Agents & automation
Cheaper tokens, broken budgets: why AI agents need spending caps
Google now lets a hard spending cap pause its work agent, which tells us agents should be run as cost centres, with an owner, a limit and a named person who can press resume.
Google Cloud says per-token prices have fallen 98% since 2024. In the same breath, its answer to runaway AI bills is a hard spending cap that pauses an agent. Both statements are true, and the gap between them is worth thinking about.
At its Gemini at Work event on 8 October, Google introduced the Gemini agent, a single agent for work. The cost controls it announced were these:
- A hard spending cap per project, set in the billing console.
- When the cap is hit, that project’s agent pauses until someone clicks resume.
- Spend tracked per project, so costs can be charged back to departments.
- Routing that sends each job to the cheapest model that can do it well.
Google’s own explanation was that “running every simple loop through a premium model quickly breaks corporate budgets.”
Two caveats before the argument
I want to be careful about what is known. Google’s launch post gives no pricing and no general availability date for the agent. And Google told the analyst firm Futurum that it does not intend to charge a per-seat licence for coworker agents. So how an agent’s activity will be metered is still open. Everything below is my reading of the design, not of a price list.
Why a falling unit price does not protect the budget
A budget is price multiplied by volume. Falling token prices help only if volume stays put, and agents are built to raise volume. An agent works in loops: it reads, plans, calls a tool, checks the result and tries again. Each loop is cheap. The number of loops is not something a person chooses one by one.
Chat use has a natural brake, which is the human typing the next message. An agent removes the brake, which is its point. So the control that used to sit in a person’s hands has to move into the system as a limit.
What the cap tells us
My read is that when an agent can be paused for going over budget, it is being managed less like software and more like a cost centre. Software is bought, installed and mostly forgotten until renewal. A cost centre has a budget holder, a ceiling, a monthly review and someone who answers for the variance.
That framing changes the questions a leader asks. “Does it work?” becomes “Whose budget does it run on, and what happens when the money runs out?”
Three decisions to make before an agent scales
I would settle three things before any agent moves beyond a pilot.
Whose budget it runs on. If the agent’s spend sits in a shared pool, nobody notices a drift until the month closes. Per-project tracking, as Google describes, lets the cost be charged to the department that benefits. That also makes the department care about the result.
Where the hard limit sits. The limit should be a number someone chose on purpose, not a default. It helps to have two levels: a soft threshold at which the agent keeps working but someone is told, and a hard cap at which it stops.
Who is allowed to press resume. This is the one most teams skip. A pause is only a control if restarting it is a decision. If anyone can click resume, the cap is a speed bump. If only the budget holder can, the cap forces a conversation about whether the spend was worth it.
What this looks like in systems I have built
None of this is new to me from the announcement; it matches how I set up agent systems that touch money.
In a marketing system I built, agents can spend a set daily budget on their own. Above that threshold they must ask for approval. People decide where the checkpoints sit: before a post goes live, and before money is spent. Users asked for “take control” and “freeze all spending” buttons, so I added them. In a Google Ads agent system built in my teaching, the agent prepares the campaign and never clicks Publish. A person does.
The principle in both is the same as Google’s cap, applied at a different point. The agent does the work up to a line, and a human owns the line.
On company accounts for commercial AI tools, I give the same advice: switch off any “extra usage” option unless you have set a cap on it. Without a cap, bills can grow far beyond what anyone expected.
The routing point is easy to miss
The fourth item on Google’s list, sending each job to the cheapest model that can do it well, matters as much as the cap. A cap limits the damage when spend goes wrong. Routing lowers the spend in the first place. If a simple formatting loop runs on a premium model, you are paying premium rates for work a smaller model could do. I come back to that in a separate piece on choosing the model tier.
A short template
If you are about to put an agent into daily use, write this on one page:
- The agent’s purpose, in one sentence.
- The budget holder, by name.
- The monthly ceiling, and the soft threshold below it.
- What the agent may do alone, and what needs approval.
- Who may resume it after a pause, and what they must check first.
- How spend is reported, and how often.
If you cannot fill in item 5, the agent is not ready to scale.
The question I would put to your team
Who would press resume in your organisation, and who would argue against it? The answer tells you whether you have a control or a formality. The person who presses resume should be able to see what the agent spent the money on. The person who argues against it should have the standing to be heard. If both exist, the cap is doing its job.