A Practical Guide to Tracking Claude Costs in Microsoft Foundry

In this blog post A Practical Guide to Tracking Claude Costs in Microsoft Foundry we will explain how to see what your AI agents cost, identify what is driving the bill and connect spending to measurable business value.

The problem is rarely a single expensive request. It is thousands of small agent actions, repeated prompts, unnecessary tool calls and failed tasks accumulating across multiple teams. By the time the Azure invoice arrives, leaders can see the total spend but may not know which agent, department or business process caused it.

That is where FinOps comes in. FinOps is a practical approach to making technology teams, finance teams and business owners jointly accountable for cloud spending. For AI agents, it means moving beyond “How many tokens did we use?” and asking “What did we pay for each successful business outcome?”

How Claude costs work in Microsoft Foundry

Microsoft Foundry is a platform for building, deploying and governing AI applications and agents. Claude is Anthropic’s family of AI models, while an agent is the wider system that gives the model instructions, access to business information and permission to use tools.

For example, Claude may interpret a customer request, search a knowledge base, call a pricing system and draft a response. One user interaction can therefore trigger several model calls rather than one simple question and answer.

Claude usage in Microsoft Foundry is billed through Claude Consumption Units, or CCUs. In plain English, Microsoft measures the input and output tokens used by the selected Claude model, applies the relevant model rates and converts that usage into a common billing unit that appears through Azure Marketplace on your Microsoft invoice.

A token is a small piece of text processed by the model. Long instructions, large documents, conversation history, tool descriptions and detailed answers all increase token usage. However, tokens alone do not show whether an agent completed anything useful.

Use three views instead of relying on one dashboard

A reliable cost model needs three connected views. Each answers a different management question.

1. Azure Cost Management shows what you were billed

Azure Cost Management is the financial source of truth. It shows Marketplace charges, spending trends, budgets and invoice-level usage.

Use it to answer questions such as:

• How much did we spend on Claude this month?
• Is spending increasing faster than agent adoption?
• Which subscription, resource group or Foundry project owns the cost?
• Are we likely to exceed the monthly budget?

Cost information can take time to appear, so it should not be treated as a live emergency brake. Review trends over days and weeks rather than expecting every dashboard to match minute by minute.

2. Application Insights shows what the agent did

Connect Microsoft Foundry to Azure Monitor Application Insights, which records operational information about applications and agents. Foundry uses OpenTelemetry, an open standard for recording what happened during an application request.

Tracing can show the model used, token consumption, tool calls, retries, errors, response time and the sequence of steps taken by an agent. This helps you find expensive behaviour that an invoice cannot explain.

For example, you may discover that a support agent repeatedly loads a 70-page policy document for simple password-reset questions. Another agent may retry a failed tool four times before giving up. Both behaviours consume money without creating extra value.

For a deeper explanation of the monitoring layer, see our guide to monitoring Claude and agent activity with OpenTelemetry.

3. Business reporting shows whether the spend was worthwhile

The most important view is often missing. Technical telemetry should be connected to a business result such as a resolved support case, completed document review, qualified sales lead or employee hour saved.

A simple internal usage record might look like this:

{
 "agent": "customer-support-agent",
 "department": "operations",
 "environment": "production",
 "model": "claude-model-deployment",
 "input_tokens": 12400,
 "output_tokens": 980,
 "tool_calls": 3,
 "retries": 1,
 "task_completed": true,
 "business_outcome": "case_resolved"
}

This does not replace the Azure invoice. It creates the allocation detail needed to explain the invoice to finance and business leaders.

Build cost ownership into the Foundry structure

Trying to separate costs after several teams have shared the same project and deployment is difficult. Create clear ownership boundaries before production usage grows.

At minimum, separate production from development and assign each agent to a named owner, department and cost centre. For larger environments, use separate Foundry projects or resource groups for major business units and high-spend workloads.

A practical naming and tagging standard should include:

• BusinessOwner — the person accountable for the outcome.
• TechnicalOwner — the team responsible for operation and support.
• CostCentre — where approved spending should be allocated.
• Environment — development, testing or production.
• AgentName — the business-friendly name of the agent.
• DataClassification — the sensitivity of the information involved.

Azure tags are useful for supporting resources, but they may not provide every piece of per-agent allocation detail for Marketplace usage. Agent names and business identifiers should also be included in telemetry.

Measure cost per successful task

Suppose a 200-person professional services company spends $6,000 per month on an internal document-review agent. It processes 10,000 requests, suggesting an average cost of 60 cents per request.

Tracing then reveals that only 7,000 requests finish successfully. The real cost is closer to 86 cents per completed review, before counting employee time spent fixing failures.

After shortening the context, removing unnecessary tool calls and fixing a retry loop, the company processes 9,000 successful reviews for $5,400. Total spending falls, but the more important result is that cost per successful review drops to 60 cents.

This is the type of metric a CIO can defend. It connects AI spending to useful work rather than presenting token counts without context.

Set guardrails before an agent runs away

Budgets and alerts should be configured at the subscription, resource group or workload scope. Useful notification levels are 50%, 75%, 90% and 100% of the approved monthly amount, with both the technical owner and budget owner included.

Remember that a budget alert is a warning, not necessarily a stop button. For stronger control, Microsoft Foundry’s AI Gateway can apply token-rate limits and total token quotas to model deployments at the project level.

Use lower limits in development, where accidental loops and bulk testing are common. Production limits should allow expected demand while still containing abnormal usage.

Security also matters. Traces can contain user prompts, model responses and business information. Restrict access, control retention and avoid recording sensitive content unless it is genuinely needed. Cost visibility should not create a new privacy risk under Australian privacy obligations.

Reduce cost without reducing capability

Once measurement is reliable, focus on the largest sources of waste:

1. Use the right model for each task. Reserve the most capable model for work that needs advanced reasoning. Simpler classification or extraction tasks may not require it.
2. Limit conversation history. Do not resend an entire conversation when a summary will provide enough context.
3. Reduce unnecessary tool exposure. Large tool definitions consume input tokens and can confuse the agent. Our guide to reducing agent token costs with Microsoft Foundry Toolboxes explains this in more detail.
4. Fix retries and failures. A cheap model call that repeats five times is no longer cheap.
5. Test quality alongside cost. Reducing tokens is not a saving if employees must redo the work. Use the approach in monitoring and evaluating Microsoft Foundry agents to balance accuracy, safety and cost.

A practical monthly FinOps routine

Good AI cost management does not require a large committee. A monthly review involving the business owner, technical owner and finance representative is usually enough to start.

• Compare actual Claude spending with the approved budget.
• Review cost per successful task for each production agent.
• Identify the agents with the largest token growth, retry rate or failure rate.
• Confirm that unused experiments and test deployments have been removed.
• Assign one optimisation action with a measurable target.
• Check whether the agent is still delivering the expected time saving or service improvement.

The aim is not to make every AI request as cheap as possible. It is to make spending visible, controlled and tied to a result the business values.

Make AI spending explainable before it becomes material

Claude costs in Microsoft Foundry are manageable when billing data, operational traces and business outcomes are designed to work together. Without those connections, organisations end up with an Azure invoice that finance questions and an engineering dashboard that business leaders cannot interpret.

CloudProInc combines more than 20 years of enterprise IT experience with practical expertise across Microsoft Foundry, Azure, OpenAI, Claude, Microsoft security and Wiz. As a Melbourne-based Microsoft Partner and Wiz Security Integrator, we help organisations build cost controls into AI agents before they become expensive production problems.

If you are not sure which Claude agents are driving your Azure bill, or whether the spend is producing enough business value, we are happy to take a practical look at your current setup — no strings attached.