Zeta Global on Why AI Agents Need Context Before Autonomy

[00:00:00] - [Speaker 0]
Scale your business with agentic AI with limited risk. With Denodo's AI data layer, your agents are provided with real time company data and guardrails for company protection. Create the business you always dreamed of with Denodo, and you can do that by simply visiting denodo.com to learn more. But now back to my guest. Can a company consume enormous amounts of AI and remain no better equipped to actually improve their business?

[00:00:36] - [Speaker 0]
Well, in today's episode, I'm gonna welcome back to the show Chris Monberg, CTO and head of product at Zeta Global. He's officially friend of the show. Spoken to him a few times now over the years. And today, though, Chris will compare the current appetite for disconnected AI tools with eating without becoming nourished. Incredibly useful warning for leaders under pressure by just buying something because it's carrying the ubiquitous AI label.

[00:01:06] - [Speaker 0]
So today, we're gonna discuss why data, identity, context, and business goals all need to operate as a connected system, explore how context graphs can create a readable history of recommendations and outcomes, and why AI agents should never grade their own homework. And Chris will also explain why familiar productivity measures can create false confidence, especially when employees produce additional code or content without improving cost, quality, or customer response, or indeed the overall health of the organization. And that difference, that only becomes visible when the entire business will perform better. And I suspect that much of what we will talk about today will resonate with many of you listening. And with that scene set, let me reintroduce you to Chris right now.

[00:02:01] - [Speaker 0]
So a massive warm welcome back to the show, Chris. It's almost a year past since we last spoke. Can you tell everyone listening a little about who you are and what you do?

[00:02:10] - [Speaker 1]
Well, first, Neil, thanks for having me back. Really love being part of your community. So I'm the CTO and head of product at Zeta Global. We're a technology company. We focus mostly in the marketing space, and what makes us very different is that we've got a proprietary data and identity asset that pairs uniquely with our artificial intelligence stack.

[00:02:31] - [Speaker 1]
We try not to disintermediate those things because AI just relies so heavily on data to begin with. And, you know, we work with brands throughout the the globe on acquire, grow, and retain use cases.

[00:02:44] - [Speaker 0]
And when we last spoke, as I said, around a year ago, we discussed building AI that is helping brands grow through repeatable programs, but most importantly, without losing that human touch there that makes them recognizable. So what have you learned since our last conversation and how has your thinking possibly changed?

[00:03:02] - [Speaker 1]
Oh, I don't know. Not much. You know, has much happened in AI in the last year?

[00:03:06] - [Speaker 0]
Agents, agents, agents, maybe. Yeah.

[00:03:09] - [Speaker 1]
Yeah. So much changes and so much stays the same. So I've been working in AI since really focused on AI since about twenty twelve, twenty thirteen. And the technology changes every month. What doesn't change is there's still a pain that users are looking through and trying to decide how much they trust systems.

[00:03:31] - [Speaker 1]
And I can't say that anybody trusts anything completely yet Mhmm. Although we are getting more and more comfortable. I was reflecting just yesterday, was looking at a World Lab company, a friend of mine is working at. It's very early stage. And I I was thinking that it was just yesterday that we'd look at AI generated image of a person and their hands never pass mustard.

[00:03:53] - [Speaker 1]
We're like, oh, their hands, they're awful. This is terrible technology, it's never gonna work. And now you can watch an entire movie trailer and for the most part people have five fingers, you know, for the most part it looks right. So the technology continues to crank and we still do need to figure out how to help organizations and individuals transform the way that they perceive the technology, the way they sew it into their daily lives, both professionally and personally. And maybe even more importantly, and we'll probably talk about this more today, is how they start to create more of a system out of it rather than a game of whack a mole with with different tools.

[00:04:28] - [Speaker 0]
Yeah. And I'm glad you've mentioned that because I think it was Gartner estimated that the worldwide AI spending will reach 2,590,000,000,000 this year. So I've got to ask from what you're seeing, how much of that investment is buying disconnected tools rather than solving the data and connectivity problems that prevent those tools from producing meaningful business results?

[00:04:49] - [Speaker 1]
Well, it's a good question. Out of 2,590,000, I'm I'm hoping a majority of that comes to Zeta Global. And if it did, the focus would be on connected tools. You know, the biggest challenge I hear in talking to brands every day is that people have this voracious appetite for AI. They're getting a lot of pressure, so the the board is saying, hey.

[00:05:12] - [Speaker 1]
We need to steer into AI. And most companies have a AI initiative around marketing. It's a pretty early adopter, which is great, but then there's a bit of mania. Somebody's been appointed the head of AI transformation, and that role didn't exist before. So it's not like this person has a ten year of experience transforming organizations around AI.

[00:05:31] - [Speaker 1]
They're trying to figure it out at the same time. They're trying to consume everything. And, you know, I've got kids. Maybe I'll I'll use one simple phrase to summarize this. Being full is not the same as being nourished.

[00:05:47] - [Speaker 1]
Mhmm. And most organizations are not nourished right now. They're eating a lot. It's like you know, my my son loves to have a like a little candy bar, a z bar after school and tastes really good, but it's it's not gonna last very long. Yeah.

[00:06:03] - [Speaker 0]
And

[00:06:03] - [Speaker 1]
the end of that is a bit of a crash. My hope is that, at least the brand Zeta works with, is that we're mitigating that crash and creating a really nice transition into a place where their tools can work together and create more sustainable and more healthy operating environment for the business.

[00:06:23] - [Speaker 0]
And before you join me today, was doing a little research on you guys and what you've been up to since we last spoke. And one of the things that stood out is you've rebuilt your data layer on the Palantir Foundry and now describe yourselves as an AI infrastructure company. So tell me a little bit more about that rearchitecture. What did it involve in practice, and were there any assumptions he had to abandon along the way?

[00:06:45] - [Speaker 1]
Yeah. Well, first, we describe ourselves an AI infrastructure company, but we've always been one. Yeah. This goes back to my company. Back then, we were writing the algorithms by hand, very literally using every resource I could find to figure out how to get those to fit into bigger models to fit into smaller memory.

[00:07:05] - [Speaker 1]
I was calling my friends down at Stanford, having them come up and spend days, mathematicians that could help compress similarity models. But we were stubborn about the way that data and AI worked together. And I continue to be most adamant about that data access layer and the way it's available for AI. That infrastructure has grown and grown year over year on under the team's watchful eye. So AI infrastructure has always been the game.

[00:07:34] - [Speaker 1]
We have found market fit with marketing organizations to date, and that will continue to evolve and be pervasive throughout organizations. Starting a couple years ago, teams outside of marketing started using us as well for a lot of different use cases. Infrastructure, think, is the missing piece for most businesses right now, and it's very complicated with not a lot of great advisers out there. And I won't purport to be the best adviser. I'm heavily opinionated as somebody who's built systems and watched these fail and be successful over the last decade and a half?

[00:08:06] - [Speaker 1]
Palantir was the second part of the question. So we've always wanted to be where our customers' data is. And we've talked about this for years. We've got partnerships with other companies. And the language we use internally or we'll use when we're talking to a prospect is we wanna wrap around your business.

[00:08:22] - [Speaker 1]
We wanna wrap around your current technology investment. We don't think that we're a big ivory tower that you're gonna move in tomorrow and everything's gonna be perfect. We know that you've got other relationships and other processes that we want to continue. And so we've got customers at Palantir, and they're working with an amazing company to help transform their business. It was really important that we supported them in that endeavor.

[00:08:42] - [Speaker 1]
Palantir is a great business. They are not marketing specialists. I don't think they'd ever tell you that they are or purport to be. And so taking Palantir's business ontology and combining it with Zeta's customer ontology creates a really unique and powerful view of the enterprise, like the broader enterprise ontology. And I think that's where the magic came together.

[00:09:06] - [Speaker 1]
So we built a version of the data cloud that operates inside of Palantir's governance, even their ethical guidelines on how they wanna deal with data and customer data and move or not move data. And so we had to create a very bespoke system that followed clear guidelines on what was right for their customers and our customers, but also gave them access to the wisdom intelligence that is endemic to Zeta's data and AI.

[00:09:32] - [Speaker 0]
And also before you came on, I was reading how you were talking online about how a context graph provides the structured memory behind an AI system. Every tech conference I've been to this year, there's been that word context has cropped up again and again and again, especially in AI conversations. So for non techies listening, can you just explain how this context connects business goals, customer identities, brand rules, campaign history, and all those previous outcomes in a way that a model on its own just can't? Because I I think this is a point that often gets missed.

[00:10:05] - [Speaker 1]
Yeah. Sure, Neil. But you're you're hitting on probably a bigger issue that I want to address for your listeners, which is context means a lot of things. A big word that people like to throw around and I'm not sure it's a us too, you know, everybody wants to be involved in this and they see the context can be a that skeleton key that opens all the doors. But I'm not sure that context is actually being deployed in the right way across organizations.

[00:10:32] - [Speaker 1]
I don't see much evidence of it. So the question was for the average non techie, what role does context play? And this is the easiest example. If I were to buy you a gift for the holidays, Neil, you know, I've met a few times. I'd probably get you something nice, maybe a nice bottle of whiskey and hopefully Take a break.

[00:10:52] - [Speaker 0]
Yeah. You know me well.

[00:10:53] - [Speaker 1]
And if you don't, you're gonna regift it, but hope hope that works. If I had the context of who you were and how you spend your time, maybe you like to go out four wheeling on the weekends or you like to rebuild old computers or you like to play guitar, my ability to give you a gift would be infinitely better, much more impactful for you. That's the way context works with AI as well. It brings context to the conversation that's more helpful. I think that's, what context does.

[00:11:23] - [Speaker 1]
The way it's built, the way it's deployed, and the way it serves a business, we can go as as deep as you'd like. And Zeta spends a lot of time researching and building these systems. I believe context should include a ledger of how it changes over time, and it should include a ledger of why AI systems recommend certain actions so that context can be accountable. And this is the big missing link for AI for the industry right now as far as I'm concerned is that very few AI systems are accountable yet. Zeta Ziz, and it's something we focus on a lot because we see it as the gap for organizations.

[00:11:59] - [Speaker 1]
But it's so easy for an AI, and you can go use any AI you want and get a recommendation for it from it. What should I do about this? You could put in some marketing information. Hey. What should I do about this?

[00:12:09] - [Speaker 1]
For it to give really accountable recommendations, a, it should probably spin up a bunch of, like, predictive machine learning models and run as much data as they can through it and make a prediction off of that. And b, it should take the outcome of the recommendation, so whatever you actually do, and feed it back in those models to see if it worked. And if it didn't, it should make those models better over time. And so context becomes a little bit of an input, but it's also I believe it has a very important role as an advocate for improving your brand, as a ledger for what has happened over time. And you'll read more and more about temporal context graphs.

[00:12:50] - [Speaker 1]
I think this is probably the next thing that you'll hear at every single conference you go to is how brains evolve over time and our thinking evolve over time. And it should serve as some kind of corporate governance and accountability so that people can go back and read things in plain text to see how we arrived at the decisions we arrived at.

[00:13:09] - [Speaker 0]
Beautifully explained. One of the reasons I wanted to ask you that is because I think many companies this year have remained stuck in pilot mode because their agents cannot act across disconnected applications or data, for example. So, what are the technical or organizational work that is separating an impressive demo from an agent that can actually be trusted in production? Where where are so many going wrong here?

[00:13:33] - [Speaker 1]
Yeah. So, you know, I I think I said this earlier. Being full is be is different than being nourished. Right? The pilot my opinion is we moved mostly from pilots to production systems sometime in the last nine months, but these production systems are still in isolation.

[00:13:51] - [Speaker 1]
I was talking with a CIO of a very impressive large publicly traded company. They've got a ton of sub brands. And they've got every vendor under the sun working with them, and they've got a bunch of AI initiatives. And they the the mood of the day was for creative. Oh, can you guys help us with creative?

[00:14:13] - [Speaker 1]
Yes. I can. We can we've got lots of AI creative tools. Nope. Love to sell you one.

[00:14:18] - [Speaker 1]
Great. But I'd rather fix your problem. Mhmm. So the challenge that organizations are going through right now is they've got a lot of disconnected tools. They don't have a central system for interoperability.

[00:14:31] - [Speaker 1]
Most technologies don't have that either. You're gonna start to see some patterns that emulate it, but they're gonna be missing some of the the crate filling, you know, of the good stuff that makes all of it stick together. But I I think that's where we are as an industry right now. And if I could sit down with every CIO, CTO, CEO, CMO in America right now, I would really harp on one thing. Start with a system that you need, and then they'd say, well, what's the system supposed to say?

[00:14:57] - [Speaker 1]
And I'd say, oh, great. Go ask your business leaders. What are their goals? What do they need to accomplish in an enduring way, not in a this week kind of way, and then you can design your system around your business. But you gotta go talk to the operators inside the organization.

[00:15:12] - [Speaker 1]
Going and buying a data warehouse and a couple of AI tools on top of it is going to be inadequate inadequate for the growth of your organization. And those companies that do transform with a solid system at the core of their business, they are gonna go to the moon. We know this is true right now because I work with a ton of startups, companies that are doing, you know, anywhere between a few million up to maybe a couple $100,000,000 in revenue. And the difference between those that are absolutely crushing it right now and those that are just plotting along are those that actually have an operating system. They've brought their AI together to help them get better outcomes.

[00:15:47] - [Speaker 1]
They can measure the company health by this AI. Very few enterprises, those that are like public can do that. They can measure the success of a pilot, they cannot measure the sex success of their business health by way of improvement from AI.

[00:16:03] - [Speaker 0]
And I also read online that you describe, every AI service as both a consumer and a producer of context. So how do you stop that continuous learning loop from reinforcing maybe poor assumptions, biased decisions, or incorrect recommendations? And the the other big phrase that we hear a lot at the moment is AI drift.

[00:16:24] - [Speaker 1]
Yeah. That's a great question, and I'm impressed that you went through and read all of this stuff. It's a good reminder of the things that we've written or have written. This has been a problem that's as old as time with AI. We always just called it the AI echo chamber Yeah.

[00:16:39] - [Speaker 1]
Where content just starts to reinforce the same behavior over and over and over again. In the early days, we actually introduced something called the serendipity quotient. Just injected randomization into it. Now things have refined and grown a lot since then, and there's tools for detecting drift. There are evals that have the expectation that everything from hallucinate hallucination to echo chamber dynamics will happen, and it will look for those and flag them so that people are alerted.

[00:17:14] - [Speaker 1]
And there's a bunch of data science approaches that are governed in a way to ensure that that doesn't happen. So right now, I'm not worried about producers and consumers working as long or both contributing AI systems as long as the same agents aren't testing or aren't grading their own homework. So I'll give you an example. In engineering, we refactored our entire engineering harness so that our engineers could do AI native development. It's been a huge overhaul.

[00:17:43] - [Speaker 1]
Again, another area I don't think the market has solved for yet, which is why we had to build our own. But those agents that build code are not also the same agents or even same LLMs that are doing QA or building evals or building the the the specs. Each one of those have a governed process that assures they've got an opinionated approach to understand and judge that code before it's deployed. That's driven by two sets of rules. One is static system rules that are built by our engineering team.

[00:18:20] - [Speaker 1]
Another is built by inferred rules built by the agents themselves because those systems need to be self improving over time. Now those are reviewed by engineers, but they're contributed to the overall criteria that dictate dictates the success of anything before it is shipped.

[00:18:34] - [Speaker 0]
I love that. And I'd love to give everyone listening maybe an actionable takeaway here. And if we do have a CTO or a CMO that is listening and they think that their company already has enough AI tools but accompanied by disappointment with the results, what should they examine first? What evidence would show that their infrastructure is beginning to turn AI spending into measurable value? Where where should they go?

[00:18:58] - [Speaker 0]
What should they be doing?

[00:19:00] - [Speaker 1]
Yeah. It's a really hard question to answer, Neil. So I'll talk about the dynamics they can look for first, but I I can't give them a golden KPI because every business is so different already. So this is what I'll say. Most people that are using AI today, engineers, I'll use examples as I go, engineers that use AI today, most of them are using AI quite successfully.

[00:19:25] - [Speaker 1]
They're not saving any time. Their code quality is not necessarily any better. So they're just at a point where they're iterating really quickly. They don't have all of their systems integrated together, and so they're spending eight, ten, twelve hours a day working, and they're getting about the same amount of workout from yesterday. There are some key indicators that tell them they're doing really well.

[00:19:48] - [Speaker 1]
We use them. We use some credible tools like Span app is a good example. It tells how many merge requests. And if you look a little bit deeper, you may like, well, how many lines of code for each merge request? And people desperately want these indicators to to claim victory.

[00:20:03] - [Speaker 1]
But if you ask smart engineers that are looking at the overall system, is it becoming healthier or not is what they want? And those indicators aren't actually changing. So they're committing more code. The code has more AI, and this the board says, yay. We're committing code that says AI.

[00:20:18] - [Speaker 1]
But is your overall code base getting healthier? Is it getting more performant? Is it getting cheaper to operate? They don't necessarily have that visibility. So there's a disconnect between what we're measuring and what business outcomes actually mean.

[00:20:30] - [Speaker 1]
And I'll go back to the example earlier of the creative tool. You have a new team that has to learn a new tool, and there will be efficiencies that come out of it. But without a transformation of your organization and going soup to nuts on creating campaigns, use campaigns as an example because it has a creative step in it, but you're going to create an audience and a campaign and you need to do QA and you need to queue it up to send them whatever channel you're gonna use, and then you need measurement and dashboards. Unless that overall trajectory that creates engagement with a consumer that was better beef than before, which creates more sales for your brand has been transformed, then you're probably, over optimizing for like a local variable. You're hyper optimizing for something small, which happens in AI all the time.

[00:21:16] - [Speaker 1]
In fact, that is oftentimes what leads to the echo chamber we talked about earlier. So maybe the question is, for individuals that are working inside a broader organization, if they feel like they could be at risk of being an echo chamber, like they're hearing their same news or they're celebrating the same wins as the people around them, maybe start asking people in other organizations, other part of the business to see how they are also using AI and see if you can create some collaboration. Take what you've learned, the data access patterns, the the efficiency wins, and connect them to other groups inside your organization to create broader change. When we do that and we start getting a a system, and when things work together, that connectivity creates a system, I believe that we're gonna see real change inside of organizations. And some companies are there.

[00:22:00] - [Speaker 1]
The very best are there. A vast majority that I speak with day in, day out, not even close yet.

[00:22:07] - [Speaker 0]
And I cannot thank you enough for sitting down with me today, especially because I know your focus has got to be all around Zetta live at the moment that is coming up next week. I know you're locked down. There's so many, different embargoes, etcetera, that possibly surround the event. But is there anything you can share around what we can expect from the event this year, which is build the most intelligent event of the year? So bold move there as well, but what can we expect?

[00:22:33] - [Speaker 1]
Beautiful. So you're right. I can't say too much. What I can say is we have an incredible team at Zeta. I have never been more humble or humbled or proud to be on a team of this caliber.

[00:22:46] - [Speaker 1]
They are moving faster than I've ever seen a team work in my life to bring real products to market. That means it goes through a beta with customers. They've got sealed lips, but they're behind the scenes helping us design these programs and build the right things that drives adoption. At Zeta Live, we'll be talking a lot about the theme of systems and how technology needs to come together. We'll certainly be hitting on context quite a bit.

[00:23:10] - [Speaker 1]
Athena, which we announced last year, is getting a glow up, one that will align well with today's conversation. And then as always, Zeta has a variety of intelligence products that are, deeply rooted in our our competitive moat as a company, but also in our DNA as a company, looking at the inner section of AI and data. We bring in a bunch of those competitive sorry, those, intelligence products to market through the lens of verticals and talking about how we can service new verticals specifically in a way that starts to give them that that underlying system that drives broader business growth.

[00:23:47] - [Speaker 0]
Exciting times ahead. And if all that geeky stuff is not enough, people listening, and this is a tech podcast so it should be enough. But there's also Olympic gold medalist Lindsey Vonn and Kevin Hart as well. So a list event there. Right?

[00:24:01] - [Speaker 1]
Yeah. She is Lindsey Vonn is something else. I really am looking forward to meeting her and hearing her talk and, you know, I've been laughing at Kevin Hart for a long time. He's in half the movies my kids watch anyways, so that'll be a great event. There's several others that will be there as well.

[00:24:17] - [Speaker 1]
Tiffany Haddish, to say the least. The last thing I'll say about Zeta Live is that it's not a pitch for Zeta. It's the way people think about industry more broadly getting together to have really intelligent conversations about how we evolve what we're doing from culture to the jobs we have and certainly about technology and Zeta's part in that is is important, but it's not the only thing that we talk about at Zeta Live.

[00:24:42] - [Speaker 0]
Awesome. And for anybody listening who wanted to find out more information around Zeta and everything that we've talked about today or some of those big announcements as they drop from Zeta Live, where where should they go?

[00:24:54] - [Speaker 1]
Once again, they should go to zetaglobal.com. They can get all the information there.

[00:25:00] - [Speaker 0]
Fantastic. Well, I love having a catch up with you. We do cannot leave it a year until we talk again, especially as you've left me with a few teasers about the event. So I think we're gonna have to get you back on later in the year or early next year. But more than anything, just thank you for for that update.

[00:25:15] - [Speaker 0]
For everyone listening, I will put the links in the show notes, so please go check that out. But I'll look forward to speaking with you again soon. Thanks again.

[00:25:23] - [Speaker 1]
Neil, it's always a pleasure. Thanks for having me.

[00:25:25] - [Speaker 0]
I think Chris leaves leaders with a test that cuts through much of the noise around AI investment right now. A team might create additional code, campaigns, images, or recommendations, but activity alone doesn't show that the company is healthier, faster, cheaper to operate, or better for customers. Because the stronger evidence only appears when data, tools, people, and goals all connect across the full process and outcomes return to the system as feedback. And we also need to highlight accountability. Agents should record why a recommendation was made, learn from what happened next, and be assessed by independent models and human reviewers.

[00:26:12] - [Speaker 0]
So, again, a massive thank you to Chris for returning to the show, sharing what Zeta Global is building around data context and AI infrastructure. You can learn more at zetaglobal.com. Follow the announcements from Zeta Live. You're gonna be some big things coming out of that. And over to you, is your organization building an AI system or just collecting tools that perform impressive work in isolation?

[00:26:38] - [Speaker 0]
Love to hear your experiences and thoughts on this techtalksnetwork.com. Remember, I'm at a lot of tech events between now and the end of the year, so have a look at the events page at tech talks network. And if you are at any of them, let me know. Be great to meet some of you in person. But that's it today.

[00:26:57] - [Speaker 0]
We're out of time again. Sorry about that. But I will be back in your podcast feed tomorrow. Same time. Same place.

[00:27:05] - [Speaker 0]
Bye for now.