AGENTIC AI IS COMING FOR MEDIA BUYING — HERE'S WHY HUMAN COMPETITIVE INTELLIGENCE STILL WINS

The buzz is deafening, and it isn't baseless. Across the advertising industry, major players are placing serious bets on agentic AI — software systems designed not just to assist human media buyers but to act on their behalf, making real-time decisions about where budgets go, which creatives run, and how campaigns optimize in flight. Stagwell has been building proprietary targeting systems around the concept. Salesforce has folded agentic capabilities into its marketing cloud. And the IAB Tech Lab, the body responsible for the technical plumbing of digital advertising, has been actively "agentifying" its standards and developing frameworks to help AI agents transact programmatically. The trajectory is unmistakable: the infrastructure for autonomous media buying is being laid, brick by brick, by the most influential organizations in the ecosystem.

But here's what the conference keynotes and breathless press releases tend to leave out — the technology is dramatically far from the autonomous vision being sold. At AdExchanger's Programmatic AI event in Las Vegas, Emily Lai, WPP Media's group director of media optimization, told attendees that somewhere between 80% and 90% of campaigns are simply too complex for AI to handle on its own. Simple programmatic guaranteed buys with minimal targeting? Maybe. But the average media plan — layered with audience segments, frequency caps, sequential messaging, cross-channel coordination, and compliance requirements — remains well beyond what any agent can manage solo.

That assessment wasn't a fringe opinion from a skeptic. It was echoed by the very institution building the rails for AI-powered transactions. IAB Tech Lab CEO Tony Katsur, after delivering a session on the evolution from machine learning to agentic AI, conceded that "we're not there yet" — and that while simple tasks are within reach, there's still an enormous amount of work ahead. Oleg Korenfeld, CTO of WPP-owned CMI Media Group, went further, calling the idea of agents communicating with each other at scale across all campaigns "just not realistic," particularly in highly regulated categories like healthcare.

The forward-looking language from vendors and analysts doesn't dispute this gap so much as quietly acknowledge it. MarTech's recent analysis of AI-native advertising describes the next phase as agentic AI systems that can make decisions autonomously, with self-optimizing agents that "experiment continuously, reallocating budget, adjusting targeting, and refining creative without human intervention." Read carefully, though, and the tense gives the game away — this is aspirational language describing a capability still under construction, not a product you can buy and deploy on Monday morning.

Meanwhile, the market's own behavior confirms the hesitation. A Digiday survey cited by AdExchanger found that only 32% of marketers use AI to actually buy ad placements, even as majorities embrace it for data analysis and content creation. The pattern is clear: the closer a task gets to real budget allocation and performance outcomes, the less willing marketers are to hand over control. They draw a sharp line between assistive AI — useful for grunt work — and autonomous decision-making, where the stakes demand human judgment.

None of this means agentic media buying is vaporware. The momentum is real, and dismissing it would be as foolish as overhyping it. But the industry narrative has lapped the technology by at least a full cycle, and that gap matters. If full automation isn't here yet — and won't be for the campaigns that matter most — then the urgent question isn't how to prepare for a robot-run future. It's what marketers should be sharpening right now, in the messy, hybrid present where human intelligence still determines who wins.

Every agentic media buying system shares a foundational dependency that its vendors rarely advertise: it learns from what already happened. The models ingest historical performance data — click-through rates, conversion logs, bid histories, auction outcomes — and optimize toward patterns that proved profitable in the past. That makes them extraordinarily good at refining what works today. It also makes them structurally incapable of identifying what will work tomorrow before the data says so. Emerging creative trends, novel audience angles, untested offers — these live in the pre-statistical wilderness where no model has enough signal to act. By the time an agentic system "discovers" a winning approach, it's already won somewhere else first.

Philip Inghelbrecht, the co-founder and CEO of Tatari, offers perhaps the sharpest framing of this problem. As he argues in his piece for AdExchanger, the industry is making a "category error" by conflating automated workflow with genuine intelligence. "You can automate a workflow in days or weeks," he writes. "You can't manufacture 10 years of outcome data." That distinction collapses the moment a platform slaps the label "agentic AI" on what is effectively automated budget pacing. Buyers hear "intelligence" and stop interrogating what the underlying model was actually trained on — whether its recommendations reflect deep outcome data or just recycled heuristics dressed in sophisticated language. Efficiency gains get mistaken for strategic insight, and nobody asks the uncomfortable question: if every advertiser on this platform is feeding the same auction signals into the same optimization engine, where exactly does differentiation come from?

This is the convergence trap, and it's baked into the architecture. When competing brands all use the same demand-side platforms, bid on the same inventory, and train their algorithms on the same pool of platform signals, optimization converges. Every system reaches roughly the same conclusions at roughly the same time. CPAs tighten across the board, but nobody pulls ahead because the machine is solving for the same objective function with the same inputs. The efficiency gains are real — and shared equally, which means they confer no competitive advantage at all.

The broader industry shift makes this even more consequential. As MarTech has reported, media buying is transitioning from impression-based models toward what some researchers call a "decision economy," where autonomous agents experiment continuously — reallocating budget, adjusting targeting, and refining creative without human intervention. But decisions still require inputs the AI doesn't generate on its own. Clearer positioning, sharper messaging frameworks, more distinctive brand narratives — these are the upstream strategic choices that determine whether an AI agent has anything meaningfully different to optimize in the first place. When execution is automated, differentiation doesn't come from faster bidding; it comes from stronger thinking before the bid is placed.

This is the dirty secret hiding in plain sight: agentic AI is a rear-view mirror with exceptional resolution. It can show you exactly where you've been, down to the pixel. What it cannot do is look around the corner. It cannot sense that a competitor is about to reposition, that a cultural moment is about to reshape demand, or that an entirely new creative angle could unlock an audience segment the historical data never imagined. Those insights require a different kind of intelligence — one that synthesizes incomplete signals, reads context, and makes judgment calls in the absence of statistical certainty. The competitive moat, in other words, cannot live inside the machine. It has to come from what you feed it.

If every AI media buying system is optimizing against the same rearview mirror, the question becomes: where does a genuine forward-looking edge come from? The answer isn't more automation. It's better reconnaissance.

The most valuable competitive signals in digital advertising don't surface in quarterly earnings calls or press releases. They hide inside the auction itself — embedded in media allocation decisions, efficiency trends, placement strategies, and channel shifts that only become visible when you know where to look. Consider the insurance vertical, where analysis has shown that a carrier like Progressive doesn't simply outspend rivals; it outbuys them, achieving lower customer acquisition costs through audience precision and diversified placement rather than sheer budget size. That's not a data point any automated bidding system will hand you. It's an insight that emerges from disciplined competitive intelligence — the kind of work that spy tools and ad intelligence platforms make possible at scale.

This is where human-led competitive research becomes the highest-leverage activity in performance marketing. Tools like Meta's Ad Library, competitive creative platforms, and auction intelligence dashboards let strategists observe what's actually winning in the market right now: new hooks gaining traction, emerging ad formats being tested, shifting offer structures, creative treatments that signal a competitor is pivoting strategy. When a rival suddenly floods a new placement type with fresh messaging, that's a strategic signal. When three competitors simultaneously test the same angle, that's a trend forming before any algorithm has accumulated enough conversion data to optimize against it. The marketer who spots it first has a structural timing advantage — a window in which they can develop their own response, test against it, and establish position before automated systems converge on the same territory.

As MarTech has argued, speed is becoming a competitive advantage in its own right. Brands that can test and adapt hundreds of creative variations quickly can respond to cultural moments, seasonal shifts, and competitive moves far faster than those locked into traditional production cycles. But speed without direction is just expensive chaos. The strategic value of competitive intelligence is that it tells you what to be fast about — which angles to test, which formats to invest in, which positioning gaps to exploit. When execution is increasingly automated, differentiation comes from stronger inputs: clearer positioning, sharper messaging frameworks, and more distinctive brand narratives. Spy tools are how you build those inputs from observed market reality rather than internal assumptions.

This reframes the entire relationship between human strategists and AI optimization. You're not competing with the algorithm — you're feeding it better starting inputs than your competitors can. An AI system initialized with a creative concept inspired by a competitive gap you identified three weeks before anyone else will outperform an identical system initialized with last quarter's best-performing ad. The machine does the optimization; the human does the seeing. And as Marketing Dive's research into agentic commerce has confirmed, even organizations at the frontier of automation overwhelmingly prefer models where AI surfaces recommendations and humans make the strategic calls. Only three percent of commerce media leaders favor environments where AI operates with a nearly free hand.

The implication is clear: the marketers who will thrive alongside agentic AI aren't the ones who delegate everything to it. They're the ones who show up each morning with intelligence the machine doesn't have yet — and use it to set the starting conditions that determine where optimization ends up.

The "automate everything" thesis sounds compelling until you examine what the people actually building agentic systems believe. They aren't building for full autonomy — and the data they're producing is the strongest possible case against the position.

Consider the most striking number in Koddi's research: only 3% of commerce media leaders favor environments where AI operates with an almost free hand and human involvement is reduced to oversight and compliance. Three percent. That's not a "the market is divided" finding. That's near-unanimity in the opposite direction of the prevailing hype. The overwhelming preference across the industry, as the study makes clear, is for AI to surface insights and propose actions — and for humans to make the final decision. Even on the consumer side, the pattern holds: shoppers are comfortable letting AI research, compare, and recommend, but they pull back sharply when it comes to completing transactions without their approval. Autonomy, it turns out, is a spectrum, and the market is clustering far from the full-automation end.

What matters more than the preference itself is the principle behind it. As Koddi's research concludes, the companies pulling ahead are not necessarily automating the most — they are building infrastructure that makes human-AI collaboration work reliably at scale. That's a fundamentally different design philosophy from the one Silicon Valley's loudest voices promote. It means the competitive differentiator isn't the sophistication of your automation; it's the quality of the human judgment your system is designed to amplify.

This aligns with a broader truth that Walker Smith, Kantar's Chief Knowledge Officer for Brand & Marketing, articulated recently: the nature of today's challenges is identical to the challenges of the past. Every theoretical concept, every practical application in marketing has been built for persuading humans — and as long as humans remain the audience, marketers have no reason to believe those frameworks have suddenly become obsolete. The real disruption, Smith argues, will come later, when AI agents become the target of persuasion rather than the tool. Until then, the fundamentals hold. And fundamentals require human understanding.

But here's where most human-in-the-loop conversations go wrong: they fixate on the approval layer. The media buyer reviewing a recommended bid adjustment and clicking "confirm" is technically a human in the loop, but that person isn't adding much differentiated value. The approval workflow is governance, not strategy. It prevents catastrophic errors. It does not create competitive advantage.

The highest-value human contribution sits upstream — in the competitive intelligence that shapes what the AI is optimizing toward in the first place. A strategist who detects that a rival is systematically shifting budget into a new channel, or that a category leader's CPMs are falling in ways that suggest a fundamentally different audience approach, is generating insight that no amount of bid-level automation can replicate. That pattern recognition changes the objective function itself. It tells the machine what to optimize for next, rather than simply refining how efficiently it pursues the current goal.

Campaign management should be automated. Competitive pattern recognition cannot be — not because the technology isn't advanced enough yet, but because the judgment involved is anticipatory, contextual, and adversarial in ways that backward-looking optimization models are structurally unable to address. The organizations that understand this distinction will build their human-AI collaboration around it. The ones that don't will keep perfecting yesterday's playbook at machine speed.