Microsoft published the fourth and final post in its “The Economics of Agent Optimization” series, focused on governing AI agent spend on Microsoft Foundry.

The governance cycle

flowchart LR
    A[See the spend: traces, monitoring, evaluations] --> B[Bound the spend: token limits, gateway policies, financial budgets]
    B --> C[Prove the return: ROI for agents connects cost to business outcomes]
    C --> D{Decide}
    D --> E[Optimize]
    D --> F[Scale]
    D --> G[Retire]

Three layers of spend limits

  1. Foundry Control Plane enforces tokens-per-minute rate limits and total token quotas at the project scope — requests over the limit get a 429 Too Many Requests, exhausted quotas get 403 Forbidden.
  2. Azure API Management’s AI Gateway applies the llm-token-limit policy across models and providers (OpenAI-compatible APIs, the Anthropic Messages API, MCP servers, agent-to-agent APIs).
  3. Microsoft Cost Management budgets use actual Azure billing data for financial accountability and escalation, complementing the earlier, faster token-based controls.

Proving ROI

ROI for agents in Foundry, currently in private preview, connects agent costs to defined business outcomes (e.g., task completion, customer satisfaction, case deflection), calculating value generated, total cost, net value, and ROI per agent.