Ad agencies race to build guardrails around AI buying tools before costs spiral

Advertising agencies rolling out AI-powered systems to streamline their planning and purchasing workflows are learning a hard lesson: autonomous technology only cuts costs if someone is watching it. Without monitoring, these tools can quickly become expensive liabilities.

As agentic AI systems designed to handle audience segmentation, campaign configuration, and media placement become more common, forward-thinking agencies are building their own tracking and auditing infrastructure. The goal is straightforward—ensure that AI systems aren't making costly mistakes or consuming excessive computational resources.

It can get out of control very, very quickly

Jonathan Whiteside, global evp of technology at Dept

The monitoring imperative

A Gartner survey of 1,300 senior marketing executives conducted in April found that 56% of companies are still deploying AI without formal usage policies in place. The research firm projects that 60% of organizations using AI will experience budget overruns tied to the technology, primarily because they lack proper usage tracking mechanisms. Some of the agencies most active in developing AI buying systems are attempting to prevent this outcome.

Rise, a performance media shop under the Quad agency umbrella, has been piloting AI media buying agents with a supermarket chain since June, according to group director Klaudia Smykowska, who declined to identify the client. The agency is leveraging an audit log system built by PubMatic, a supply-side platform that has collaborated with several agencies—including Butler/Till and Abovo Maxlead in the Netherlands—to test AI media buying capabilities. The tool flags when agents deviate from their initial parameters and creates a complete record of all actions.

I can go and ask, 'why did you make that change? What was the thought process based on that initial brief?' I can make sure everything is recorded and that we can reconstruct what happened and why

Klaudia Smykowska, group director at Rise

According to Harry Tong, director of sales engineering at PubMatic, the system captures every modification an agent makes across any environment with a timestamp and comprehensive documentation of how the change was executed—whether through the user interface or another method.

Tracking agent behavior during experimental phases is essential, but this oversight will likely become standard practice as these tools move beyond testing. Rise is not alone in this effort. Brainlabs monitors how its teams develop and deploy AI agents throughout the organization. Dept is constructing monitoring systems that generate decision logs and quality control reports in formats that humans can easily understand. Whiteside emphasized that accountability remains with people, not machines.

Every deliverable has an accountable human. It's not an excuse to say, 'Oh, AI did it.' You are still accountable whether AI did it or if it was done by a junior.

Jonathan Whiteside, global evp of technology at Dept

Controlling computational costs

PubMatic's current auditing system does not yet calculate the token expense associated with a particular agent pathway. Instead, Rise calculates the cost of its agent experiments by combining three metrics: monitoring token and licensing expenses, measuring time saved or spent by staff, and determining what percentage of campaign budget was allocated to actual media spend. According to George Forge, svp of client technology and product development, this calculation method remains imprecise.

Model selection represents a critical variable in controlling expenses. Computationally sophisticated AI models demand more processing resources and generate higher token costs. Defaulting to the most capable models across the board can unnecessarily inflate expenses.

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

Nicole Greene, Gartner analyst

This presents a challenge for advertising firms. Many marketing agencies have permitted staff to select the model they believe fits their task, reasoning that this freedom encourages experimentation and professional growth. Brainlabs operates a tiered token allocation system where employees can request additional tokens after demonstrating that their usage is justified. The company's philosophy balances spending with purpose.

We want people to spend. We just want them to do it usefully

Daniel Gilbert, founder and CEO of Brainlabs

Gilbert characterized such monitoring as essential to the business, not merely a nice-to-have feature. Some agencies are adopting more restrictive approaches. Dept implemented an "AI gateway" that strips individual employees of the ability to select which model handles their task.

We've had some people burn through 1.5 million tokens in a day

Jonathan Whiteside, global evp of technology at Dept

Under this system, staff submit their prompt or launch their agent, and a centralized team determines which model processes the request based on business or legal requirements. Whiteside noted that some clients mandate the use of specific models to ensure their data remains within their country's borders, which sometimes means selecting models that cannot be hosted in the United States.

PMG created a tool called "Alli4U" that establishes daily caps on token consumption per user. Employees who consistently reach their limits may receive a larger allocation or guidance on selecting alternative models, according to Dillon Larberg, consulting and strategy director.

We can assign token limits or cost limits associated with tokens to users… that provides us a human in the loop moment to meet with teams or meet with individuals

Dillon Larberg, consulting and strategy director at PMG

Other containment strategies exist as well. Both ChatGPT and Claude offer "Skills," a feature that allows users to save specific documents or sequences and restrict an agent to using them, thereby limiting the agent's ability to devise its own approach. PMG leverages this capability to "bake best practices" into agent design, using skills to "codify the steps we want the agent to take to ensure that it stays on the rails."

While agencies must remain vigilant about hallucination risks and client data protection requirements, the core issue is fundamentally about design efficiency—and specifically, selecting the appropriate model for each task.

I don't drive an 18-wheeler to work, and I don't go cross-country in a hybrid

Dillon Larberg, consulting and strategy director at PMG