Dept Built a Meter for Its Own AI and Found 1.5 Million Tokens Gone in a Day

A single Dept employee burned through 1.5 million tokens in one day. Tokens are the units an AI model charges by, roughly the pieces a sentence gets chopped into before the model reads it, and they are the meter running behind every AI tool your team opens. The agency knows the number because it built something to watch it. Before that, the first place a day like this would have shown up was the invoice.

Digiday reported this week that Dept, Brainlabs, PMG and Rise are all building audit tooling for the AI they run on client work, as agents begin buying real campaigns rather than drafting decks about them. Jonathan Whiteside, who runs technology globally at Dept, described the failure mode in six words:

“It can get out of control very, very quickly.”

We covered the finding that made this inevitable a month ago. KPMG surveyed 2,145 senior leaders and found the companies that could see their AI bill were five times more likely to have provable returns than the ones that couldn’t. That story was about who measures. This one is about what the measuring equipment looks like once somebody builds it, and the answer is more specific than a dashboard.

What an agency meter turns out to be

Dept’s version is an internal AI gateway. Every request from every employee routes through it, and the gateway picks the model based on the commercial and legal requirements of the task rather than whichever application the person happened to open. The routing layer is the cost control. It works by taking model choice away from the individual.

Dillon Larberg of PMG framed model selection the way a procurement team would, talking to PPC Land:

“He does not drive an 18-wheeler to work, and he does not go cross-country in a hybrid”

Most marketing teams do exactly that all day. They send a two-line rewrite to the most expensive reasoning model available because it’s the one in the tab, and reasoning models charge for the thinking as well as the answer. We’ve written about where that goes: agents use five times the AI a person does, and most of it is re-reading.

Daniel Gilbert, who founded Brainlabs and runs it, was blunter about where this sits on the priority list:

“Monitoring is not a nice-to-have. It is existential.”

That’s a strong word from someone selling agency services, and it’s worth asking what he gets from saying it. Agencies that meter their AI can bill for governance, and every quote here comes from a company that would like you to believe your in-house setup is riskier than theirs. Hold that against the next quote, which cuts the other way.

Nobody actually knows what this costs

George Forge, who runs client technology and product development at Rise, said the part most vendors won’t:

“Calculating what any of this actually costs is, at this point, a very rough math equation.”

Rise is building the tooling and still can’t price the output cleanly. That’s the honest state of the practice in September 2026. Any vendor telling you they’ve solved AI cost attribution is claiming to be ahead of the four agencies who’ve been building the meters.

The mechanism is simple enough to explain to a finance lead. Traditional media buying had a plan approved in advance, so the spending decision happened before the money moved. An agent that reasons in a loop decides for itself how much work to do, consumes tokens doing it, and produces the number afterward. There’s no approval step, because there’s no moment where a human gets asked. The invoice is the alarm, and the invoice arrives monthly.

Whiteside’s rule for the governance side is the most portable sentence in the reporting:

“Every deliverable has an accountable human. It’s not an excuse to say, ‘Oh, AI did it.’”

The policy gap underneath

Digiday and PPC Land both cite a Gartner survey from April of 1,300 senior marketers finding that 56% of companies deploying AI lack clear usage policies. We couldn’t open Gartner’s own write-up to confirm the figure or its wording, so treat it as reported by those outlets rather than something we verified at the source. The direction matches what the agencies describe, which is teams running production AI with no rulebook.

The cost story and the policy story are the same story. A usage policy is what tells someone which model to use for a rewrite. Without one, the routing decision falls to whoever has the tab open, and the meter runs at whatever rate that person’s habits set.

Five things worth doing this month

1. Find out whether anyone at your company can produce your AI spend broken out by team. Not the total on the corporate card, the split. If nobody can, you’re in the 3% cohort by default.
2. Ask your agency what their per-account AI cost was last month and who pays for it. This is now a contract question. Agencies building meters are going to start showing you the readings, and you want to be asking before you’re being told.
3. Set a model default in writing. One paragraph naming which model handles routine copy and which handles genuine reasoning work does most of what Dept’s gateway does, for free.
4. Put a daily cap on any agent that spends money or runs unattended. A cap turns a runaway into an interruption. Without one the ceiling is your credit limit.
5. Attach a named human to every AI deliverable that reaches a client or a customer. Whiteside’s rule is free to adopt and it’s the only one here that also covers the mistakes that aren’t about money.

Dept found its 1.5 million token day because somebody built the thing that would find it. The four agencies here are ahead on exactly one count: they can read their own meter. Most of their clients still can’t read theirs.