AI Agents Are Overspending. Agencies Are Building Audit Logs

Give a machine a budget and no supervisor, and you will get a very fast, very confident overspend. That is roughly the lesson media agencies are learning as they push AI agents into planning and buying workflows, according to reporting by Digiday.

What is actually happening

Agencies rolling out agentic tools for audience targeting, campaign setup and execution are discovering the savings do not arrive automatically. Left unmonitored, agents hallucinate, drift outside their brief, and chew through tokens. Jonathan Whiteside, global evp of technology at Dept, put it bluntly: it can get out of control very, very quickly.

So the tooling arms race has shifted from “can an agent do this?” to “can we prove what the agent did?” Rise, part of the Quad agency group, has been testing AI media buying agents with an unnamed supermarket client since June, using an audit log feature built by SSP PubMatic to catch drift and record every action. Group director Klaudia Smykowska says the appeal is being able to ask an agent why it made a change and reconstruct the reasoning. PubMatic’s Harry Tong says each change is time-stamped and stored with the details of how it was made.

Brainlabs tracks agent use and development across teams. Dept is building monitoring tools that produce decision logs and quality control reports.

The numbers that should worry you

• 56% of companies are deploying AI tools without clear usage policies, per an April Gartner survey of 1,300 senior marketers.
• Gartner estimates 60% of organisations using AI will hit cost overruns caused by a lack of usage tracking.
• Marketing leaders are less likely than other functions to attach financial controls to their team’s AI usage.
• Dept has seen individuals burn through 1.5 million tokens in a single day.

This is a behavioural problem wearing a technical costume

Three well-documented biases are doing the damage here.

Automation bias. People over-trust confident outputs from systems. An agent that returns a tidy media plan feels verified in a way a junior’s spreadsheet never does — even though it isn’t.

Diffusion of responsibility. When a machine executes, nobody feels like the owner. Whiteside’s rule is the correct antidote: every deliverable has an accountable human, and “AI did it” is not a defence.

Zero-price and abstraction effects. Tokens don’t feel like money. They are invisible, fractional and denominated in a unit nobody has an intuition for. Rise’s team currently triangulates cost by watching licensing and token spend, hours saved or lost, and how much of the campaign budget ended up as working media — described by svp George Forge as rough math at this point.

Model selection is the hidden cost centre

Gartner analyst Nicole Greene points to the real driver of rising token bills: people simply aren’t picking the right model for the job. More powerful models cost more, and defaulting to the biggest one inflates spend without improving the output.

Agencies are responding in two directions. Brainlabs runs a tiered token allowance — go over, get reviewed, get more if it’s justified. CEO Daniel Gilbert frames it as wanting people to spend, but usefully, and calls the monitoring “existential” rather than a nice-to-have. PMG’s “Alli For You” sets daily token caps, which creates what consulting and strategy director Dillon Larberg calls a human-in-the-loop moment to coach teams.

Dept has gone furthest, removing individual model choice entirely through an “AI gateway” that routes each request centrally based on commercial and legal factors — including client demands that data not be hosted outside their country.

What to copy this quarter

You don’t need an agency-scale stack to apply the principle. Make the invisible visible and the anonymous accountable:

• Attach a named human owner to every AI-assisted deliverable.
• Report token or credit spend in currency, weekly, next to hours saved.
• Default to the cheapest adequate model; make the expensive one an opt-in with a reason.
• Use constraints like Skills in ChatGPT or Claude to codify your process so agents stay on the rails.

Larberg’s analogy is the one to remember: he doesn’t drive an 18-wheeler to work, and he doesn’t go cross-country in a hybrid. Right-sizing the tool is the whole discipline.

Source: Digiday