News

Microsoft Targets the Small Decisions That Add Up to Big AI Bills

Decision-1.1 is designed to handle routine choices inside AI workflows. Its low token price highlights how much those choices cost once you include errors, retries, and review.

Microsoft is targeting a less visible part of enterprise AI spending: the repeated decisions an agent makes before it produces an answer or completes an action.

The company made Microsoft-Decision-1.1 generally available on November 17, pitching it for routing, classification, prioritization, verification, and workflow control. The model accepts images as well as text and is available through Microsoft Foundry, OpenRouter, and Vercel.

The proposition is that many steps in an AI workflow need a bounded judgment rather than an extended generated response. Using a specialized model for those steps could reduce processing costs and delays, provided its decisions are sufficiently reliable.

“For us, it is about mapping intelligence to a given task, with the best economics,” Microsoft Chairman and CEO Satya Nadella said in the company's edited version of his opening keynote remarks at Ignite in San Francisco.

Decisions Between the Bigger Tasks
A hypothetical support workflow illustrates the distinction. A generative model might draft a reply, while a decision model determines which department should receive the request and whether it needs escalation.

Microsoft's existing Decision-1 documentation describes returning numerical judgments instead of prose. Applications provide defined questions and answer options, then use probabilities, selected choices, or ordered scores to guide subsequent steps.

Those operations can recur within a single assignment. Their economic importance depends on how often they run, how much context each requires, and what happens when the wrong option is selected.

Microsoft says Decision-1.1 runs more than twice as fast as its predecessor at the same cost. It also reports improved calibration, meaning closer alignment between confidence estimates and observed correctness. Those are company-reported improvements, not a guarantee of performance on a customer's workload.

The approach sits within Microsoft's broader push toward specialized models. Mustafa Suleyman, CEO of Microsoft AI, described the strategy in a Reuters interview published November 17.

“This is a system, a hill-climbing machine that we have built, which co-optimizes performance and price,” Suleyman told Reuters.

A Small Price With a Larger Denominator
Microsoft lists Foundry input pricing at $0.042 per million tokens for Global Standard deployments, $0.046 for the U.S. Data Zone, and $0.05 for EU and APAC Data Zones. Output tokens are free.

For illustration, one million requests averaging 1,000 input tokens would consume one billion input tokens. At the Global Standard rate, the model-input charge would be $42. That calculation excludes other models, tools, infrastructure, and operational work.

The comparison also needs an appropriate baseline. A decision that can be handled adequately by an existing rule or conventional classifier may not require another model call. Where a model is useful, a lower request price still needs to be weighed against mistakes and their consequences.

Microsoft Is Using It to Route AI Work
Microsoft identifies an internal use for that approach in its Copilot Cowork announcement.

“On every step, Auto uses our latest decision model, Microsoft-Decision-1.1, to pick the right model and the right amount of reasoning,” the company wrote. The enhancements are available in its Frontier program.

That application makes the model's role more concrete: it can help determine where other AI spending occurs. The announcement does not establish a general percentage reduction in customers' workflow costs.

Microsoft also announced Decision-1.1 as an available judge in Foundry Evaluations. Model-based evaluation can supply another signal about output quality, but it is itself a judgment that requires validation.

Confidence Still Needs a Threshold
In the Decision-1.1 launch post, Microsoft describes training the model to recognize uncertainty and abstain when information is incomplete.

“This allows developers to set their own thresholds for when to act, flag, or escalate,” the company wrote.

The distinction is operational: a probability becomes useful only when an application has rules for responding to it. Sending an uncertain request for review can prevent an expensive mistake, but also introduces time and labor that belong in the cost calculation.

Measure the Completed Workflow
The existing Decision-1 guidance recommends testing on representative labeled examples and setting confidence thresholds according to the consequences of false positives and false negatives. It also describes escalating uncertain decisions to a person or another model.

For buyers, that makes cost per successfully completed workflow a more informative measure than token price alone. A useful comparison would track incorrect routing, repeat processing, escalation rates, and the time people spend correcting results.

Microsoft uses similar language in its broader Microsoft IQ announcement: “Teams define success, compare quality, latency, and cost per successful outcome, and approve changes before deployment.”

That describes an evaluation process, rather than evidence that Decision-1.1 will produce a particular saving. A cheaper decision can improve the economics of an agent. Establishing that improvement requires following the decision through to its outcome.

About the Author

John K. Waters is the editor in chief of a number of Converge360.com sites, with a focus on high-end development, AI and future tech. He's been writing about cutting-edge technologies and culture of Silicon Valley for more than two decades, and he's written more than a dozen books. He also co-scripted the documentary film Silicon Valley: A 100 Year Renaissance, which aired on PBS.  He can be reached at [email protected].

Featured