One staffer, 1.5 million tokens, and a meter nobody was watching
A single employee at Dept burned through 1.5 million tokens in one day. The agency's answer was not a policy document, it was a meter.
The meter nobody was reading
Foxes do not count the chickens, they count the exits. Agencies spent eighteen months counting what their agents could do and almost none of that time counting what those agents cost, which is how a single staffer at Dept came to spend 1.5 million tokens inside one working day.
Read nextMost of It Went Through the Mesh. They Counted What Stayed.Agentic browsing just won its first appeal
That figure comes from Digiday's 4 September report on agencies building their own audit tools, and it was picked up the same week by PPC Land. The detail that matters is not the number. It is that nobody noticed until afterwards.
tokens spent by one person at Dept in a single day, before any cap existed
An agent that reasons in a loop consumes tokens the way a tap consumes water, quietly and continuously, and no line in the standard agency operating model was ever built to watch that dial in real time. That absence is what ai agent governance now has to fill, and it is a commercial job rather than a compliance one.
Call it ai agent governance and a room reaches for a framework. Call it a meter and the same room reaches for a switch, which is why the plumbing word is the better one here.
The reflex reaction is a policy. The useful reaction is plumbing. Dept's answer, reported by Digiday, was an internal AI gateway that takes model choice away from the individual: staff send the prompt, and the routing decision is made centrally on commercial and legal grounds rather than personal preference.
Why this landed on media buyers first
Media buying was always going to be the first room where this bill arrived, because ai media buying is the one agency function where an agent can spend twice: once on inference, and once on the media it just bought. Every other department risks a wasted afternoon. This one risks a wasted budget.
PubMatic's AgenticOS gave one worked example. A direct-to-consumer brand, Rouge Care, ran two agents through it in the second quarter of 2026 on a 25,000 dollar spend and reported over 125,000 dollars in attributable sales, a five-fold return, as PPC Land recorded. Read that as one vendor-supplied case, not as a category benchmark, because it is.
Half the market already pulled back
The anecdote is vivid. The survey behind it is the part a finance director will read twice. KPMG's Global AI Pulse polled 2,145 C-suite and senior leaders across 20 countries, territories and jurisdictions, and found that 49 percent had narrowed, delayed or paused agent deployments once operating costs outran the value delivered.
Those numbers are published by KPMG's Global AI Pulse and were summarised by PPC Land's report on the KPMG findings and KPMG's own release on cost visibility. Note the wording carefully: narrowed, delayed or paused covers three quite different decisions, and a fox reading that hedgerow does not pretend it saw one animal.
| Item | Value |
|---|---|
| 49% of senior leaders narrowed, delayed or | 49% of senior leaders narrowed, delayed or |
| paused AI agent deployments once running costs outran th | paused AI agent deployments once running costs outran th |
Two further readings from the same study sharpen it. Only 26 percent of those leaders report full, real-time visibility of what their AI systems cost to run. A further 42 percent report partial visibility, and a third say they do not properly understand the cost structures involved, token pricing included.
Put those three bars beside the Dept anecdote and the shape of the problem is obvious. This is not a story about models behaving badly. It is a story about a meter running in a room with the light off, and the fix is a switch rather than a sermon.
There is a second-order effect worth naming. When leaders cannot see the bill, they cut the programme rather than the waste, because cutting the programme is the only lever they can find. Visibility is therefore not a reporting nicety, it is what keeps the work alive.
Five controls that actually hold
Here is the practical part, drawn from what agencies have actually built rather than from what vendors suggest. None of it needs a new platform, and all of it survives the next model release, because it governs behaviour rather than technology.
Take model choice away from the individual. Dept routes every request centrally on commercial and legal grounds, so the expensive model is a decision rather than a habit.
PMG's Alli For You pools access to every major model under a daily token cap per user, with an escalation review when someone hits the ceiling.
Rise uses PubMatic's audit log to watch for behaviour drift, with time-stamped agent changes stored in full. Spend tells you how much. The log tells you what happened.
Divide tokens by the thing produced: a brief, a build, a campaign change. A per-outcome rate is the only figure that survives a model price cut.
One person owns the meter and the escalation. An unowned cap is a suggestion, and suggestions do not appear on invoices.
Notice what is missing from that list. There is no committee, no maturity model and no framework with five pillars. The agencies that got this right built a gateway and a cap, which is roughly two weeks of engineering, then argued about the policy afterwards.
What a sensible cap looks like
PMG's published number is a 50 dollar daily token ceiling per user across the pooled model set. Treat that as a starting posture rather than a standard: the right ceiling depends on whether your people are drafting copy or running agents that hold context across an entire account.
This is where ai agent pricing stops being a vendor's problem and becomes yours. Headline rates are not running costs. Context tiering and cache reads move the real number a long way, so a rate card quoted per million tokens tells you almost nothing about the bill at the end of a month of agent work.
The 49% pulled back AI agents over cost headline hides two different decisions. In KPMG's survey of 2,000+ executives across 20 countries, 24% narrowed deployments. Another 25% paused further rollout after costs stopped matching expected value.
That reading is right, and it is the one most coverage skipped. Pausing a pilot because the maths did not work is a healthy decision. Pausing it because nobody could produce the maths at all is a governance failure wearing the same coat.
What this changes for marketing teams
For a marketing leader the practical question is narrow: what do you fit before the next invoice. The answer is a cap, a log and a per-outcome rate, in that order, and the whole thing is cheaper than one month of unmetered enthusiasm.
agentic ai marketing has a specific hazard that general enterprise AI does not, which is that the agent's output is a spend instruction. An agent that drafts a bad email wastes an hour. An agent that widens a match type or lifts a budget cap wastes money at auction speed, which is why the audit log matters as much as the token ceiling.
| Control | Owner | Evidence it is real |
|---|---|---|
| Central model routing | Engineering | A request log showing model choice made outside the user's hands |
| Daily token cap | Operations | A user who hit the ceiling and an escalation record for it |
| Agent audit log | Operations | Time-stamped agent changes retrievable for any campaign edit |
| Per-outcome rate | Finance | Tokens per brief or per campaign change, tracked monthly |
| Named owner | Leadership | One name on the escalation path, not a mailbox |
Ask the same question about ai agent costs per month that you would ask about any other retained line: what did it buy, who signed it off, and what happens if it doubles. If the answer to any of those is a shrug, the meter is not being read.
An unmetered agent is not a productivity gain. It is an invoice with a delay on it.
If you want the meter fitted rather than described, that is the work folkfox AI consultancy does, and it sits alongside the buying discipline in PPC and paid social. Rates are published on the pricing page, because a consultancy that hides its own numbers is a poor guide to reading anybody else's.
Where AI agent governance meets the rate card
Vendor rate cards are where ai agent governance quietly fails, because the published figure is a price per token and your invoice is a price per task. OpenAI sets out its own tiering on its pricing documentation, and Anthropic documents cache reads and long-context behaviour on its pricing pages. Both are accurate. Neither is your bill, and a fox that mistakes a scent for a supper goes hungry.
The commercial weather around this is worth tracking too. Digiday has reported agencies chasing an edge from agents while fending off clients planning to bring the same tools in house, in its coverage of the in-housing squeeze, and AdExchanger has been logging the same cost anxiety in its daily roundups. Read together, they describe a market where ai agent governance is becoming a pitch credential rather than an internal chore.
That is the useful reframing for anyone selling services. A client asking what your agents cost to run is not being difficult, they are asking the question their own board asked them last quarter. An agency that can answer with a per-outcome rate has an advantage that no demo matches, and one that cannot is on the wrong side of the same hedgerow.
None of this needs a philosophy. Fit the meter, watch the dial, keep the trail of what the agent changed, and ai agent governance stops being a term and starts being a number you can defend on a Tuesday morning.
The parts nobody can prove yet
Two honest caveats, because a piece about cost discipline that overclaims would be a poor advertisement for it.
First, vendor case studies are vendor marketing. The Rouge Care result came through a platform with an interest in the number, and one campaign is not a category. Second, the independent evidence on what agents can reliably finish is thinner than the discourse suggests. METR's own write-up of its task-length work is careful about what its curve measures, and its task set skews to software and research work rather than to the messy middle of a marketing account.
Third, and most practically, an agent with write access to an ad account is a different risk class from one that reads. That distinction is already visible in how platforms ship their connectors: Snapchat's advertising server launched read-only, with write access held back, as its own Snapchat for Business post sets out. Copy that posture before you copy the enthusiasm.
The trail here is not hard to follow, it is just unlit. Fit the meter, name the owner, keep the log, and the rest of the argument about whether agents are worth it becomes answerable with evidence instead of adjectives. That is the whole of ai agent governance in one sentence, and it is why folkfox writes it as plumbing rather than philosophy.
We have written before about what happens when agents act on the open web without a ledger behind them, in the agentic commerce attribution gap and in the rules for agentic browsing. The pattern repeats: the capability arrives first, the accounting arrives late, and the brands that fare best are the ones that build the accounting anyway.
One last practical note on scope. ai agent governance covers three separable questions: what the agent may spend, what it may touch, and what it must record. Teams that collapse them into one policy document end up enforcing none of them, because each has a different owner and a different failure mode.
Take them in that order, and the undergrowth clears quickly. Spend is a cap, access is a scope, and the record is a log. Every serious ai agent governance programme folkfox has seen working is those three things wearing a more expensive name.
Treat ai agent governance as a running discipline rather than a launch task. The meter is read weekly, the log is sampled monthly, and the per-outcome rate is argued about quarterly. Done that way, ai agent governance costs an hour a week and answers the only question a board actually asks about agents, which is whether the spending bought anything.
Frequently asked questions
What is ai agent governance in a marketing team?
It is the set of controls that decide what an autonomous agent may spend, which model it uses, and what record it leaves. In practice that means central model routing, a daily token cap, an audit log of the changes the agent made, and one named owner for escalations.
How much does it cost to run an AI agent for a month?
There is no single figure, because cost follows context rather than headcount. PMG operates a 50 dollar daily token ceiling per user across pooled models. The honest way to budget is a per-outcome rate: tokens spent per brief, per build or per campaign change.
Why did so many companies pause their AI agent projects?
KPMG's Global AI Pulse found 49 percent of 2,145 senior leaders narrowed, delayed or paused deployments when operating costs outran the value delivered. Only 26 percent reported full real-time visibility of what their AI systems cost to run, so many were cutting blind.
What are ai agent costs per month made of?
Inference tokens are only the visible part. The real drivers are context length, how often an agent re-reads its own context, cache behaviour and retries. Two teams on identical seat counts can differ several times over on the same rate card.
Should an AI agent have write access to our ad accounts?
Not on day one. Start read-only, watch the audit log for a full billing cycle, then grant narrow write scopes with a spend ceiling attached. Snapchat shipped its own advertising connector read-only first, which is a reasonable posture to copy.
Is agentic ai marketing worth the overhead of governing it?
It can be, but only where you can show the maths. The controls cost roughly two weeks of engineering. Running unmetered cost one agency 1.5 million tokens from a single person in a day, which is the cheaper argument of the two.
Most of It Went Through the Mesh. They Counted What Stayed.
The same accounting lag, one layer further down the funnel: agent purchases arriving faster than merchant reporting can see them.
Read the pieceAgentic browsing just won its first appeal
What a court decided about user-directed agents, and the access rules your own site should adopt.
Read the pieceOpenAI's Own Agents Formed a Swarm. Here Is What Agentic AI Security Missed.
When agents coordinate without supervision, the audit log is the only thing that explains what happened.
Read the pieceWant the meter fitted, not described?
folkfox builds the cap, the routing and the per-outcome rate that make an AI programme defensible to a finance director, then keeps the media buying honest alongside it.
Want folkfox in your Google results and AI answers? Set folkfox as a preferred source.