Microsoft Decision-1: new AI model targets fast routing tasks

What happened

Microsoft has introduced Decision-1, a new AI model purpose-built for fast classification and routing tasks, marking the company's entry into the rapidly expanding category of "decision models." Built on the Qwen3.5-9B architecture, Decision-1 is designed to make rapid, structured judgement calls rather than generate open-ended text, positioning it as infrastructure for systems that need to triage, sort or direct information at speed.

According to Microsoft's own testing across 36 benchmarks, Decision-1 achieves 83.5 percent accuracy with an average latency of 85 milliseconds. The company frames this combination of speed and accuracy as the model's core value proposition, aimed at use cases where split-second routing decisions — such as directing a query, flagging an exception, or classifying an input — matter more than generating lengthy, nuanced responses.

Why it matters

Decision-1 reflects a broader shift in the AI landscape away from large, general-purpose models toward smaller, task-specific ones optimised for narrow but high-frequency operations. Classification and routing sit quietly beneath many customer-facing and operational systems — determining which queue a support ticket lands in, which workflow a transaction triggers, or which response path a chatbot follows — and a faster, more accurate decision layer could materially improve the responsiveness of the systems built on top of it.

For organisations running AI-driven service or automation pipelines, this signals growing competition and choice in the "decision model" tier of the stack, separate from the large language models used for conversation and content generation. Enterprises building multi-model architectures now have another option to evaluate specifically for speed-sensitive routing tasks, rather than defaulting to a single general-purpose model for every function.

By the numbers

• 83.5 percent accuracy achieved by Decision-1 across Microsoft's internal benchmark suite
• 85 milliseconds average latency reported for the model's decision outputs
• 36 benchmarks used by Microsoft to evaluate the model's performance
• Qwen3.5-9B the underlying architecture Decision-1 is built upon

The Renascence take

It's tempting to read Decision-1 purely as a speeds-and-feeds story, but the more interesting implication sits in how invisible this layer is to the people it ultimately affects. Decision models don't write the message a customer reads — they decide which message, which queue, which path that customer gets routed into, often without anyone noticing the choice was made at all.

Most organisations obsess over the model customers can see — the chatbot, the generated reply — while underinvesting in the quieter routing logic that decides who gets fast-tracked and who gets stuck in a loop. A 85-millisecond decision is only a win if the decision itself reflects the right priorities, not just the fastest ones; speed without a clear service-design rationale behind the routing rules just lets organisations make bad calls more efficiently. Before adopting a dedicated decision layer, experience leaders should map exactly which judgement calls it will be making on their behalf — and audit those rules with the same rigour they'd apply to a human agent's training.