Microsoft urges businesses to manage AI agents as a continuous discipline

From pilots to production economics

In a blog post dated November 5, Microsoft argues that the central challenge has shifted from proving that AI works to showing that it can scale economically. The company points to findings from the IDC InfoBrief Scaling AI Without Scaling Costs, sponsored by Microsoft, to illustrate the pressure on organizations to demonstrate value.

According to the cited research, nearly 90% of organizations apply financial thresholds before approving AI initiatives. However, only 17% build models to project how costs will scale in production, even though about 78% prioritize cost efficiency when selecting a model. The research also says that 86% expect financial results within six to 24 months.

Microsoft says those figures leave roughly one budget cycle for evidence of value to emerge. It recommends evaluating each proposed agent through four questions covering how to fund, run, improve and prove the system.

Four controls for managing agent costs

The first question is how to predict and track costs at scale. Microsoft recommends projecting usage at three times and 10 times current levels, then tracing spending to an application, team and agent. It also advises comparing models on cost, quality and latency, while using reserved capacity for steady demand.

The second is how to optimize the cost of each request. Microsoft says businesses should measure the cost per outcome rather than the price per token, since one successful result can require a dozen model requests. Routine work should not automatically use a frontier model, and evaluations should confirm that lower-cost choices maintain quality.

The third question concerns continuous improvement. Microsoft identifies retrieved documents, tool definitions, instructions and conversation history as potential sources of unnecessary cost and weaker answers. Retrieving only relevant passages, reducing repeated tool descriptions and retaining only important information from transcripts can lower spending, according to the company.

Finally, Microsoft recommends connecting spending to outcomes such as a resolved case. It says agents should have circuit breakers, evaluation suites, cost and latency budgets, rollback paths and regular reviews.

Microsoft Foundry as the operating platform

Microsoft presents Microsoft Foundry as the platform layer intended to make these controls repeatable across an organization. The company describes Foundry as an enterprise platform for building, grounding and governing AI applications and agents, with security, compliance and policy controls across the agent lifecycle.

Microsoft recommends model routing across the over 11,000 models available in Microsoft Foundry, along with Microsoft IQ to improve relevance and reduce costs. Its proposed operating loop has three parts: Plan, to compare models and capacity options; Control, to monitor cost alongside quality, latency and safety; and Optimize, to route requests to an appropriate model, provide context and retrieval centrally, and detect quality regression.

For users and AI leaders, the message is that scaling from one agent to 20 should not require rebuilding cost tracking and governance each time. Microsoft says platform-level evidence can support funding decisions, while shared controls give teams practical levers to optimize or retire agents as their economics become clearer.