812 | AI Agents Change EVERYTHING in 2026 (And Most Marketers Aren’t Ready) | Scott Brinker

Agenic AI is at least as big as the Internet. It’s moving faster, like, you know, riding a psychotic horse through a burning barn. That’s Scott Brinker, XVP of Platform Ecosystem at HubSpot, editor of chiefmarktech.com and the mind behind Marketing Technology Landscape that’s defined the industry for 14 years. Right now, he’s on the front lines of the next shift. Agenic AI, MCP-driven tool use, and first-mile edge enrichment where data becomes decision fuel the moment it’s created. In a world where buyers show up with their own agents, Scott’s here to tell us what changes, what doesn’t, and how to win. But what’s his play?

Fourteen years of mapping those thousands of logos. Not me, the tech is 10%. The advantage now is we have this ability to dramatically expand the scope of things that are being… In this episode, you’ll learn what are AI agents and the three kinds of agents changing marketing right now. You’re starting to see agents that the marketers don’t control, that it’s the buyers who control these. Why APIs and orchestration, not models, will define the future stack. Can I have like a super agent that acts as the orchestrator organizing these other agents in a network of agents? I can’t think of anything in the arc of human history that even comes close to this.

How first-mile data turns raw clicks into decision-ready signals. The engines themselves are effectively zero differentiation. All the competitive advantage is entirely a function of like, okay, what data do we feed into these things? And how do we leverage these things in a process that is unique to our business? You know, I’ve been saying that the agents now are a different type of animal. And so just to have a bit of a level set, map out what you mean, like how agents are fundamentally different from chatbots and from rules-based automations. Yes, 10 different people, what they mean by agent or agent. And you might get then slightly different answers.

It’s probably easiest to think of it as a continuum. Where on one end of the continuum, you have this like very deterministic rules-based automation, which is what we’ve kind of got pretty good at. We’re all pretty comfortable with that now. You know, on the other far extreme would be these AI agents that operate with a high degree of autonomy. They can take on pretty major, large-scale tasks, like you say. Oh, go find me five customers and research everything about them and, you know, prepare all the campaigns. Off it goes. All is a single thing. Along that continuum, there’s a lot of different steps.

We are not at that far end yet where you just turn things over to an AI agent and it does all your marketing for you. And probably a good thing for all the marketers in the firm. I think word weighs away from that. But you’ve got these sort of steps along the continuum where we’re like, okay, well, could I have an AI agent go out and do research on the web for a particular customer? And that research might happen in a very rules-based workflow of saying, oh, someone came in. They filled out this form. I saw this signal from the mouse. We are very deterministic things. And then I want to put a little bit of a very hyper-specialized agent in that workflow to say, okay, at that moment, you go off and you learn and discover about this customer. And that you’re actually giving it a certain amount of autonomy to decide, like, okay, well, where is it going to find that out? And when it finds things, what path does it follow to, you know, find more things? So I think of these three categories, three domains, sort of three areas where we’re seeing agents in marketing. The first are these agents that are helping marketers with their work, you know, the production of assets, you know,

being able to do analysis on how campaigns happen, what went well, that like customer research, well, many agent we were talking about, you know, so these are all, we see a lot of these tools, you know, that are emerging to like help marketers accelerate the things they do. There’s a second group of agents that are agents that are customer-facing. Now, the marketer, the company still controls them, you know, but they’re interacting with our prospects and customers. And the most obvious example is, of course, are these, you know, chat bots that we’ve had on our websites forever to help people answer questions or resolve customer service issues. And these LLMs that they now use, you know, are really good at these conversational interactions.

We’re getting better at hooking these things up to the right sort of data sources behind the scenes for like knowledge bases and ticket histories and things like that. So they’re actually getting pretty good. But you have other kinds of AI agents that are interacting with customers too, you know, in B2B, for instance, like these AI, SDR, BDR things that, you know, either engaging inbound or sometimes outbound emails. We’ll leave the value judgment aside on for that moment. But again, these are a set of agents that marketers are deploying to interact directly with the outside world. And the third category, which to me is the most interesting, is we are starting to see agents that the marketers don’t control, that it’s the buyers who control these. And probably the best example of this would be something like these deep research functions that are available in ChatGPT or, you know, Anthropics Clawed, where I can say like, you know, I am interested in evaluating, well, let’s say I’m interested in evaluating, you know, customer service chat bar solutions from my website, you know, and it can go off and it can like track down the information and

who are the vendors and where have I seen other comparisons for it? And what are people saying in, you know, social media or forums, you know, and organize is really nice report for me. That agent is acting on my behalf as the buyer. But as part of it, it is interacting with the websites, you know, other sorts of content resources. And so this is sort of a really exciting thing for marketers that we’ve always been used to dealing with a little bit of a computer agent out there as an intermediary. It was Google, right? The Googlebot was, you know, the one other agent that we’re like, yeah, I guess we’re doing this for humans, but kind of want to do it for the Googlebot too. I think we’re on the cusp here over the next year or so of just seeing easily a dozen of these other different kinds of agents operating on the customers, the buyers behalf that we’re going to have to interact with.

As agents sort of proliferate, I think, you know, it feels like there’s got to be some sort of center of gravity across the stack. Is that going to live in the CRM and the website, the messaging apps or in some other layer that connects everything and makes it talk together? The traditional Mar-Tech stack is not dead. If you’re finding value in this video, make sure to hit that subscribe button and don’t miss out on more tips coming up to help grow your agency. In fact, if anything right now in the state of AI, it’s more valuable than ever for exactly the reasons you say, you know, which is, hey, it’s great that we have these like AI agents or these HNN capabilities, but they work best in two scenarios. One, where we can point them at very clear authoritative data, you know, data that they can feed into what they’re doing, whether it’s decision-making or content personalization. You know, what you’re using is your system of record for that. You might not even just have one. You might have a CRM, you might have a dam, you’re using for certain content assets. You see a lot of larger companies, you know, have moved to putting a lot of that, you know, both customer data and company-wide

operational data into a cloud data warehouse, you know, and being able to use that as the source of truth that you can feed into agent work clothes and things like that, you know, but it’s very important, like you’ve got to have your systems of knowledge, your systems of truth, your systems of record that these things can work on. Data is still pretty much a mess. Either, you know, we still have a lot of data that’s just, like, siloed off in different places or even the data we have with them, particularly silo, it might not be very well, like, maintained hygiene of it, as I would say. And then even in cases where you have companies that are starting to connect data across the org through, like, using a cloud data warehouse layer or something like that, in theory, okay, this is great. We’re now sharing data across, you know, product and marketing and sales and operations.

And then you realize, wait, the way in which, you know, product refers to these things is under a completely different language than how the customer service team does and so these things. You know, so we’ve got a lot of work to do. Anyone, you know, can sign up and get open AIs, APIs or Anthropics or Google Gemini or whatever. The engines themselves are effectively zero differentiation for any company that’s using them. All of the differentiation, all of the competitive advantage is entirely a function of, like, okay, what data do we feed into these things? And how do we leverage these things in a process that is unique to our business? And there there’s, like, phenomenal opportunity for differentiation and competitive advantage. But if you don’t have your data in a good place where both it’s good clean data and you can pipe it to the right place at the right time, you’re just not going to be able to tap that power. The second thing is, you know, again, we were talking about that continuum between purely rules-based deterministic everything. And then Wild West, I’ve turned it all over to the agents. I don’t even have a job

anymore. You know, and where we’re seeing a lot of the practical use cases, they’re saying like, okay, I’m using my marketing automation platform or customer engagement platform or whatever it is that, you know, I’ve been using to set up these structured workflows, you know, maybe it’s an iPass-like tool or workflow automation tool. But inside those well-structured workflows and repeatable processes, I’m now embedding pieces of more agentic AI to do certain tasks.

Essentially, then, that workflow automation engine is becoming your orchestrator. It’s sort of keeping the pieces together. You know, now there’s a emerging capability that people talk about of like, well, wait a second, can I have like a super agent that acts as the orchestrator organizing these other agents in a network of agents and like Google has that A to A agent to agent protocol, but you know, as we may get there right now, most of that stuff is still basically a science fair of state. So not quite prime time. Let me ask you about like APIs. APIs are going to be like super important for the future of the marketing stack. Integration seems to be like almost a deeply religious thing for me. You know, for all the years of maintaining that marketing technology landscape, you know, I mean, the fundamental challenge people was like, hey, it’s great, you have all these pieces of innovation happening all different ways all over the place. But you know, to actually run a business, I gotta have this stuff work together and it doesn’t work together very well. I mean, AI feels like it just even like raises the stakes even more that, you

know, we have this ability to dramatically expand the scope of things that are being automated or run programmatically, you know, but at the end of the day, it’s got to be able to integrate with our data sources. It’s got to be able to integrate with the services we run to actually execute things like, you know, enroll someone in the subscription list or, you know, processes transaction. I am a believer that the best way to do that is through APIs or this model context protocol, which is kind of like a layman’s terms would be it’s almost like a standard by which these AI assistants and AI agents can do an integration that is very API like.

I think that’s that’s best because, you know, I mean, it’s very clear what you do, what you expect, it’s easy to sort of maintain, you know, those things of all. Now, there is another school of thought. There is a school of thought that, hey, actually, these AI agents that run in the browser, you know, they can just go and interact with the company or their website or their, you know, web app through the user interface that a human would have done and they simply simulate the process of being a human doing that. You see this now with some of these, you know, AI browsers that are out there, like the perplexity one comment, you know, what was that the open AI when they did that operator, you know, Asian thing early in the year, these are things that basically like, no, no, we don’t need to worry about APIs or anything that will just interact with the web like a human would. I think it’s really interesting in the sense that the upside of that is it does let these agents interact with things, even if they don’t have an API, even if they don’t have a, you know, defined MCP. The downside to that is even as a human user of these things, you know how confusing it can be when you’re

used to using an app in particular way and then the app developer like changes things in the menus or, you know, I mean, it’s just like the AI agents are going around the same. It’s just not as, in my opinion, it’s just not as reliable as having like, hey, listen, we’re going to well-defined APIs, well-defined MCP servers explicitly for those AI agents and AI systems to use. I just think that’s a more stable approach. There’s some agents that are really good at doing like 90 to 95% of the things that they’re supposed to do. But, you know, they’re saying about trust, right? It didn’t take a long time to build trust in just a moment to break it. Sometimes that 5% that it can’t do can break trust. So let me ask you what, like, what are the guardrails that you think should be in place before some of these agents can do, you know, really important types of tasks? As human beings, generally, if we’re getting 90 or 95% of things right, actually a pretty good, you know, so we have a certain set of expectations and goes back to that thing. We’re so used to computers and software being, you know, those rules-based

deterministic, it’s always going to work this exact same way. You put in X, you get Y, you know, so moving to this world where, okay, actually, you know, you put in X and you might get Y, 90% of the time, you might get Z, 5% of the time. We’re just not used to even thinking that makes a lot of people very uncomfortable. But, being said, I think it’s important to understand like, okay, well, what’s first of all the range of those outcomes? Like, for instance, if we’re talking about personalizing a message, you know, writing an email message, what does it mean to be 90% right? Or what’s the case where it gets wrong? Like, is there a question of how well the message is written? Did it like pull in the right information?

Was it making up like, to decide, hey, actually, I’ll give you this product for, you know, 80% off? Wait, wait, wait, do you know how loud you don’t have authority to do that? You know, and so a lot of these things, to be honest, the range in which something might be, it’s not that it’s necessarily going to go wrong. It’s just that, you know, 80% of these were really good 15% were so so. Oh, that last 5%. Yeah, and that wasn’t very great at all. But like, you have to ask yourself, like, in certain contexts, like, you know, I mean, if it’s summarizing a ticket history, or is it trying to just personalize an email a bit?

But that being said, your original question of like, okay, well, how do we put the guardrails in place? Well, I mean, again, this has a lot to do with what is currently now called context engineering, making sure like both the instructions that we’re giving to a particular AI agent or task, very well defined, we control the exact data that we’re feeding into it. We then might have things on the other side that sort of monitoring, you know, what was done allowing me to go back to audit it. You know, there’s like a series of this whole thing around like AI observability and AI governance. It’s maturing. But I think, yeah, companies need to get those systems in place ultimately before they’re going to feel really comfortable scaling up more and more of these agent work flows. Yeah, I mean, Ethan Malick, you know, who is like probably one of the best writers of like trying to just keep up with, you know, what’s what’s the state of what’s happening here in AI, you know, has referred to this as the jagged frontier, which is, you know, the puzzling thing about so many of these AI assistants and these agents is it can be ridiculously good at something that you would think would be like

very hard. And you’re like, Oh my goodness, that’s pretty amazing. I was just using this the other night to me learn calculus to help my daughter on a calculus class. And oh my God, I’m asking it to do things. And it’s just it’s phenomenal. And then there’s other things you’re like, well, surely if it can do that really hard thing, then this really simple thing should be a piece of cake and you give it to it and totally crafts the bed. I mean, this is one of the things that just makes it very difficult to be able to say, okay, like, can we, you know, is it what is AI going to replace what should, you know, humans do?

I think for now, the more discrete the task is, you know, that’s the metaphor people say, like, you know, treat it like an intern, very bright, but entirely inexperienced intern, those sorts of things that you would tend to turn over to it. Yeah, in a lot of cases, it is pretty good at those where I think you’d be very careful is at the level above that, you know, that the interns aren’t the ones actually setting the strategy. The interns aren’t the ones who are actually making sure all the operational pieces, you know, fit together the right way. The interns aren’t the ones who are actually evaluating to make sure this is working the way it is. You know, I think those are all things that, yeah, you need human leadership, human judgment, human insight on, and actually I think you’re going to need that for quite some time. How do you see orchestration working? Again, when the tools are spinning up their own agents and you’re trying to manage workflows across all of these, is that being managed inside of applications like CRMs, like Vendastas and other CRMs that are out there, is there another orchestration layer that’s needed? It’s early, probably the most mature right now.

Are the existing platforms, you know, like CRMs or, you know, very smart thing, automation or engagement platforms, is even some of these like, again, like the workflow, automation products, you know, the eye passes that they’ve sort of evolved to embrace, you know, AI and the agents. These are where most people are doing the orchestration because, again, the way they’re doing the orchestration is they’re still sort of keeping the scaffolding of a very deterministic workflow. And a lot of the pieces tend to be deterministic and reliable and like, this is going to work the way I expected to work. And then just inserting like, you know, very discrete oriented AI led tasks where you could actually turn to the AI to do something that was very hard for us to do before. For instance, like the very common case of this is how we leverage unstructured data, you know, it was always very hard for us to like programmatically in these automations, like interpret, you know, emails or call recordings or things like that. I mean, not impossible, but so difficult and expensive is to be unusable by most companies. Now, boy, it’s like it’s a piece of cake. You just feed that call

transcript or that email, you know, into the LLM and say, Hey, listen, I need to pull out, you know, what are the next three action items you think we need to do to do that comes out with those action items. And then the execution of those action items again, once again falls into that very deterministic, you know, workflow process. So it’s providing that orchestration. I think the question is, will that change dramatically? Like, will we get to the place where, oh, I spin up a master agent, and it’s coming up with all these workflows and orchestrating agents at that time? Maybe. That’s the sort of like truly disruptive version of what is the MarTech stack? What is the companies like tech operations going to look like? We’re just not there today. And I don’t think we’re going to be there. You know, I can get in so much trouble to try and make these predictions in such a fast changing world. But I don’t even think we’re going to see that over the next six to 12 months. I think this sort of combination of well structured workflows with then AI capabilities sort of embedded in constrained ways within them,

it’s just the safest and most reliant, like the highest ROI, like the risk-reward ratio. That’s what I’m thinking of, like relatively low risk, pretty high reward. Let’s get that working before we leap off into space. What’s going to be the difference between organizations that just add on AI to what they’re doing versus ones that are truly native AI companies? Yeah, I mean, probably depends a lot on the business. You know, if I am a restaurant, if I’m a chef, I might be leveraging AI and various things around, whether it’s my marketing or my supply chain or things like that. But the core of my business is probably, I don’t know, I might get myself in trouble for saying this one too, but it feels like one of those things like, well, it’s probably going to look about the same 10 years from now that it looks now. If I’m a software company or like a company that provides a digital service okay, that’s a whole other thing because, yeah, I mean, like the whole way in which, you know, the service is created and offered and how people consume it, like if they start consuming it through their own, I mean, yeah, that whole world, you know, feels like it could change very

rapidly. So I don’t know, I think it is okay for most businesses that right now we’re in the AI bolt-on stage because we’re still learning, you know, we’re still learning like, how does this stuff work? What is it good at? Was it not good at? Where are the things that we’re missing, right? Like the data realization, you’re like, okay, well, I’m glad I’m figuring out my data problems. Well, this is still just an add-on, you know, like before I turn my entire operations over to it. But these are steps along the path where eventually we will become AI native to the point that even the phrase AI native probably won’t be much, just like we no longer say digital marketing. Yeah, like, yeah, you know, just marketing. But yeah, we’re in the horseless carriage stage right now.