Microsoft has set out new tools within its Foundry platform aimed at helping enterprises control AI agent spending and measure whether agents are delivering enough business value to justify their cost.

In a blog post published on 10 September, the fourth in a series on agent economics, Steve Sweetman, Microsoft's VP of Product Management for Foundry Models, said good governance needs to make agent consumption visible, attributable and bounded, since traditional cost tools track billing data after spending occurs rather than during a live request. Microsoft said its Foundry Control Plane can enforce token-based rate limits and quotas at the project level in real time, rejecting requests that exceed them, while a separate policy tool can apply consumption limits across multiple projects or model providers, including OpenAI-compatible APIs and Anthropic's Messages API.

The company also introduced ROI for Agents, a feature currently in private preview, which tracks the costs an agent incurs against defined business outcomes such as task completion or customer satisfaction, calculating net value and return on investment. Microsoft said this allows teams to identify low-return conversations and trace them back to specific inefficiencies, such as an oversized model or repetitive tool calls, rather than relying solely on token usage as a proxy for cost.

Microsoft said the new capabilities were separate from Microsoft Agent 365, which is aimed at IT and security teams managing an entire agent estate, while the Foundry tools were designed for developers building and optimising individual agents.


AI Governance Reality Check: Most Enterprises Can’t Answer an Auditor’s Questions
When an auditor asks how your AI made a decision, can you answer? For most enterprises right now, the answer is no. AI adoption has outpaced risk management since ChatGPT’s arrival. Boards now recognise AI as a systemic risk to reputation, intellectual property, and regulatory standing — not just a productivity tool. Yet most organisations remain at Level 1 governance: policies on paper, basic intake processes, and zero visibility into what their agents are doing in production. Speakers: Mahesh Varavooru, Founder at Secure AI, and Stewart Tinson, Project Director at AI-360 You’ll learn: • Why paper-based governance will fail an EU AI Act audit — and what Level 3 looks like in practice • How runtime guardrails work as an AI-era firewall, intercepting every prompt and LLM response in real time • How to defend against prompt injection, jailbreaks, hallucination, and PII/PHI leakage in production systems • Why multi-agent systems amplify governance risk — and how to govern them at scale • How to reach Level 3 maturity in weeks to months — and make governance an enabler, not a blocker Key topics: AI Governance Maturity (L1–L3) • Runtime Guardrails • Prompt Injection Defence • Hallucination Management • Shadow AI & Data Loss • Multi-Agent Security • EU AI Act Compliance • Board KPIs • Human-in-the-Loop • DevSecOps Integration Essential viewing for CISOs, CIOs, Chief Risk Officers, and compliance leaders scaling AI in regulated environments.
Share this post
The link has been copied!