Ad Agencies Build Audit Tools to Control AI Agent Costs

The hidden cost of autonomous AI

Media agencies racing to deploy AI agents for campaign planning and media buying are confronting an unexpected challenge: the technology designed to save money can quickly become a budget drain without proper oversight.

Several major agencies are now building dedicated tracking and auditing systems to monitor their AI agents, concerned that autonomous tools left unchecked will hallucinate, drift from assigned parameters, or consume expensive computing tokens at unsustainable rates.

"It can get out of control very, very quickly," said Jonathan Whiteside, global executive vice president of technology at Dept.

Why it matters

As AI agents move from experimental pilots to production systems handling real client budgets, the lack of cost controls represents a significant business risk. Gartner estimates that 60% of organizations using AI will face cost overruns due to inadequate usage tracking—a problem that could undermine the ROI case for AI adoption in media agencies.

Tracking what agents actually do

Performance media agency Rise, part of the Quad agency group, has been testing AI media buying agents with a supermarket client since June. The agency uses an audit log feature developed by supply-side platform PubMatic to monitor when agents deviate from preset parameters.

"I can go and ask, 'why did you make that change? What was the thought process based on that initial brief?'" said Klaudia Smykowska, group director at Rise. "I can make sure everything is recorded and that we can reconstruct what happened and why."

PubMatic's system timestamps every agent action and records how changes were made, creating a complete decision trail. The SSP provider has worked with several agencies to test AI media buying tools, including Butler/Till and Abovo Maxlead in the Netherlands.

Brainlabs tracks AI agent use and development across its teams, while Dept is developing monitoring tools that produce decision logs and quality control reports. "Every deliverable has an accountable human," Whiteside said. "It's not an excuse to say, 'Oh, AI did it.'"

The model selection problem

Beyond tracking agent behavior, agencies face a more fundamental cost challenge: employees often default to the most powerful—and expensive—AI models regardless of task complexity.

"A lot of the [rising] token consumption costs are because people are literally not choosing the right model to meet the need of the activity," said Nicole Greene, an analyst at Gartner.

Agencies are responding with different approaches. Brainlabs uses a tiered token allocation system, granting employees additional tokens after review if their usage proves valuable. "We want people to spend. We just want them to do it usefully," said founder and CEO Daniel Gilbert, who called such monitoring "existential."

Dept has implemented an "AI gateway" that removes individual model selection entirely. After some employees burned through 1.5 million tokens in a single day, the agency centralized model selection based on commercial and legal considerations. Some clients require specific models to ensure data remains within their country of operation.

PMG developed a tool called "Alli For You" that caps daily token usage per employee. Staff who consistently hit limits may receive larger budgets or guidance on model selection. "I don't drive an 18-wheeler to work, and I don't go cross-country in a hybrid," said Dillon Larberg, consulting and strategy director at PMG.

Guardrails through design

Agencies are also using features like "Skills" in ChatGPT or Claude—saved documents or sequences that constrain agent behavior—to reduce the risk of agents inventing their own approaches. PMG uses these features to "codify the steps we want the agent to take to ensure that it stays on the rails," Larberg said.

The challenge reflects a broader industry reality: according to an April Gartner survey of 1,300 senior marketers, 56% of companies are implementing AI tools without clear usage policies, and marketing leaders are less likely than other executives to assign financial controls to AI usage.

These details were first reported by Digiday.