the same thing. I'm trying to deflect you. I'm trying. You know, it's just like you know, in 2026. Yeah, I think

Neeraj Verma  16:14  
that's the reality. But you know, if you if you ask me what is the ultimate customer experience, it's it's this where I don't have my phone on me. I don't have a problem because my problem's already being solved by something.

Maribel Lopez  16:24  
Right,

Neeraj Verma  16:24  
right.

Maribel Lopez  16:24  
You're not sitting here worrying about. I don't have to worry about a problem. I don't have to

Neeraj Verma  16:27  
make calls. I don't have to make contacts with anybody. My problems are solved. That is customer experience to me, right? And ultimately, I think AI agents can offer that. We're still, in my opinion, a year to year and a half, maybe two from that. It's it feels like forever, but the amount of infrastructure needed, all the way down from the businesses to the applications that make the changes is pretty intense.

Maribel Lopez  16:54  
I love the fact that you put that reality check in it because from where we sit as analysts, you know, we go to all the conferences, right? And from where you sit as vendors, you're messaging really something that's typically a year or two out from customer deployment. Right, you're saying here's a vision, here's what we're building for you. It's getting better all the time. The it used to be we used to announce something once a year. Right now things are moving so fast. Like every couple of months, there's something really great and new and interesting in the field, and I think we're starting to get a little fatigued by that. But what I think is interesting is this is a very powerful technology wave, and it will happen faster than any other technology wave we've happened, we've had happen before. But it's still it's going to take a while, right? We still have to put you know all this stuff we're talking about with harnesses and everything around that, so I think it's important that we acknowledge that it is still a system that is evolving and in flux. But that doesn't mean you shouldn't do anything, right? Right? You still exactly what you said. Said you have to go and use a technology and understand what it can do and try to build things and sort of see where they break, see where they surprise and delight, because sometimes you get surprise and delight as well, and that's that's a great thing. You're like, well, I didn't even know it could do that. So, yeah.

Neeraj Verma  18:09  
And I talk to leaders about this all the time. You know, even when you deploy these types of technologies, you still have to be so open, right? Models get deprecated so fast in today's world. Yeah, that I spent a year creating this agent, and now it's useless. Right, I have to create it again. So it's like you always have to be proactive and pragmatic, and understand. I'm telling you, you have to. The people that are the most successful in this are the people that actually understand the technology, and everybody else is struggling to keep up. And I see this, by the way. I have decision fatigue all the time when I read all this news. I have so much fatigue from just trying to understand when these models are out, when they're going to be deprecated, and what what they're valuable for. It's it's a it's an interesting world for sure.

Maribel Lopez  18:53  
You know, I'm really happy you brought this up because I think it gets to the new world order of what it takes to be successful, and it requires continuous innovation, and you have to be comfortable with that. Like as, and I think that's a big organizational change. If you think of the way you know certain technology stacks used to work in the past, you you had a roadmap and you you know executed on it over you know two three year time frame, and then you sort of felt you were done for a while. Right, we've done the rollout. It's like, well, I don't know if you're going to be done for a very long time, and you have to be okay with that. You have to put in processes for that. You have to have a more experimentation mindset, not just for experimentation's sake, but to see if we can push it forward. You always are going to be trying to push things forward now, and I think that is different than what we've had in the past, and requires a different style of leadership and a different style of openness. And you know, some people call it growth mindset. Maybe that's the term. I think we get a new term in 2026. But I do think that that's something that you know you've been talking about a lot. Actually, you're on stage. It came up from the customers about how they realize that they're you know not going to be done. They're just going to keep doing more things and finding more things to do. So it's the

Neeraj Verma  20:08  
outcome mindset. I think I'm trying to tell people it's the outcome mindset because hopefully the outcomes that you want always improve.

Speaker 1  20:16  
Yes, and

Neeraj Verma  20:16  
never resolve. Right.

Maribel Lopez  20:18  
Yeah. All right. Final piece of advice for the audience:

Neeraj Verma  20:23  
Be open, experiment, try, use tokens. I'm telling you, it's it's amazing. I you know, I told this in the last podcast. We have 4000 employees on Claude Code today.

Maribel Lopez  20:37  
Yeah,

Neeraj Verma  20:37  
I'm in the top 10.

Maribel Lopez  20:38  
Nice,

Neeraj Verma  20:39  
and not just because I want to be in the top 10. Because I, but I think it's so interesting. I think it's so interesting to build and experiment and solve problems, and to lead the team by example.

Maribel Lopez  20:49  
Yeah, I love that leading the team by example, outcome-based mindsets. Perfect, Niraj. Thank you for having the time, spending the time with me today, and sharing all your insight here at Nice World, and I hope we'll get to have a cup of coffee and talk more about this soon.

Neeraj Verma  21:05  
Yeah, be great, be fun.

Maribel Lopez  21:06  
Cheers. Thank you. Thank you.

Transcribed by https://otter.ai