AI agents are moving from answering questions to carrying out work. The bill is moving with them.
Google Cloud's 26 August launch of expanded billing and cost controls for agent workloads is a useful market signal: pooled quotas, runtime estimates, hard monthly caps, consolidated usage views, and options for deferred work. OpenAI's 12 August enterprise report points in the same direction from the demand side. Agentic work is spreading beyond engineering, and longer delegated tasks naturally generate more output than ordinary chat.
The operator implication is simple. Token volume is becoming an operating expense, but a token dashboard is not an operating model.
SMEs need Agent FinOps: a practical discipline that connects the cost of digital coworkers to workflow behaviour, accepted outcomes, risk, and human review.
Token volume is activity, not value
It is tempting to treat low token consumption as efficiency. That can be badly misleading.
An agent may use fewer tokens and still choose the wrong tool, miss an exception, create rework, or push a weak output into human review. Another workflow may use more tokens because it retrieves evidence, checks policy, and validates the result before an approver sees it.
The first looks cheaper on a model-usage chart. The second may be cheaper for the business.
Volume does not prove capability. It does not prove waste either. The question is what the organisation received for the total cost of the work.
The real cost sits across the workflow
For a digital coworker, model calls are only one line in the cost stack. SMEs should account for:
- Model usage: planning, generation, checking, and retries.
- Tool usage: search, databases, CRM or ERP actions, browser work, file processing, and paid APIs.
- Orchestration: routing, memory retrieval, handoffs, logging, and evaluation.
- Human work: approval, correction, exception handling, and escalation.
- Failure cost: rollback, customer recovery, compliance review, and repeated work.
This is why cost leakage is usually an orchestration problem before it is a prompt problem. Weak context creates retries. Loose scope creates tool sprawl. Missing stop rules create loops. Poor approval design moves cheap machine work into expensive human cleanup.
Use cost per accepted outcome
The right unit depends on the workflow. Useful examples include:
- cost per reconciled invoice;
- cost per renewal brief accepted by an account manager;
- cost per support case resolved without reopening;
- cost per qualified lead package approved for human outreach;
- cost per management report accepted without correction.
A practical measure is:
Cost per accepted outcome = total model + tool + infrastructure + human-review cost ÷ outputs accepted without reopening.
Add first-pass acceptance, correction time, and exception rate beside it. Otherwise, teams can make the AI dashboard look efficient by quietly transferring work back to people.
Build five controls into every digital coworker
1. A work-package budget
Set a maximum cost, runtime, retry count, and number of tool calls for the task. A monthly platform cap matters, but the useful control sits closer to the workflow.
2. A stop rule
Decide what happens when the budget is reached or confidence falls: route to a cheaper model, request missing data, defer non-urgent work, ask a person, or stop. A threshold without a response is just a warning.
3. Workflow-level telemetry
Log the workflow version, model route, tool calls, retries, human interventions, final status, and outcome. Cost spikes become diagnosable only when spend is tied to behaviour.
4. Approval thresholds based on risk
High-risk finance, customer, HR, and compliance actions may justify stronger models, independent checks, source verification, and maker-checker approval. Low-risk drafting can run with lighter controls. Cheapest is not the same as best value.
5. A review cadence
Review digital coworkers like managed operations: which workflows are producing value, where corrections cluster, which agents exceed their budget, and which work package should be redesigned or stopped.
A simple SME Agent FinOps scorecard
| Measure | What it tells you | Operator response |
|---|---|---|
| Cost per accepted outcome | Whether spend produces finished work | Compare with the current human or automation baseline |
| First-pass acceptance rate | Whether output quality survives review | Fix context, rules, examples, or evaluation |
| Retry and tool-call rate | Where orchestration is leaking cost | Add limits, improve routing, or simplify the workflow |
| Human review minutes | Whether AI is shifting work rather than removing it | Redesign evidence packets and approval thresholds |
| Exception and rollback rate | Whether autonomy exceeds readiness | Narrow permissions or move the boundary back to “ask” |
What SME leaders should do this week
- Pick one agent workflow already in use or planned.
- Name the accepted business outcome.
- List the full cost stack, including human correction.
- Set a workflow budget and stop rule.
- Instrument retries, tool calls, review time, and final status.
- Compare cost per accepted outcome with the current way of working.
Do not begin by squeezing tokens out of every step. Begin by making the work inspectable. Optimisation without auditability often removes the evidence needed to know whether the system improved.
From chatbot budget to digital coworker economics
The shift from chat to execution changes the management job. Leaders are no longer buying only seats or API calls. They are commissioning work from a mixed workforce of people and agents.
That requires owners, budgets, permissions, evidence, approval gates, and clear outcomes. Agent FinOps is not a finance-only function. It sits where operations, domain expertise, technology, governance, and human judgment meet.
Orchestrate, don't operate—and measure the orchestration by the outcomes it produces.
