Something bothers me about much of the conversation surrounding AI in martech: there is a lot of talk about content generation, large-scale personalization and “copilots” that help teams save time. You connect an agent to a CRM, give it access to some data, ask it to draft a few emails and then call it a “revolution”.
Personally, I find that there are more compelling areas to explore elsewhere. What fascinates me about “agentic marketing” isn’t a model’s ability to generate a campaign in thirty seconds but it’s the potential to build a system that observes a marketing environment, formulates hypotheses, makes decisions, executes actions across various tools, measures the outcomes and adjusts its strategy.. all without any human intervention.
Furthermore, the real intrigue lies in understanding what happens when several decisions, each reasonable and logical on its own, begin to interact with one another.
Let me try to explain my somewhat chaotic line of thought: an agent drafting an email, that’s fairly straightforward to understand and anticipate. It’s almost as predictable as a standard computer program. But what about an agent that decides a customer shouldn’t receive any emails at all, because it anticipates that such an interaction might diminish the customer’s long-term value? Aha! See what I mean? Now, that’s much more interesting!
Wait, hold on! It’s not over yet!
Now, what’s even more fascinating is a system that has to strike a balance between a given decision, revenue targets, a company’s business strategies, commercial pressure constraints, product availability and, above all, the simultaneous actions of other agents, each more unpredictable than the last.
Your head’s spinning a bit, isn’t it? Come on, admit it, you’ve never thought about AI in this way before.
I can see the look on your face through the screen: “She said ‘unpredictable’, but an AI isn’t unpredictable!”
Sorry, but it actually is. No matter how hard you try to control an AI, you can never completely eliminate its hallucinations. Do you know why? Because an LLM is a language model based on mathematical probabilities. By definition, probabilities are never certain, unless you reach 100%, which is impossible when it comes to AI.
Personally, back in 2023, I even managed to convince an AI that 1+2 = 1. Okay, it was still the early days of the AI era, but I still pulled it off!
1- The real change isn’t generation, it’s the shifting of decision-making power.
I like to distinguish between three levels in the evolution of AI-augmented marketing:
- AI that produces: You provide a brief and it generates content. You remain the one who thinks, makes judgments and decides. (Sorry to break it to you, but if this is your situation, you should start looking for a different profession because yours will disappear in less than three years.)
- AI that recommends: It analyzes performance, detects anomalies, compares various options, and suggests a course of action. You remain the executor. (Sorry to say, but you, too, will disappear.. along with all other execution-based marketing roles.)
- AI that acts: You entrust it with an objective, a scope for decision-making and a set of constraints; it then selects specific actions, executes them, and adapts its behavior based on the results. (Well, you have a bit more time, but honestly.. you’re managing an AI that’s smarter than you. Just think about what happens to managers who are less capable than the people they manage… now imagine managing AIs!)
To be honest, I have no idea..I’m just being provocative; I don’t know whether you’ll keep your job or not. But that’s beside the point. Let’s look instead at the third level.
I believe we gloss over a fundamental question: at what point does a marketing system stop being merely an execution tool and become a true decision-making system?
Take a deliberately mundane scenario: an e-commerce brand notices a drop in conversion rates for a specific product category.
- A conventional system might trigger a rule: if conversion drops, increase promotional pressure.
- A predictive system might estimate which customers are most likely to make a purchase.
- An agent, in theory, can go much further. It can seek to understand whether the decline stems from a pricing issue, a shift in traffic sources, an out-of-stock situation, a degraded user experience, or a change in demand. It can then weigh various hypotheses and prepare different courses of action based on its findings.
But there is one detail that changes everything: the agent might also conclude that the best course of action is to do nothing at all.
Maybe the decline is temporary. Maybe users acquired through a recent campaign simply have a lower purchase intent. Maybe triggering a discount would boost conversion in the short term, while simultaneously eroding margins on sales that would have happened regardless.
I find this concept far more compelling than an AI that merely optimizes a conversion rate in a mechanical way.
Agentic marketing truly becomes strategic when it’s no longer tasked simply with boosting a specific metric, but rather with choosing between competing, potentially contradictory outcomes.
And this capability requires something that many projects underestimate: an extremely precise definition of what the company is actually trying to optimize. It’s great for companies, sure! I’m not saying otherwise! Deep down, I’m not bitter. But my god, what a shame it is for marketers!
2- A Fascinating topic: That agent that chooses not to send a campaign
I truly believe that one of the most underrated use cases for “agentic marketing” involves the decision to intentionally forgo a marketing action!
It’s quite remarkable when you think about it: it’s as if you were transforming your marketer into something akin to a prophet.
For years, we’ve built systems designed to determine whom to contact, when to reach out and what message to send.
We’ve spent time optimizing segments, engagement scores, automated workflows and send windows.
Yet, as “human” marketers, we’ve largely overlooked a fundamental question: “When should we decide not to contact someone?”
Let me be clear: when I say “prophet,” I’m not referring to magic or divination of any kind. Rather, I’m talking about an agent’s extraordinary ability to synthesize all available signals (interaction history, frequency of engagement, recent purchases, unsubscribes, complaints, and the performance of past campaigns) to achieve the goal of maximizing long-term customer value.
For instance, an agent might decide that a customer who is typically very responsive to promotional emails should not receive another offer today.
Why?
Because the customer has already made a recent purchase, another communication is scheduled for tomorrow and the likelihood of securing an additional conversion may not justify the risk of “marketing fatigue.”
A human marketer is capable of this.. it’s not extraordinary in itself. But a human can only manage this for, say, 1, 2, 3 or even 100 customers over the span of three or four months. They certainly cannot do it simultaneously and automatically.
That’s why I used the term “prophet”!
To be precise, however: an LLM does not spontaneously possess a reliable understanding of customer value. It must be provided with the right data, estimation models, explicit rules and a method for evaluation.
I would therefore describe the agent less as an omniscient brain and more as an orchestrator : one capable of leveraging various systems to make a decision rooted in context. It’s an incredible, sudden development.. so unexpected that it has caught many marketers off guard, leaving some struggling to keep up.
In fact, they are struggling so much that they fail to realize the true role of a marketer today: empowering the agent to compare the potential outcomes of different actions.
Send now. Wait. Switch channels. Offer assistance rather than promotion. Or do nothing at all. From this perspective, the absence of communication becomes a marketing decision in its own right.
I find this to be a rather profound shift in philosophy: moving from a logic of constant activation to one of justified intervention.
Furthermore, from the standpoint of a marketer’s actual work, the role shifts.. moving away from that of an executor, strategist, or manager.. toward that of a “prompter,” where the only remaining sign of one’s marketing identity is the company badge received upon hiring, which displays one’s job title.
3- Unpredictable… yet entrusted with the power to make autonomous decisions! The new marketing contradiction:
This may well be the most underestimated contradiction in the future of marketing: if AI is by nature unpredictable and prone to “hallucinating”, then how can we truly grant it the power to make autonomous decisions?
Let’s be logical! We can’t, on one hand, acknowledge that its reasoning can be flawed and its outputs merely probabilistic, while on the other hand delegate to it decisions that directly impact a company’s very survival!
This is precisely where the issues of reliability and governance become central:
- Who defines the limits of an agent’s autonomy?
- Who verifies that its choices actually serve the company’s long-term interests, rather than simply satisfying a poorly defined objective or acting on biased data?
- And, above all, who bears the responsibility when a series of individually coherent decisions leads to a disastrous outcome for the entire system?
The true revolution doesn’t really lie in letting AI make decisions for us, but in our ability to determine which decisions can reasonably be entrusted to it, under what conditions and with what safeguards.
And between us, this very contradiction might just be what saves the role of the marketing manager (yes, I stand by my claim that those who merely execute tasks will lose their jobs.. sorry folks!). The real challenge of “agentic marketing” may not be building AIs capable of deciding on their own, but rather designing systems where their decisions can be evaluated, challenged, corrected and (if necessary) halted.
Because we, humans, will always need another human to vent our rage upon when the ship starts to sink.
Now, for my final word: I must admit that an LLM.. probabilistic as it may be.. can actually be highly reliable when designed correctly and properly, with the right safeguards in place (such as validation, thresholds and auditing). However, one needs to be aware of this.. and I’m not sure many companies fully grasp the importance of this point.
