‘Intentionally being cautious’: Why the ad industry isn’t ready to let AI agents spend ad dollars

In practice, that switch remains conspicuously off. 

For now, LLMs are being used as accelerants, not decision makers. They compress workflows. They shrink timelines. They surface insights faster. They do not spend the ad dollars – and that is not an accident. 

At the Consumer Electronics Show in Las Vegas last week, one ad tech trading automation platform cut through the ambient AI enthusiasm with a blunt reality check. It can reduce two-hour workflows to 10 minutes with near-zero errors but stops short of actually spending ad dollars. “At QuantumPath, we want to automate the workflow, not the buying decisions,” said its CEO Jeffrey Hirsch. 

It was not a contrarian take. It was the prevailing one. 

Across agencies, platforms and infrastructure providers, an invisible line has been drawn between automation that helps humans move faster and automation that replaces humans at the point of spend. Granted, some of that is institutional self-preservation – turkeys, after all, rarely vote for Thanksgiving. 

But the resistance runs deeper – rooted in the mechanics of auctions, the limits of measurement and the unresolved question of who carries accountability when machines mishandle inventory. 

Let’s start with the most common: the tech. LLMs are designed to operate in open-ended semantic space and sampled probabilistically – a mismatch for programmatic auctions that demand fast, repeatable, deterministic logic. That keeps them parked at the edges of transactions – planning, setup, reporting and analysis – rather than at the crux of it.

That boundary, however, is not being treated as permanent.

Ad tech executives, who operate at the heart of programmatic advertising’s infrastructure, openly frame decision-level autonomy less as a technical limitation and more as an engineering timeline. They expect more advanced autonomy to move out of experimentation as computing becomes cheaper and infrastructure matures – even if, as Michael Richardson, vp of product at Index Exchange, put it, “it’s not going to be broadly deployed” yet because of cost, readiness and unresolved use cases.  

In other words, the models will get there. 

The harder question for ad execs is not whether LLMs can spend ad dollars autonomously. It is whether they should. 

“That’s the big concern for me: unreliable inputs produce unreliable decisions,” said Tom Swierczewski, VP 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 point being that advertising’s modern data foundation is still riddled with distortions: last-click bias, siloed walled gardens, platform-reported metrics that are difficult to audit and a persistent lack of incrementality adjustment continue to shape how performance is measured. Train autonomous systems primarily on those inputs and the system does not get smarter. It scales its blind spots. And because it learns continuously those distortions do not merely persist, they become self-reinforcing. 

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

Which is why so much of today’s LLM investment is focused on plumbing, not pilots. Platforms are modernizing infrastructure, containerizing auctions, opening APIs and lowering tech switching costs – all prerequisites for autonomy – while deliberately keeping the bidder itself grounded in the same narrow, rule-bound machine learning that clears the market today.

Yahoo DSP, for instance, is welcoming LLMs into the orchestration layers and interfaces of its platform while keeping the core bidder rooted in deterministic bidding logic. LLMs may drive dashboards and workflows. They are not being positioned as the engine that decides what to buy, when to buy it or how much to pay. 

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

Even the most bullish LLM builders are now recalibrating how they describe what their systems actually do vs. what their branding suggests. PubMatic’s work with independent agency Butler/Till is a case in point. 

The companies positioned the effort as an end-to-end “agentic” campaign. Directionally, that framing holds. Operationally, it flattens important nuances.

Buter/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 generate concepts and draft strategy. The resulting plan was passed to PubMatic, whose own AI systems mapped the intent to inventory, channels and audience segments within its platform. Final parameters were reviewed and approved by Butler/Till staff before launch. 

“We’re intentionally being cautious on what we’re directly and entirely attributing to agentic systems at this stage,” said Nishant Khatri, evp 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 LLMs in digital advertising really look like. 

The transformation underway is quieter – labor compression, infrastructure rewiring, slow shifts in power across the ad stack. Simply put, the industry isn’t waiting for smarter machines. It is deciding who controls the machine that controls the money. Until that fight is settled, LLMs can draft plans, build workflows and run dashboards.

They just won’t be handed the keys.