Short answer: Microsoft's new write-up on agent economics is worth reading past the Azure branding, because it names the three things every organization running AI agents eventually needs: seeing the spend, bounding it, and proving the return. Their tooling is Azure-specific; the framework is not. Here's the platform-agnostic version.
Source: The Economics of Agent Optimization: How AI agent governance controls cost and proves ROI — Microsoft Azure Blog, September 10, 2026.
The three speeds of agent cost control
The piece frames optimization as happening at three time scales, and that framing alone earns its keep:
| Speed | Examples | Who owns it |
|---|---|---|
| Per request (runtime) | Model routing, caching, token rate limits | Engineering |
| Days to weeks | Context engineering, memory, tool selection | Product + engineering |
| Continuous | Cost attribution, spend limits, ROI measurement | Finance + platform |
Most teams get the first row and skip the third — which is exactly how agent programs end up with a surprising invoice and no story about what it bought.
The governance controls that actually matter
Microsoft's implementation maps to controls any platform can express:
-
Attribution before anything else. Every agent's token consumption and cost estimate, tagged to a project, so finance can filter and allocate. On Azure this is Foundry cost management plus API Management's AI Gateway emitting token metrics per API key and subscription. The principle: if you can't say which team's agent spent a dollar, you can't manage it.
-
Limits in the request path, not the invoice. Tokens-per-minute caps (the request gets a 429) and token quotas (403 when exhausted), enforced at request time — because, as the article puts it, an agent retry loop does not wait for the next budget evaluation. Billing-data alerts fire after the spend; request-path limits fire during it. You need both, and the difference between them is the whole point.
-
Gateway policy per consumer. Token limits keyed by subscription, team, or workload identity, working across OpenAI-compatible APIs, the Anthropic Messages API, and MCP servers. One honest caveat from Microsoft: concurrent requests can produce small temporary overages — enforcement is approximate at the margins everywhere.
-
ROI as a measured quantity. Their Foundry ROI view (private preview) computes value generated versus total cost, net value, ROI, average value per conversation, pass rates, and agent-version comparisons. The specific metrics are less important than the discipline: treat each agent as an investment with a measurable return, and retire the ones that can't clear the bar.
The gap Microsoft admits
The most useful paragraph in the piece is the one admitting the hole: platform controls enforce tokens, while finance plans in dollars. A token quota doesn't map cleanly to a budget line, and dollar-denominated agent budgets are only now being built. Any organization designing its own agent governance today should expect to write this translation layer themselves — tokens times price per model, rolled up to something a CFO can recognize.
What this means if you're not all-in on Azure
Strip the product names and the checklist survives intact:
- Attribute spend per agent and per consumer before scaling the fleet.
- Enforce limits where the spend happens (the request path), monitor where the bill lands.
- Evaluate quality and cost together — a smaller model that passes the quality bar is a cost decision, not a compromise.
- Build the tokens-to-dollars translation early; finance will ask for it, and retrofitting attribution is far harder than instrumenting for it up front.
The teams that struggle with agent costs are almost never the ones lacking rate limits — they're the ones that skipped attribution and can't connect spend to a business outcome. That last mile, turning raw usage telemetry into finance-grade answers about what each agent delivers and what it should be allowed to spend, is a data and analytics problem — the same discipline that turns warehouse telemetry into executive dashboards. It's precisely the workflow Dataonia automates with its autonomous finance and operations analytics agents: continuous diagnostics over your own operational data, delivered where decisions get made, instead of in a dashboard nobody opens.
FAQ
What is "agent optimization"? Running AI agents as managed investments: attributing their spend, limiting it where it happens, and measuring the value they produce — rather than treating model usage as an unmanaged utility.
Why enforce token limits if we already have budget alerts? Budget alerts act on billing data, which arrives late. An agent in a retry loop can burn significant spend between an alert and a human response. Request-path limits (429/403) stop the burn during the burn.
Do we need a gateway to do this? Some enforcement point that sees every request — a gateway, a proxy, or platform-native controls. What matters is that limits apply per consumer and can't be bypassed by the agents themselves.