The Future of Marketing Briefing: Epsilon’s quiet bet against the LLM goldrush

“They’re going to get crushed,” he said. “Crushed by something that has been doing exactly this job, at exactly this speed, for over a decade.”

That something is Epsilon’s proprietary stack of roughly 15 machine learning models, embedded deep inside its ad server – a system the company built from scratch precisely so it could make real-time buying decisions at the level of individual people, not just sites or audiences. In the U.S. alone it tracks up to 270 million unique individuals and every bid decision is built up from that level. Most platforms do the opposite. They buy against broad contextual signals and layer targeting on top. Epsilon’s identity-centric approach meant it had to solve the AI problem differently. Every day, those models process 800 billion bid requests. For each one, they make approximately 15 AI-driven decisions. All of it happens in under 10 milliseconds. 

It is a performance no LLM could replicate. Not because the tech isn’t impressive but because it is the wrong kind of impressive. The “P” in GPT, Nowlan is fond of reminding people stands for pre-trained. These models are frozen in time, shaped by human language scraped from the internet, incapable of the rapid, continuous adaptation that a live market demands. Asking one to manage real-time bidding, he argued, is like asking a brilliant essayist to day-trade derivatives. The skills simply do not transfer. 

But here is where Epsilon’s argument gets more nuanced (and more interesting). The company is not anti-LLM – far from it. Loch Rose, chief analytics officer at Epsilon, noted that the industry is awash with announcements about LLMs doing exactly the kind of work that once kept human media managers glued to dashboards all day: adjusting bid frequency, swapping out underperforming sites, tweaking targeting parameters. That, he said, is actually  a perfect use case. Repetitive, contextual and describable in plain language – everything an LLM does well. Epsilon’s own workflows are proof of that.

Take its “index reports” for instance. They’re dense audience studies that compare consumer populations against broad national metrics, and were previously a drain on analyst hours. An LLM now drafts the summaries. When it got things wrong, Epsilon built a second model to check the first, cross-referencing outputs against the source documents before anything went out.

A human then reviews the lot. That basic structure – specialized machine learning models doing the heavy computational work at the base, LLMs handling the language-driven tasks above and a verification layer across both – is how the whole stack operates. No single model runs the show. Each does what it is actually good at. 

This, Rose said, is where the industry is quietly headed, whether it admits it or not. LLMs handle the human-facing, language-heavy strategy layer. Specialized machine learning models handle the real-time, high-frequency decisioning. And humans stay in the loop wherever the cost of a mistake is too high to accept. 

There is also a less technically sharp warning buried in this outlook.

If you hand your marketing strategy to a general purpose LLM then you are, by definition, handing it to the same model as every one of your competitors. Pre-trained systems give pre-trained answers. Differentiation – the thing marketers are actually paid to create – cannot be outsourced to something that treats every prompt the same. 

Numbers to know

35%: Percentage of platform-agnostic user-generated content that is part of influencer marketing campaigns in 2025, surpassing TikTok’s 21%.

40%: Percentage of listening time that podcasts accounted for this year, surpassing spoken-word radio at 39% 

108 billion: Total mobile app time spent on social media in 2025

49%: Percentage of marketers that believe when it comes to AI, data security risks will be a significant or critical concern in the next year or so.

What we’re reading

OpenAI Held Early Talks With The Trade Desk to Sell Ads

Following the announcement that OpenAI has partnered with Criteo, the platform has held talks with The Trade Desk for a similar deal, according to The Information.

Government Agencies Raise Alarm About Use of Elon Musk’s Grok Chatbot

The safety and reliability of Elon Musk’s AI tools has been questioned by officials at multiple federal agencies, including questioning whether his chatbot Grok’s looser controls present risks, according to Wall Street Journal.

Musk Says Judge’s Bias Forced Him to Pay Full Price for Twitter

During his testimony on Wednesday, Elon Musk told jurors that Delaware Chancery Court’s chief judge, Kathaleen St. J. McCormick’s bias against him forced him to pay the full $44 billion price for Twitter, but his lawyers warned he was unlikely to win the case, according to Bloomberg.

TikTok starts court battle to save China ties

TikTok’s battle in the U.S. might be over, but this past week the platform went head-to-head with the Irish Data Protection Commission to defend how it handles Europeans’ privacy and data, according to Politico.

What we’ve covered

‘The conversation has shifted’: The CFO moved upstream. Now agencies have to as well

VaynerMedia’s CEO Gary Vaynerchuk isn’t the only one seeing CFOs become part of the conversation much earlier – other agencies are seeing the shift too.

Yahoo pauses IAB membership amid a series of quiet cost-saving measures

Sources have revealed to Digiday that Yahoo has quietly withdrawn from several IAB boards globally, signalling a strategic shift away from formal industry governance in multiple regions.

‘Not a big part of the work’: Meta’s LLM bet has yet to touch its core ads business

Meta’s CFO Susan Li made the point that the team isn’t using LLM architecture to do ranking and recommendations work (yet) during the Morgan Stanley Technology, Media and Telecommunications conference on Wednesday.

WTF is Meta’s Manus tool?

Meta has recently added an AI agent tool called Manus to its Ads Manager – Digiday explains what it is, and whether or not it works.