Last week Google did for enterprise AI budgeting what most agencies still have not done for themselves: it installed a meter. At Gemini at Work 2026 on October 8, Google Cloud introduced a single “universal” Gemini agent that plans work, connects to Workspace, Microsoft 365, Slack, Jira and Postgres, and chooses a model per task. Two supporting features matter more to an agency’s P&L than the agent itself: smart routing, which sends each job to the cheapest model that can do it, and real-time spend caps — a hard per-project limit that pauses the agent’s work when it trips. Google’s own justification is the tell: “Running every simple loop through a premium model quickly breaks corporate budgets.”
Your agency has the same problem, unsolved. Client work runs on variable inference costs that hide inside a fixed retainer, pooled behind one shared API key, priced as if tokens were free. The fix is settings plus one sentence in a statement of work.
What Google Actually Shipped
Three cost controls arrived together, per Google’s announcement:
- Model choice per task. The agent “runs each job on the model that fits best,” across Gemini and Anthropic’s Claude models. Google’s framing: “The best model for the task is not always the largest one.”
- Smart routing. An automatic triage layer that assigns each workload to the model delivering maximum performance at the lowest cost.
- Real-time spend caps. A hard limit per project in the billing console; when it trips, the agent pauses. Because tracking is per project, Google notes companies can charge AI costs back to specific departments.
One caveat: licensing for persistent “coworker” accounts and usage charges for long autonomous runs remain undefined, as Yahoo Finance notes. Unclear vendor pricing is the reason to meter per client.
Why This Lands on Your Margin, Not Your Client’s
Agency pricing was built when the marginal cost of a deliverable was a person’s time. AI flipped that: the retainer is fixed, but research runs, drafting, transcription, tagging and QA passes are metered. Pool every client behind one key and three failures arrive at once — you cannot tell which account is expensive, you cannot stop a runaway job, and when one client’s usage quietly triples you find out on the invoice. We walked through that leak in Agency AI Spend Is Up. Profit Isn’t.
Clients are also getting their own agents, pushing work into shared surfaces where both sides see costs and outputs, as we covered in When Clients Get Their Own AI Agent. Showing per-client AI cost and a written routing policy marks you as a partner whose operations are under control.
Step 1: One Client, One Credential, One Cap
Attribution is an engineering problem before an accounting one. Give each client its own API project — or at minimum its own key and billing tag — so cost rolls up per account instead of being reconstructed from memory at month-end. Cap each project at roughly expected usage plus 30 percent. When a cap trips, the run stops and a human reviews: the difference between an incident you handle and a surprise you absorb.
The same isolation is a data boundary and an audit trail: one client’s prompts never share a key with another’s.
Step 2: Route by Task, Not by Habit
Most agency work is not frontier-model work, and paying frontier prices for it is a habit rather than a decision. A workable three-tier default:
- Tier 1, small and cheap: classification, tagging, extraction, deduplication, transcript cleanup, first-pass summarization, routing. High volume, low judgment.
- Tier 2, mid-size: drafting from an approved brief, translation, reformatting, turning structured data into copy. Human-reviewed before it ships.
- Tier 3, frontier: strategy synthesis, adversarial QA of your own draft, anything client-facing that ships unreviewed.
Two rules keep routing honest: pick the tier by the cost of an error, not the size of the task, and log model and token count on every job so tier boundaries are evidence rather than opinion. Re-test quarterly — the cheap tier keeps absorbing work that needed the expensive one six months ago.
Step 3: Name a Billable Unit and Put It in the SOW
Nobody buys tokens, and nobody wants a usage meter stapled to their invoice either. Price the deliverable — a research brief, ten creative variants, a monthly reporting cycle — and write the allowance in: “includes up to N runs per month; additional runs billed at $X each.”
Per unit, the numbers are small: a brief costing a dollar in tokens is a rounding error in inference, not the product. Margin is made on the tail, where a client asking for 400 variants a month is no longer a rounding error — and the allowance clause is what lets you say yes instead of absorbing it.
Step 4: Meter Three Numbers Every Week
- AI cost per client, by project.
- AI cost as a percentage of that client’s monthly fee.
- Cost per deliverable, trended week over week.
Then set a threshold and honor it. For most shops the trigger is inference spending crossing roughly 5 percent of a client’s fee, at which point you re-scope or reprice at renewal. It is the same discipline that makes guardrails work, laid out in Agentic Ad Ops: Set the Guardrails Before Agents Spend: one owner, narrow measurement, a decision attached to the number.
Don’t Build a Price on Someone Else’s Subsidy
On October 6, Anthropic expanded its Claude Startups program with a free year of Claude Team for up to five premium seats and $1,000 in API credits, plus partner discounts worth up to $45,000 (TechCrunch, CNBC). Agencies founded in the last five years or funded in the last two could move a slice of their tooling onto someone else’s balance sheet.
Then the terms moved. Anthropic’s program page now says it is “over capacity” on the Claude Team and credit offers after “hundreds of thousands of applications,” and will re-review all applications, which “may result in a change in the status of some applications.” Claimed offers are honored; the rest are in limbo, days after launch.
A subsidy is inventory, not infrastructure. Spend credits on the metering and routing above, keep a second provider warm so no engagement depends on one vendor’s promotional calendar, and price clients from steady-state rates. A free year is margin — not a discount you hand the client in the proposal.
What to Do This Week
- Split last month’s model spend by client. If you cannot, fix attribution first.
- Create one API project per active client and move new runs off the shared key.
- Set a spend cap at 130 percent of expected usage, alerting one owner.
- Write the routing table: task, tier, and who reviews the output.
- Add an AI usage allowance and an overage rate to your next two renewals.
None of it requires a new platform, and all of it is copyable from what Google shipped for enterprises this week. Metering per client is the difference between AI work that compounds your margin and AI work that quietly eats it — and if the plumbing is where your team loses the hours, that is the kind of operations work we do.