Wrapping marketing in agents, augmenting not replacing
We're in the free trial period of AI right now, so it's important to be frugal and use just the right size of model for the right task. The true promise won't come from reducing costs; it will come from reaching into the next layers of possibility and solving problems that were completely unreachable.
Ankur Mathur is CTO and Head of AI at Experiture, an omnichannel marketing platform that runs email, direct mail, and SMS campaigns for brands, many of them agencies with decades-old relationships. His team has wrapped the core platform in a layer of AI agents that give marketers a language-based interface, governed by what Ankur calls an autonomy slider: dial it all the way up and the agents handle everything from design to journey orchestration to analytics, flagging you only when something interesting happens with your audience. Paid media still stays with humans, but composition and the surrounding workflows increasingly run on agents.
Ankur has spent close to a decade in machine learning, long enough to have lived through an AI winter and several hype cycles. Before Experiture he was Head of AI at Iterable for more than five years, where he founded the company's AI and Experimentation group. He holds a master's in computational physics from Cornell, a background that shows in how he talks about models as organisms to be probed from the outside rather than black boxes to be trusted. Alongside his day job, he advises startups across creative tooling, recruiting, and robotics.
Ankur's engineering worldview centers on the harness around the model, not the model itself. The LLM, he argues, is a small part of the solution; the real craft is context and memory management, retrieving the right data at the right moment from vector stores and data lakes, then logging every interaction so a fleet of background agents can evaluate how the AI performs in the real world. He favors a multi-model, cost-aware setup, using the smallest model that fits each task and reaching for private models when customer IP cannot leave the building.
Ankur is betting that AI's real payoff is not efficiency but reach. He frames the current moment as the free trial period of AI, one that rewards frugality now and will pay off later by solving problems that were simply unreachable before. Drawing on Kahneman, he sees a chance to offload overloaded deliberate thinking to machines so people can step back and build what actually matters. He wants to hire engineers who think like product managers, and he insists the goal is to augment marketers, not replace them, moving people up the value chain rather than out of it.
Read full transcript of interview
In this conversation: Josh Rubin (Host, CTO Studio) and Ankur Mathur (CTO & Head of AI, Experiture).
So what do you do?
I'm the CTO of a marketing technology company called Experiture, and also their head of AI.
AI is obviously affecting the heck out of marketing. What kind of AI stuff are you incorporating into those systems?
Yeah, that's been a very interesting journey. So we've put a layer of AI agents around our core platform that is now allowing a language-based interface for our marketers. We have this concept of an autonomy slider, which lets the marketers decide how much AI they want in their workflows. And if you slide it all the way to the right, the agents are going to do all your marketing, from design to journey orchestration and even analytics, and notify you when something interesting happens with your audience.
Spend? Is it going to buy ads for you as well?
So this is mainly email and direct mail, but we also do some SMS. We haven't quite identified the advertising side of things. The paid media is still handled by humans, but with AI help. That's still going through that phase; a lot of that does require some careful attention. A lot of the agentic work right now is focused on the email workflows and direct mail composition and stuff like that.
So are you building your own AI agents based on something like a Claude MCP, or are you using ChatGPT? Are you hitting multiple different models? Because, for example, if you want to write an email, one model generalizes a bit better at writing emails than another, but that could change tomorrow. How are you approaching that?
Yeah, that's a great question. We do have a multi-model approach, and we use the best model with the most economically feasible footprint for each type of task. So for composition, yes, we'd use something like a Claude Sonnet or a GPT-4.5 or a Gemini. And we have an interface where you can state your preferences, assign profiles and different types of budgets and policies to what you want to do and how much you want to spend. But then there are certain heavier cognitive tasks, like analytics, where you want it to sift through potentially terabytes of data and come up with insight that's going to be interesting. That's where it blends into classic ML work that I've been doing for almost a decade now, where you can actually do predictive analytics and automatically segment your audience. So when you get into that realm, you want to use heavier models, or even private models, because you don't want to send a lot of that core IP-type data out into public models. So we've got a pretty stratified setup in terms of using LLMs.
Where does the memory live in that case? If an individual customer is dumping in all of their private data, all the emails they've written in this tone, and you're using multiple models, where does that core bit of content live?
That's where most of the engineering comes in. This context and memory management is essentially what this harness engineering thing is doing. The LLM is a small part of the solution; managing the memory, being able to retrieve the right type of context and feed it to the LLM at the right time, that's where all the art and the science and the engineering happen. So yeah, it's in databases, vector stores, and even large data lakes. That's where all the raw data is. And then there are traditional pipelines purifying that data and processing it, and then storing the signal in a spot where it can be easily retrieved by these agents. But that's kind of the more interesting, geeky side of marketing tech.
This is also empowering a lot of... just as we're seeing maybe less hiring of engineers, or old-school coders being replaced by this, this is replacing old-school marketers in some ways. Are you seeing your customers embrace the AI efficiencies?
Yeah, that's where this autonomy slider idea comes in. Our intent is not to replace, but to augment. And I'm sure you hear this a lot, but we really mean it. We've got really long relationships. This is a 20-plus-year company, and it has very loyal relationships with our customers. A lot of them are agencies, and print agencies particularly tend to have old-school marketers, but they're learning AI through our platform and learning to trust it, to embrace it. And that's another rewarding side of my job that I enjoy a lot, because I've seen the AI winter, I've been through a number of hype cycles. I fundamentally believe that after this kind of turmoil, things will stabilize to a point where humans are able to devote their attention to things that actually matter and that bring them joy and satisfaction, and leave a lot of the tactics and executional overhead to the AI. So that's how I approach this.
So the challenge with what you're doing is you actually have to interface with these newspaper-style, old-school print media guys and teach them how to use AI, how to engage with AI. That's not easy, but that's kind of the core product work that happens. What kind of resistance have you faced? AI has a terrible problem of communicating its value outward, and you're essentially having to do that one person at a time as you develop your product.
Yeah, the whole spectrum of people, and one has to approach it from different angles. So there are some people who just will not open that little popup that says, hey, here's my agent, Zaira, would you like some help? They're like, no, I want to do it my old way. That's fine. Maybe one of their colleagues or interns will come in and open it. And then we refocus on the people who have some openness to it, and we ask for their feedback and we want to make it better, more useful. This is kind of fundamental to AI engineering: we always gather the feedback automatically. So we're logging everything that's happening, when people are abandoning that conversation, or when their sentiment turns negative. We have an entire array of agents in the background doing this backend work, including the data processing I mentioned earlier, but also this feedback processing and evaluation of how the AI is doing in the real world. That's critical to understanding how humans are reacting to it. So we use some of that insight, and we basically look for champions within these groups and try to get their feedback and get them to teach their colleagues.
That's interesting, the dark, negative way of looking at this: if you're using AI, we know you're using AI, we know how you're using AI. It's not that we're spying on you, but everything is being watched, evaluated, judged for sentiment analysis, all of those things, which I imagine makes the actual project validation much easier to do than it's ever been historically.
Yeah, that's true. Logging and telemetry is an old game, it's been around for a very long time. But the ability to sift through all that data and use AI to analyze that exhaust of all the interactions is new and very valuable. And, to clarify, all of this is sitting in their private data warehouse or data lake. So it's not getting leaked to anyone else. They've already trusted us with that data, and we treat this the same, with SOC 2, HIPAA, all these compliance frameworks, because it's their data. But it's about their usage, and not just about their customers.
But you're the one sitting on top of all of that, so you can draw conclusions from it. I imagine you can roll out new products much faster than you ever could before. The difficulty is, how do you keep yourself from blue-skying yourself to death? This is something I'm trying to get to the bottom of. In a world of abundance where I can make anything, how do you choose how much time to spend on something?
That is such an interesting question, and I think about this a lot. I actually think one has to take a big step back and really think about... Kahneman brought this system one, system two way of thinking. System one is our instinctive, pattern-matching, reflexive, almost limbic thinking. And system two is slow, deliberate reasoning. System two has been overloaded with the advent of technology and the pace of technology. We're overwhelmed with this rational, slow thought. And now, finally, there's an opportunity to delegate some of that to machines and actually look at what we should build to make life better, and what will actually bring the species forward, and not just be worried about the tyranny of the urgent, just thinking about what needs to be fixed to get through the next quarter and meet Wall Street expectations. It's an opportunity to really look at blue sky, like you say. And there's so much. That's why I spend part of my time advising startups in various domains. It's been really interesting. People are reimagining all kinds of things, from creative design. There's a company that my friend runs called Uni that is focused exactly on this: how do you take the creative process and augment it with AI in the most free way possible, so that creativity is harnessed and used in a way that brings joy to the creator, and not captured or replicated by the machine. That's one of them. And then people are looking at the recruiting and hiring process, how you find people and match people in a more authentic way. And of course there's tons of opportunity in the physical world, which is also moving pretty quickly with robotics, that would really open up so many new avenues for elder care and even...
Yeah, the palliative care aspect of all of this. I thought palliative care was the only real use case for VR I've actually seen, getting a nursing home strapped in.
You talk about hiring. How big is your team?
My team is around 50.
So we're obviously seeing things like Intuit and Facebook laying off at the same time. And Dropbox is actually profitable, whether or not we believe that profit is real, given the definition of a profit. But companies of your size, mid-tier, mid-level companies, or say Series B companies, they're not firing people. They're not necessarily hiring either, but the reason they're not hiring is they don't exactly know yet what they're hiring for. So that's what I'm trying to understand. As an engineering org, as a CTO, my productivity is way up. Where do I need more bodies? What type of person is the new type of person, and how do I find that type of person? And for you: you have all the money in the world, and you can hire the perfect person to join your org. Who is that person?
Yeah, that's a really good question. And I'll touch on the points you made earlier. The reason I think the hiring is slow or ponderous is because there's an intent, at least that's how I see it, to try to re-skill people, to take your existing people and move them up the value chain so they're not just coding, testing, and taking orders, but they're thinking like a product manager. And that's the answer to your question for me: I want to hire people who could be product managers but have been engineers, at least in the tech org. They understand how to build the software, but they're more interested in solving the business problems. They can think like a business person, and actually use AI to build it, and they like to free up their time. They're not people who get so caught up in the art of programming that that's all they care about. There are people like that too, and I respect them. But for businesses right now, at least in the application layer, the most valuable people... And there are still people building the core foundation models who need to be those kinds of craftsmen, who all they think about is high-dimensional math, the people who are into the core engineering problems. There's definitely a place for that. But at least in the application development space, you need people who are more user-focused.
Which then leads to some other interesting transitions. Where's your team based?
They're distributed. So a lot of it is in New York, some are in South America, and a lot in India.
So I think what's going to be really interesting, if you look at Tata, if you look at companies like KMC, all of these groups: when you're asking people to think like a business person, you're not just asking them to think like a business person, you're asking them to think like the business person or the user in that specific cultural context. And cultural context matters. It's not that people in Latin America or the US or Europe or India or the Philippines are any smarter or less smart. But cultural context is a real thing. How do we navigate that? How do these orgs navigate that?
Yeah, that's a great question. I think it's actually important to respect that diversity of thought, approach, and culture, and use that to create products that adapt to it, too. That's again where, once you start thinking in terms of software that adapts and can respond differently to different types of stimuli, the cultural aspect becomes interesting to take into account. So to make it more concrete: if you're building this product for an Indian audience, the agent and the memory and everything should adapt to it. And over time, as it gets more feedback, it starts adapting and behaving a lot differently than what a European or an American agent does. So it can evolve and learn those characteristics.
And it's fascinating to watch. Language is software. How we think is really influenced by the language that we operate in. And that's why some of the Chinese models are able to, from what I've heard, the character count is much more truncated, they're actually able to squeeze more into certain things, but they lose nuance sometimes. I don't know how true that is, but just as different programming languages are different, different languages are different.
That's a really interesting question, and I need to maybe do a little more research on it, but that makes intuitive sense to me based on my understanding of language models. If they're trained on a corpus of Mandarin text, the embeddings and the internal structure of that knowledge encoded into the LLM is going to be significantly different, and the associations it makes are going to be different, and it would generate a different kind of text. So I'm sure that's true. And with a lot of regional training that's happening right now, even in India, all of that is going to start emerging. It's going to be very interesting to see. And at the end of the day, we have to probe this thing from the outside, almost like an organism, because once you get inside it, it's all high-dimensional numbers, right? They're just vectors, and it doesn't make sense anymore. So you have to go outside. And that's what some companies are starting to do for explainability. I just read this week that the ex-CEO of Twitter started this company called Parallel Systems, and he's created this product called Index that actually probes the LLM. It trains it on text from big publishers like The Atlantic and others, and then it creates this explainability layer, which tells you: this answer was 20% due to this text from The Atlantic and 30% from The New York Times. So now it can attribute the answer and assign value and pay them based on it. It's a whole new way of looking at things.
It's also like a 23andMe of that.
Exactly, yes, that's a really good way to put it. So I'm very excited about those approaches.
This is where we'll end up. All of this only matters, or is only useful, if we can validate the outcome. Ultimately, what validation is for different companies and different people changes, and you have to figure out what that is in the beginning. But we're only now at the beginning of having the frameworks to determine the validation of these new tools. Are you confident that all of these AI tools you're putting into your systems are going to net out? You're going to ROI, you're going to get the value, it's valid? Do you have the framework in place to evaluate that?
Yeah, that's the big burning question for most execs. I feel like in a lot of ways we're in the free-trial period of AI. And soon the bill is starting to increase, the cost of infrastructure and everything, so that's something I personally keep a close eye on. I think it's important to be frugal and use just the right size of model for the right task. There's a lot of work going on in that space too, to make these models more efficient, and Meta and others are constantly coming up with models that are doing more with less horsepower. So I think the framework is in place. Right now, I don't think it's net positive or green for most people. But it's also a very experimental stage where people are willing to take a bit of... people are investing heavily because the promise is there. And I think the true promise will come from being able to solve problems that were completely unreachable so far. That's going to unlock the value, and not reducing operational costs so much. I think it's going to be more about reaching into the next layers of possibility.
Thank you so much.
Be part of the
conversation.
Whether you're a CTO who wants to be featured, a company looking to sponsor, or an engineering leader wanting a seat in the room — there's a place for you here.