Ten years ago, businesses would open their cloud bill at the end of the month and often get a surprise. FinOps emerged to help them understand that spending, assign costs to teams and determine whether the expense was justified by the business value.

AI agents are bringing back the same problem, only faster. Requests, tool calls and reasoning loops rack up token usage. Without monitoring, the bill may be the first sign that something needs attention.

In The agentic enterprise, published by AWS Marketplace, Sreenivas Gadhar of Marsh makes the connection: token governance plays the same role for AI that FinOps has played for cloud spending. Here’s how to put it into practice.

Why agents cost more than you expect

An agent’s cost goes well beyond the initial prompt. A single user request can trigger ten model calls. Four things can quietly drive up the bill:

• Reasoning loops. The agent plans, acts, observes and starts again. Without limits, an ambiguous task can keep it running for a long time.
• Growing context. Conversation history and tool results accumulate at each step. The tenth step can cost much more than the first.
• Retries and evaluators. A tool call fails, so the agent tries again. An evaluator agent reviews the answer. Each pass adds to the cost.
• Multi-agent architectures. An orchestrator, a planner and three execution agents multiply the number of calls. The contributors to the collection agree: start with a single agent.

A common mistake is to choose a cheaper model when the real waste comes from unnecessary complexity. As Ram Kumar of Prudential points out, simplifying the workflow and removing unnecessary reasoning steps can save more than switching models.

7 practices to put in place now

1. Tag each agent as a resource

Before you can optimize spending, you need to know where it comes from. Treat each agent as a workload with an owner, a team, a use case and an environment.

On AWS, Amazon Bedrock application inference profiles let you tag usage by agent or team. Activate those tags as cost allocation tags in your billing settings.

Action: Define a minimum tagging standard (agent, team, use-case, env) and require it before any deployment.

2. Measure cost per verified outcome, not per request

Heavy usage doesn’t necessarily mean an agent is worth its cost. A more useful metric is total cost divided by the number of successfully completed and verified tasks.

For example, a support agent costs $4,000 a month and resolves 2,000 requests without escalation. That works out to $2 per resolution. If a person handling the same request costs $9, the savings are clear. If the agent resolves only 300 requests, the economics look very different.

Action: Define what counts as a successful outcome for each agent before putting it into production.

3. Set execution limits

An agent with no limits can keep spending long after it should have stopped. Three technical controls can help prevent runaway costs:

• A maximum number of iterations per task, with a fallback such as escalation to a person.
• A token limit per task and per session.
• A timeout for each tool call, with a limited number of retries.

Add AWS Budgets and cost anomaly detection to monitor the spending attributed in step 1. These alerts can help you catch problems before the monthly bill arrives.

4. Match the model to the task

Not every task needs the most powerful model. A model cascade routes routine work, such as classification, extraction and short summaries, to a smaller, faster model. More complex requests go to a more capable one.

Bedrock offers intelligent prompt routing to make that selection automatically within the same model family. You can also build your own routing logic by classifying requests before sending them to a model.

5. Avoid paying for the same work twice

Many interactions involve information the agent has already processed. There are three ways to reduce that repeated work:

• Prompt caching for system instructions and reference documents reused across calls. Reading cached tokens costs considerably less.
• Response caching for frequently asked questions and tool results that remain valid.
• Precomputed information, such as dashboards, product summaries or key metrics generated in advance instead of with every request.

6. Keep context lean

More context doesn’t automatically produce better answers. It does increase costs and can make it harder for the agent to identify what matters. Nan Li of Citizens Bank describes this as memory hygiene: give the agent only the information it needs for the task.

Use retrieval-augmented generation (RAG) to find the relevant passage instead of feeding the agent entire documents. Summarize long conversation histories. Add long-term memory only when it provides measurable value.

7. Batch work that can wait

Document analysis, data enrichment and report generation don’t always need an immediate response. For eligible models, Bedrock batch inference costs roughly half as much as on-demand inference.

Action: Review your use cases and classify each as either “real time” or “can wait a few hours.”

Who owns the costs, and how will they be reviewed?

Technical controls need people responsible for maintaining them. Cloud FinOps has already shown this: tools help, but clear responsibilities and regular reviews are what change spending habits.

Assign an owner to every agent

Every agent in production needs someone accountable for its value, cost and behaviour. Keep a simple inventory that records the agent, owner, use case, success metric and monthly budget.

Include token budgets in project planning

Estimate an agent’s running costs during design. Each AI project brief should include an expected operating cost alongside the development effort.

Give teams visibility into their spending

Teams can make better decisions when they can see what their agents cost. Start with showback, which reports costs to each team, before moving to chargeback, which bills those costs to their budgets.

Review agents every month

Set aside 30 minutes a month to review each agent’s cost, usage volume, cost per outcome, escalation rate and trends. Then decide whether to keep it as is, optimize it or retire it. Retiring an agent that delivers too little value is a sound spending decision.

Bring FinOps and AI teams together

The FinOps team understands cost management. The AI team understands agent architecture. Both need to be involved to understand what drives spending and where changes will help. Include FinOps from the start of each new agent project.

Know what each agent is worth

Agentic AI needs the same spending discipline as any cloud workload. Track what each agent costs, whether it completes its tasks successfully and what those results are worth to the business. That gives you a basis for deciding where to invest more and where to cut back.

Deploying your first AI agents on AWS? Let’s talk about getting the costs and controls right from the start.