Ad Agencies Hit Token Limits as AI Costs Outpace Proven Value

Advertising agencies are discovering that unlimited AI experimentation comes with a very real bill. As generative AI moves from pilot projects into daily operations, marketing firms are scrambling to control compute costs while struggling to prove the technology delivers measurable business value.

Laura Higgins, chief brand and innovation officer at Dollar Shave Club, ran through her first month on the job using Claude, ChatGPT, Higgsfield and Gemini without tracking expenses. Within weeks, she built a triage system: cheaper models for routine tasks, premium tools only when necessary. The shift from unmetered exploration to token discipline took about 30 days.

That personal reckoning is now playing out across the industry at scale.

Why it matters

The advertising industry adopted AI faster than it built systems to measure effectiveness or manage costs. Agencies are now caught between client expectations of lower fees and the reality of rising infrastructure expenses, forcing them to price outcomes they cannot yet quantify. Without agreed-upon metrics for AI-generated work, the industry risks repeating the waste and opacity that plagued early programmatic advertising.

Token caps emerge as usage governance

Agency PMG rolled out its Alli For You platform company-wide last month, pooling staff access to major language models under a $50-per-day token cap per user. The limit followed months of testing that began in January, when early users operated on what executives called "free token range."

PMG says it rarely hits the ceiling, but the cap exists for peak demand scenarios. "If it's Black Friday and you're launching ads, you're doing reporting, you're leveraging an agent, we want to make sure we have enough tokens to handle all that," said Kaitlin McGrew, PMG's head of search engine marketing.

The challenge extends beyond simple cost control. Tracking token spend is straightforward; determining whether that spend delivered results is not. Caroline Giegerich, vice president of AI and marketing innovation at the IAB, noted the industry started with time-saving metrics but is now trying to measure actual business impact. The problem: there's no standard way to tell if heavy token usage signals effectiveness or inefficiency.

Pricing models fracture without measurement standards

Agencies have adopted conflicting approaches to billing for AI work. Dept refuses to pass token costs to clients, arguing itemization reduces the value of human expertise to a compute metric. S4 Capital's Monks builds tokens directly into subscription pricing. Major holding companies are folding AI expenses into broader principal media deals.

All three strategies attempt to work around the same missing piece: a way to define and measure what AI actually delivered.

Cheil Agency Network CEO Joe Maglio is moving all agencies toward output-based pricing, with half of existing clients already transitioned. The shift isn't driven by solved measurement problems but by competitive pressure, as AI capability has become a selling point regardless of proven ROI.

Publicis CFO Loris Nold told investors last week that productivity benefits from AI tools offset costs, pointing to 17 basis points of margin improvement in the first half of the year. However, more than 30 basis points of savings had already been reinvested in AI tools and training before reaching the bottom line.

The honeymoon ends

Off-the-record comments at Cannes revealed growing skepticism. One executive reported three recent conversations where clients found offshore engineers cheaper than heavy AI reliance. A holding company leader warned that infrastructure costs can quickly outpace human labor expenses without discipline.

An industry analyst was blunter: "The honeymoon around AI and agencies is essentially over. CMOs were sold on cost efficiencies and speed to market, and now, as they review and renew scopes of work, they're still not seeing the cost savings in terms of fee."

Coca-Cola's Christmas ad used 70,000 individual prompts last year, a figure celebrated at the time. That same statistic would likely face scrutiny today as token volumes and compute costs sit on CEO and CFO desks.

These details were first reported by Digiday.