#297 Jeff Lunsford: How Agentic AI Will Redefine Every Digital Interaction

0:28This is happening so fast. software to collect your first-party data, to find your attributes that you want to, you know, your own language to enrich this data with. Telium's original mission was to help the largest companies in the world collect and enrich and then create experiences around their first-party data, their customer data. So as software ate the world and the customer experience was created by more and more specialized software packages, each of those software packages had its own data set around the customer. And we felt the industry needed a neutral provider to come in kind of the way TIBCO did in the mainframe era or WebLogic did in the app server era and be a neutral provider that connected to all these upstack systems and to all the downstack transactional systems and listened to the event level data from customer experience systems, resolve that event level data around visitor profiles.

1:30So we basically turn machine data into human data and then allow companies to enrich those profiles with their own language. Right. So a big pharma company might have different attributes they would assign than maybe a big airline would or a big bank. You know, and, you know, Telium works with 10 of the 20 largest banks in the world, six of the 10 largest telcos, three of the five largest airlines. So these companies have customer data flowing everywhere. And, you know, so we'll perform that identity function. We'll allow them to enrich that data. This all happens, Craig, on a high performance real time platform.

2:05And it all is a governed and trusted platform because, you know, we've had privacy regulations around the world. And so now a big airline, big pharma, big hotel chain, big e-commerce company can collect all this data, follow privacy rags, and then use trusted and compliant customer data to fuel customer experiences. And now this was all happening in the MarTech, AdTech, CRM landscape prior to AI. It turns out that as these big companies embrace AI and start using AI to fuel customer experiences, this is the data set that they need to populate the context window, right? To give you this real-time context, the behavioral data around your customer's activity that the LLM needs to understand, right?

2:59Right. And the LLM likes structured data. It likes labeled data. And that's exactly the stuff that Telium produces or helps our customers produce. So that's kind of a long introduction, but it's all about basically a trusted platform. And whether companies use Telium or not, they have to solve this mandate, right? They have to follow privacy regulations. They really need to get this customer visibility into real-time customer data because that real-time data gives you the highest intent signal, whether you're an e-commerce platform or travel platform or whatever. And those intent signals are what that data is, what the models can use.

3:45And by the way, this isn't always an LLM. It could be a fraud algorithm model. It could be a cat boost model. It could be an LLM. But basically, more data equals better outcomes with AI. but it's not just a completely unlimited firehose of data. You know, we're now learning in AI that giving too much context can lead the models to wander, right? So, you know, one of our founders talks about instead of finding the needle in the haystack, how about creating smaller haystacks? And so, you know, you kind of want to give that tighter, higher resolution profile to the model so that it can do its thing.

4:30Yeah. And let me ask some questions about that. That's interesting. And I haven't sort of broken down this data aggregation and rationalization process. So how many different channels can you pull data from and how many different formats, not specifically, but... Sure. So basically anything that's digital, So the web, mobile apps, call centers, point of sale, set top boxes, over the top television, kiosks, ATMs, store data, you know, from Wi-Fi routers and stores, basically any kind of event level data. And we take that again, that machine level data, and we resolve it then around identity so that you get an actual visitor profile.

5:26and it all happens in real time. So it's a big, you know, compute load at real time. And then we allow that profile to be available back to, you know, like let's say, put AI aside for a minute, if you're just doing classic online advertising, so Meta, Google, Snapchat, Pinterest, Reddit, they all have conversion APIs. And when a customer's visiting, You can send them data in real time and you give them feedback on what is converting. And then they actually use AI behind the scenes to basically help companies dial in the focus for their audiences that they're targeting. And we're seeing an average of 26 % improvement on return on ad spend for companies using these conversion APIs.

6:17And so, I mean, you know, you talk about there's 500 billion or whatever digital ad spend just in the US and Europe this year, probably. So 26 % of that is a pretty big number. And then when you add AI to it, you know, we're seeing even tighter and more impressive outcomes. Like we had an auto dealership that was using Snowflake and Telium to get their data act in order. And, you know, they went from a 16 % conversion rate at the bottom of the funnel, people that bought a car, to a 30 % conversion rate, right? So more than doubling. Yeah. And when you say that all that data coming in, it's got to be put into a machine-readable format, right?

7:07If you're getting call center data, how does the system convert that? Is it converting it into vectors or is it using natural language?

7:26We have, think of us as before like the vector store or the graph database. Now, for AI purposes, we definitely are helping vectorize this data and using graph databases to show the relationships of the attributes. But prior to that, just the core tooling platform is a deterministic, not a probabilistic model. And we just understand all, you know, we have 1300 connectors that we call them connectors, but where we can ingest inbound data. And part of building one of those connectors is understanding the data format you're going to get from that source. right? From, you know, Samsung TV or from a mobile app or whatever.

8:08And then you do, you map field by field, these elements that are going to come in on the event level data. And then our customers can use, we call it an identity key ring, but, you know, multiple different

8:23dimensions can be combined in identity key ring to resolve around identity and to turn this event level data and associated with a human, right? And so it could be email address, it could be device ID, you know, it could be some other signature URLs, all that kind of stuff. And so that is all, that all happens deterministically. And this is a pretty key point, Craig, because there are a lot of folks out there talking about probabilistic matching and things like that. But when you, when you look at global privacy regulations, you cannot have probabilistic anything going on because if a customer one of the most important things we do is we capture the consent status of a customer of a visitor to any of these digital you know channels and then we propagate that downstream to all the systems that can benefit from knowing that consent status right because this company over here can target if they target someone who just opted out, they can get fined.

9:28And, you know, it's very important that you don't start saying, oh, we think this person might be, or these events might come from that person. So we're probabilistically going to stitch them together. And then you start targeting that. And if your probabilities were wrong, you are now in violation of privacy regs, right? Yeah. Yeah. And all this is happening in real time in the cloud. I mean, are you a SaaS platform? Yes, we are. So the way you think of it is, Telium, I would say like we're just software, like think, you know, when you're talking to a customer, think of us as running inside your four walls.

10:09But yes, it runs in the cloud. And so you're just going to use our software to collect your first party data to find your attributes that you want to, you know, your own language to enrich this data with. And we think of it kind of in three tiers, right? So you've got all your customer experience systems that are typically outside the firewall. Also, now they're mostly cloud based systems. Every once in a while, you still have a big bank or something running a CMS on-prem or something like that. then you have telium which creates this this trusted data layer and think of it as kind of like a high performance cache that sits at the edge of the network and then in the back office you will have your point i mean sorry your transactional systems like a dda system or credit card processing system and you might have like your data warehouse right like a cdw like a, you know, Snowflake or Databricks or Redshift from Amazon or Google, BigTables, and then our BigQuery, sorry.

11:13And then those are what I, what we call kind of like the stateful at rest back in customer data systems where you want to run your heavy workloads and your heavy analytics and things like that. We are the high performance cache where when someone is visiting one of your properties, you want to have that profile at the ready so you can populate, you know, send it to the LM in an AI use case or send it to the media property in an ad tech use case in real time with the most up to the millisecond behavioral data, right? Because it's what I'm looking at in the current session gives you much more intense signal than what I was looking at two days ago or two years Right.

11:57Yeah. Wow. And presumably you've been developing your platform or your software as AI has gone through these successive advances. And right now we're in the agentic phase. I'm a little bit of a skeptic on agentic AI. I think there's a lot of, what are they calling it, agent washing, where there's not necessarily some real agency there. But can you talk about how your system has adapted to those changes? Sure. So first of all, yes, there's probably a lot of agent washing going on, Craig, but one of my favorite science fiction writers has a saying, which is the future is already here. It's just not evenly distributed.

13:03Right. Yeah. And if you look at, for example, in crypto, there are already entire agent economies like on virtuals and pump dot fun where agents are interactive with other agents. And, you know, Pump has 750 million a day of activity and Virtuals has 125 million per day, like their own economies. And so what we see, we believe, and this year, by the way, you're already going to have 136 billion of agentic commerce in 2025. And that is predicted to go to 1.7 trillion of agentic commerce in 2030. Can I just stop and have you, I'm not familiar with Pump or with the other platform. Virtuals. virtuals and when you say agentic commerce just give me an example of a transaction that you're talking about yeah so a transaction it's like pump.fun they're trying to compete with twitch where a creator can go on to pump and do their stream just like you would with your your podcast and they get paid based on activity how many people are watching them in the pump token and And, you know, it's kind of wild, but it is going.

14:30That's the bleeding edge. Now, the mainstream is going to grow rapidly because we see the biggest players in commerce and media promulgating protocol standards that are either being developed or are in production. the most uh high vis one so far was the open ai uh agent commerce protocol that they just they and stripe announced this week and i went on to open ai and i you know shopped on etsy and went all the way to the shopping experience still within the open ai so you've got this agent commerce protocol acp we all know what mcp was uh model contact protocol so i say mcp basically was giving agents hands, you know, so giving them the ability to go do something like maybe make a reservation, but it didn't give them the ability to do a financial transaction.

15:29So now you have all these different protocols coming at you. So the Ethereum group has promulgated ERC-8004 as a standard for agent identity. So this is a new trust layer that will go on. And then you've got Coinbase, which promulgated X-402, which is a payment standard. And then you have Google, which promulgated A2A, agent protocol. And they then have agent commerce protocol, which is inclusive of and supportive of the x402 from coinbase you have google and coinbase basically setting a standard you got open ai and stripe setting payment standards and so agentic commerce is simply where i can have my agent like chat gpt or any other go do a transaction for me in in in in the future so there's a couple things there so i i have spent 30 years in fintech and I'm on the board of the bank and helped the team launch the world's first internet bank back in the late 90s.

16:39So, you know, the intersection of technology and finance, which is now known as fintech, has been going on a long time and it's very fascinating to watch. But, you know, what we saw over the last 30 years, Craig, between the internet revolution and the mobile revolution, we're seeing even more change happen now over the next four years than we saw over the past 30. And because all the rails are built, you know, and so we believe there will be 1.7 trillion and other, you know, third parties, not Telium, you know, are predicting that by 2030. That's about, it's a little over 20 % of global digital commerce.

17:18So, I mean, if you're that, I think that's kind of a mid to bearish forecast. If you're bullish on it, it could be half because, you know, like Gartner saying, 95 % of B2B buying journeys in next year, 2026 are going to be initiated through agents. Now that sounds kind of crazy, but if you think about it, think about your own behavior, right? How, how much are you now starting it? This is B2B. Think about it. So I'm here at Telium and I want a new HR system. I'm not going to Google it anymore. I'm going to ask Chad CPT what the leading vendors are and which is the best fit for Telium. And so I could see where probably that 95 % next year is pretty accurate.

18:10So at the top of the funnel, the shopping or the searching will start with agents. And so it's just logical connection that these intelligent agents, once they have like an identity trust layer, like what ERC 8004 provides and payments capability, what like A2P or A2A provide, um, then we are going to see agent of commerce. And so we're preaching to our customers. We have about 700 customers around the world that, you know, this is coming. And so you need to build agentic front doors to your business. Right. And our belief, again, having watched the web and the mobile for the last 30 years and how that changed everything and anticipating what's going to happen over the next four years, our belief is that your current, let's say you're a company that has a million digital interactions a day to day.

19:13Okay. So that's your, that's your total digital customer engagement number. Well, in 2030, we think you'll have 2 million double, double the number of digital interactions. We think the current million aren't going to really go away because, you know, human behavior is, is, you know, it's a, it's a slow slope. It's not an event. And so you're, so you need to preserve everything you're doing to be successful and competitive with the current digital world. The other million are going to be agentic, right? And so you need to have that agentic front door that, by the way, it doesn't look like a website, right?

19:56It's not going to be agents coming to use browser interfaces. Yes, agents can use browser interfaces. So yes, that your website maybe will work to convert some visitors. But if you have a high performance agentic front door that supports all these protocols, the trust, the identity, the payment protocols for agents, then you're going to get more of the business than your competitor who doesn't have that. Right. So, you know, we're preaching, you're riding the risk, if you're not thinking about this now and architecting for it right now, you run the risk of losing a generation, right? So, you know, 10, 15 years ago, if you did not launch a mobile app, you'd lost a generation.

20:43You don't have any of my kids that are in their twenties and early thirties. Um, and in 20 years ago, if you didn't launch a website, you lost a generation, you lost, you know, my generation. And so, um, and even before that, If you didn't launch a voice response system, you lost a generation. So companies will lose an entire generation if they do not lean into this. Now, what's very interesting, Craig, so when you support enterprise companies that have to follow global privacy regulations, while we are evangelizing, they lean into it. They also now increasingly are setting up basically AI risk committees or AI approval boards where before you do anything with AI, you have to kind of send it up to the board.

21:32And that's actually slowing big companies down and giving smaller, more agile companies or more risk oriented companies an advantage. And so now what we're seeing is there was this concept of privacy by design that kind of came 10 years ago and when Europe started with GDPR. And we think the same, the best practice now in the big companies is they're actually taking these people from the AI approval board and embedding them in the development teams that are building these projects. because it's kind of like having an embedded media person like you out on the battlefield, right? So that helps bigger companies speed this up.

22:22But we believe it's just what I call a strategic mandate. Embrace it and build for it or you'll miss a generation. And so you're talking about not just preparing training data. you're talking about governance and and and how agentic needs real-time policy enforcement is that right yeah it's governance and it's um so the training data you can't really use unless you've anonymized it you can't really use customer data in your training because of the privacy problem, right? Because those customers may opt out at any time and then that taints your whole model. So for customer specific AI use cases, you need to wait and populate the context window at time of inference.

23:18And that is the data that I'm talking about where this high performance trusted cache can give you the visitor profile at time of inference comes in really handy. Yeah, that's amazing how this stuff is developing. And consumer data platforms, are you providing dynamic control systems for this AI-first world that we're looking at? so um yeah the product category we invented back in 2013 is called customer data platform and it oh i'm sorry not consumer yeah yeah that's right it um it was such a good idea that you know four years later we had 130 competitors other companies saying oh we do customer data platforms so um you know the i the simple idea was what i started with just okay you got all this data everywhere about your customer and you need to resolve it into visitor profiles and be able to leverage it, but also govern it.

24:37And so then when AI kicks in,

24:44and if you think about this agentic world, prior to AI, there was this concept in Martech called journey orchestration. You could send a customer down different paths. And what we always preached was, look, you can't do some beautiful WYSIWYG customer journey map in a software interface because you just run at it. There's too many permutations of possibilities on most digital. And so what you really do is you orchestrate customer journeys by orchestrating data. And so when a customer does this, you enrich the profile, then you have rules upon which the customer, you know, you can enhance the customer experience based on what they just did.

25:31And then you have more rules and more rules. And so through data orchestration, you can do amazing journey orchestration. And so thinking about agent orchestration is sort of the next logical step, right? So it's now, whereas these business rules that you were doing data orchestration with and journey orchestration with were deterministic, within the agent world, everything is a little probabilistic. Now, it's getting more and more tight, right? Like, I tell my team all the time, like, you have to fly the airplane looking out of the front of the cockpit. I was a naval aviator way back. And, you know, it's like you don't fly the airplane looking in the rearview mirror.

26:22And so you have to assume this technology is getting better and better. So three years ago, we saw 15 percent hallucination rates. And so all the naysayers were saying, well, you can't use AI. Well, it was it's now less than one percent in 2025. You know, it'll be less than point one percent 2026. So you have to assume these things are getting closer to deterministic.

26:53And what will happen is you'll have an agent that, you know, you'll ask it a question. Here's an example. So one of our customers did this. They said, OK, so this visitor is on our website. This is a large global chip company. They're on our website and they've looked at these three products and our 80 ,000 parts skew. Speculate for me what industry they work for, what kind of project they're working in, and in what three other skews our sales team should call them about tomorrow to include in their project research. So kind of a B2B use case. And then they would ask the LLM that. The LLM would answer with the three parts and they'd put that in the CRM and then the sales rep would get a notice.

27:39Right. So of, okay, now call this person. Now it's very simple to say, just, let's just put that back in the website in real time right now too. Right. You know, and so that kind of thing is what we're seeing, you know, very simple, but very high value AI use cases with this data. yeah that yeah and could that could then another that could be the first agentic output and then there could be another one and another another one if you think about it and that would guide the customer journey but but who's can is this all um uh being orchestrated by a reasoning agent or or do you have uh sort of rules in place what's the control mechanism for for for deciding what you're going to send to the website, which you're going to send to, you know, the CRM and that sort of thing.

28:37Yeah. So the control mechanism pre-AI is we have rules and, you know, we'll say, you know, if a customer abandons their shopping cart with three items in it, then you would add them to a retargeting audience on Google or Crateo or whatever. Right. And then, or if a customer converts and buys something, then you would suppress future media. I mean, how many times you've been retargeted for stuff you already bought? You're like, these people are wasting their money, especially for a long considered purchase. Like I just bought at least a car. Well, that's going to be a three year lease. So why, you know, why waste money advertising could be for the next two years at least.

29:15And so those things were rules that our customers set up. Now, and again, the action then happens through these connectors we talked about. So now it'll simply be you'll ask a model, you know, what you should do. And you can ask the model any question you want, as you know. So you can say, hey, given, you know, this audience profile that this person fits in and their recent activity, you know, what is the next best product recommendation we could put to them the next time they visit or even on the next page in session? right that's where when you get into the real time it's really the in-session change of journey and if you take this to its logical conclusion correct uh every like the concept of an audience we think is dead we've been preaching that for a long time basically every every visitor is an audience of one and you're going to create a personalized experience for them and then ultimately and i don't know if this is one two or three years if you think about like what Sora can do and all these other real-time content generations, like every visitor should have their own visual and audio experience.

30:33Right. Now, remember, that's still the first million interactions. Now for the other million interactions where their agent comes on their behalf to transact, that'll be different. That'll be through, you know, APIs and there'll be data flowing back and forth, but it will still, You can still determine and modify what you send to their agent that is working on their behalf to drive conversion. And, you know, a conversion isn't just an e-commerce term, right? We work with large banks, large lenders. Just filling out a loan application is a conversion, you know, that's super high value for a bank.

31:18and any almost every business in big pharma where you know you're you're dialing in different content to different audiences right so if someone and by the way as controlled by geography so if you're going to tell a cancer patient you're going to give them different con content, then you are a cancer researcher, then you are a cancer doctor, then you are a pharmacist, right? You're going to have different content to each of those audiences, depending on who they are and what they're looking at. And so super high value use cases like that, um, that then the pharma company can actually kind of connect the dots on, okay, when I put this in and I started dialing in my content, we saw a script lift of, you know, 32 % or whatever in Spain when we started targeting content for this new oncology drug we're launching in Spain.

32:21Right. But that moment when the system is figuring out next steps, is that a human operator that's looking at the data and entering a prompt or is that a reasoning? That's a reasoning engine. And that's the point if it's just going to get faster and faster. Right. And, you know, obviously inference, everybody is working very hard to make inference faster. And inference is getting, you know, 30 times faster. The latest NVIDIA chips, the inference is 30 times faster. And that's not just optimizing at the chip, it's optimizing through networking and algorithms. and that's 30 times faster in one year now think of that compared to moore's law you know moore's law is uh what doubles every two years or something like that and and so the the exponential pace at which ai performance is improving is kind of hard for the human mind to comprehend And so inference, that's what I mean by it's pretty clear that we're going to go to where reasoning right now could take 30 seconds.

33:44Reasoning will take 0.3 seconds and you'll still get all the benefit of that reasoning. And there are companies, hardware-specific companies like Grok with a Q and Cerebrus, which is building a super large chip, that are basically optimizing just for inference. And then there's NVIDIA and others that are optimizing for training plus inference. but it's all just going to get faster and faster and faster where the end state is you get your personalized reasoned output as a consumer that is only for you. Yeah. I mean, we started talking about governance and governance. And you were saying how it's slowing some companies down and that's creating opportunities for others.