AI Can Write the Media Plan But It Still Can’t Pull the Trigger on the Spend

Read Time: 6 minutes

Inside ad tech’s quiet decision to keep large language models on the safe side of the budget line.

Large language models—the ChatGPT class of artificial intelligence—have become ubiquitous in modern office life. They draft emails, summarize decks, generate campaign concepts, translate meeting notes into next steps, and turn messy spreadsheets into coherent narratives.

They are welcome, in other words, almost everywhere.

Almost everywhere except the one place that matters most in advertising: the moment where real money gets spent.

That sounds contrarian in an era when “agentic AI” is marketed as inevitable—software that doesn’t just assist humans, but acts for them. Planning, buying and optimization, we’re told, will soon run autonomously. A brief goes in. A campaign comes out. The machine does the rest.

In practice, the switch remains conspicuously off.

For now, large language models are being used as accelerants, not decision makers. They compress workflows. They shrink timelines. They surface insights faster. They make the work feel lighter.

They do not spend the ad dollars—and that isn’t an accident.

At the Consumer Electronics Show in Las Vegas last week, amid the usual fog of AI enthusiasm, one ad-tech executive offered a reality check that landed precisely because it wasn’t dressed up as a moonshot. His company’s trading automation platform, he said, can take a two-hour workflow and reduce it to ten minutes, with near-zero errors.

Then came the boundary.

“At QuantumPath, we want to automate the workflow, not the buying decisions,” said Chief Executive Jeffrey Hirsch.

It wasn’t a lonely stance. It was the prevailing one.

Across agencies, platforms and infrastructure providers, an invisible line has formed between automation that helps humans move faster and automation that replaces humans at the point of spend. Some of this is institutional self-preservation—turkeys rarely vote for Thanksgiving. But the resistance runs deeper than job protection. It’s rooted in how auctions work, what measurement can actually prove, and the unresolved question of who carries accountability when machines mishandle inventory.

The result is a paradox that defines the current moment: the industry is investing heavily in AI, while simultaneously refusing to let AI drive the one decision that turns strategy into cash flow—how much to bid, on what, and when.

Why the most powerful AI is being kept at arm’s length

Start with the technical mismatch. Large language models operate in open-ended semantic space. They generate responses probabilistically. They are brilliant at pattern recognition across human language.

Programmatic advertising, by contrast, is a domain of split-second decisions, hard constraints, and a relentless demand for determinism. In the real-time bidding environment, the system must evaluate an opportunity, decide whether it matches a campaign goal, assign value, and bid—all in milliseconds.

That means the core of ad buying still belongs to narrow models designed for one job: bidding logic that can be measured, tuned, and repeated consistently.

LLMs excel at narrative. Bidding requires math and discipline.

That’s why, in practice, large language models are being deployed around the edges of the transaction—planning, setup, reporting, analysis—rather than at the crux of it. They help humans get to the decision faster. They just don’t make the decision.

Yet few in ad tech are treating this boundary as permanent.

Executives who live inside programmatic infrastructure don’t describe decision-level autonomy as impossible. They describe it as not ready. Or not economical. Or not worth the risk—yet.

Michael Richardson, vice president of product at Index Exchange, framed the direction plainly: more advanced autonomy will move out of experimentation as computing gets cheaper and infrastructure matures. But, he cautioned, it’s “not going to be broadly deployed” yet—held back by cost, organizational readiness, and still-fuzzy use cases.

In other words: the models will likely get there.

The harder question is no longer whether LLMs can spend ad dollars autonomously. It’s whether they should.

The data problem: AI can’t rise above what it’s fed

The most forceful arguments against autonomous buying aren’t philosophical. They’re empirical.

“That’s the big concern for me: unreliable inputs produce unreliable decisions,” said Tom Swierczewski, vice president of media investment at Goodway Group. “For LLMs to buy autonomously in programmatic media, they’d need bidstream data—and that data is deeply flawed.”

His critique cuts to the heart of advertising’s modern operating system. The industry still runs on measurement that is, at best, uneven—and at worst, structurally biased.

Last-click attribution remains overvalued. Walled gardens are difficult to audit. Platform-reported metrics can be opaque. Incrementality is discussed frequently and implemented inconsistently. Cross-channel truth is promised, but rarely delivered cleanly.

If you train—or simply guide—autonomous systems on distorted inputs, the system does not become intelligent. It becomes efficient at repeating the distortion. It scales the blind spots.

And because these systems learn continuously, the distortions don’t just persist. They can become self-reinforcing.

In a world like that, “autonomous optimization” can quietly become a machine that grows very confident while getting very good at the wrong thing.

“The industry needs AI to manage complexity and move faster,” said Paul Boruta, chief executive and founder of ad-tech platform Slingwave. “But it should not hand that intelligence to systems that are optimizing toward the wrong signal.”

The phrase “wrong signal” matters. Advertising doesn’t suffer from too little information. It suffers from too many metrics pretending to be truth. The machine can optimize forever—if the target is flawed, the optimization is just a faster route to mediocrity.

Why the industry is investing in plumbing, not autopilots

That’s why much of today’s LLM investment is not going into full autonomy, but into infrastructure that makes autonomy possible later.

It’s a quieter transformation: modernizing platforms, containerizing auctions, opening APIs, lowering the friction of switching costs, cleaning up data paths, improving workflow orchestration.

The goal is not to replace human decision-making overnight. It’s to rebuild the factory so that, someday, the assembly line could run itself.

For now, though, companies are deliberately keeping the bidder grounded in the same narrow, rule-bound machine learning that clears the market today.

Yahoo DSP is a case in point. The company is welcoming LLMs into its orchestration layers and user interfaces—places where language models shine. They can power dashboards, generate recommendations, and streamline workflows.

But the core bidder remains rooted in deterministic logic.

“Nothing that we’re doing at the moment would suggest that agentic or an LLM will take the place of bidding logic,” said Adam Roodman, general manager of Yahoo DSP. “I mean there could be parts of it eventually but at its core it will still be machine learning.”

This distinction is more important than it sounds.

In the marketing copy, “AI” often gets presented as a single, unified brain. In reality, advertising technology is becoming a layered system: language models up top to translate human intent into structured instructions, and specialized models beneath them to execute transactions predictably.

The LLM becomes the front office. The bidder stays the factory.

Even “agentic campaigns” still have human hands on the wheel

Even the most bullish builders are starting to recalibrate how they describe what their systems actually do, versus what the branding implies.

PubMatic’s work with independent agency Butler/Till provides a telling example. The effort has been framed as an end-to-end “agentic” campaign. Directionally, the term fits. But operationally, the reality is more nuanced—and more cautious.

Butler/Till used an agent built on Claude to translate a human-written brief into a structured media plan. Like ChatGPT, Claude is increasingly used to draft strategy, generate concepts, and accelerate the front end of planning.

That plan was then passed into PubMatic, where its own AI systems mapped the intent to inventory, channels and audience segments inside the platform.

But the final parameters were reviewed and approved by humans at Butler/Till before launch.

“We’re intentionally being cautious on what we’re directly and entirely attributing to agentic systems at this stage,” said Nishant Khatri, executive vice president of product management at PubMatic, in an email. “As the campaign continues, we expect greater clarity into efficiency and performance trends. Directionally, these results align with what we would expect from an early agentic campaign operating at a national scale.”

This is what LLM adoption in digital advertising really looks like today: not a robot running campaigns while humans go golfing, but a machine that makes the human faster—and a human who still has to sign their name at the bottom of the page.

The real fight isn’t about intelligence. It’s about control.

The transformation underway is quieter than the hype suggests, but no less consequential.

It’s labor compression. It’s fewer hours to build a plan. It’s fewer people needed to pull reporting. It’s faster iteration across creative variants. It’s a rewiring of infrastructure and a slow shift in power across the ad stack.

In plain terms, the industry isn’t waiting for smarter machines.

It is deciding who controls the machine that controls the money.

Because the moment you let an autonomous system place bids, you’re not just making an operational change—you’re moving agency, accountability, and leverage. You’re deciding who owns performance. You’re deciding who gets blamed when it goes wrong. You’re deciding which players become utilities and which become decision engines.

Until that fight is settled, large language models can draft plans, build workflows and run dashboards.

They just won’t be handed the keys.

Source: https://digiday.com/marketing/