Your AI agents need a budget owner
Agencies testing AI buying agents found the tools do not save money unattended. The lesson applies to every company deploying agents: an owner, a cap and a log.
Digiday reported on September 4 that media agencies testing AI agents have hit a problem nobody put in the business case. The agents spend money on their own, and nobody was watching. At Dept, an executive said some people had used 1.5 million tokens in a single day. Dept's answer was an AI gateway: staff no longer pick the model. They submit the task, and the model is chosen centrally on commercial and legal grounds.
One person's token bill is small. Multiply it by every seat and every agent running all day and it becomes a P&L line with no owner.
On the surface that is a media story. Underneath, it is cost governance, and it applies to every company that moved AI from pilot to production this year. CFO Dive reported a July survey of 300 executives by WitnessAI, an AI security vendor: 68% of US companies said at least some AI initiatives ran over budget in the past year, and a third said overruns happened mostly or always. Only 9% said more than three-quarters of their AI initiatives delivered a measurable financial return.
Why it happens
Token pricing looks like cloud pricing, and most companies handled cloud badly the first time too. Usage-based cost. No cap. No owner. No visibility until the invoice. The unit price falls, which is what vendors point to. Consumption rises faster, which is what the CFO sees. Flexera's 2026 cloud survey found self-estimated cloud waste rising again for the first time in five years, and it tied the rise to AI workloads.
What the agencies built
| Control | Who Digiday says is doing it | What to copy |
|---|---|---|
| Audit log | Rise (Quad), using a PubMatic tool | Every agent action time-stamped, with alerts when it drifts outside set parameters |
| Allowance per person | Brainlabs | A token allowance per user, with a review before anyone gets more |
| Daily cap | PMG | Daily token or cost limits, with heavy users moved to cheaper models where they fit |
| Central model routing | Dept | People send the task; the system picks the model on cost, data residency and fit |
| Constrained agents | PMG | Saved, step-by-step instructions the agent follows instead of inventing its own path |
| A named human | Dept | Every piece of work has one accountable person, whether or not AI produced it |
None of this is exotic. It is the same discipline good companies apply to cloud and media: an owner, a budget, a cap, a log, and a monthly number someone has to explain.
The part that lands on your agency fee
Your agency's AI costs are real now. They are being recovered somewhere: from your fee, your media, or the agency's margin. Ask which.
In the audits we run, AI increasingly shows up as a productivity gain in the pitch and a cost line on the invoice. It can't be both. Three questions for your next quarterly review:
- How are AI costs on our account recovered, and where do they appear on the invoice?
- Is usage on our work capped and logged?
- Can we see the log?
If the answers are vague, the cost is probably landing on you.
Sources: Digiday, media agencies build audit tools to prevent AI agents from overcharging, CFO Dive, nearly 7 in 10 firms report AI cost overruns, Flexera, what 2026 data tells us about cloud spend.
Why do AI agents cause cost overruns?
Usage-based pricing with no cap, owner or visibility until the invoice. Unit prices fall, but consumption grows faster. A WitnessAI survey reported by CFO Dive found 68% of US companies had at least some AI initiatives run over budget.
What controls keep AI agent spend in check?
An audit log of agent actions, allowances or daily caps per user and agent, central routing that picks the model on cost and data rules, constrained agent instructions, and a named person accountable for each piece of work.
Who should own AI spend?
One named executive with a budget and a monthly number, the same way cloud or media spend should be owned.
Should we ask our agency about its AI costs?
Yes. Ask how AI costs on your account are recovered, whether usage is capped and logged, and whether you can see the log.
That is usually the first sign there is money in it. If AI spend has appeared in your P&L without an owner, or in your agency fee without an explanation, we can show you where it is going.