Decision Model
What it means in practice
Many AI systems use a large language model for jobs that are really just choices: is this email a complaint, which team should this ticket go to, does this draft breach policy, should this agent be allowed to take this action. A language model can do all of these, but it does so by writing an answer word by word, which is slow and expensive when it happens thousands of times a day.
A decision model skips the writing. The question and its permitted answers are set up front, and the model reads the input once and scores every option in a single pass. Because nothing is generated, answers come back in tens of milliseconds and at a fraction of the cost. And because each answer arrives with a confidence score, the system can be told what to do when the model is unsure, such as passing the case to a larger model or to a person.
The category took its name in September 2026 when TypeSafe launched Jev, and within weeks Cloudflare (Clef), Amazon (Strands Decider), OpenAI (Decisions API) and Perplexity had released their own. Several are open-weight, so they can run inside an organisation’s own environment.
Where it fits
Decision models suit high-volume, well-defined judgements that sit between the steps of a workflow such as routing requests, triaging cases, choosing which tool an agent should use, checking outputs against a rubric and acting as a guardrail before an action is taken. Open-ended work such as drafting, summarising or reasoning through an unfamiliar problem stays with generative models.
Insight
The lasting significance lies in the pattern these models make possible. Well-designed systems are starting to cascade. A small, fast decision model handles the bulk of routine calls and only the uncertain or high-stakes cases escalate to a frontier model or a human reviewer. The confidence thresholds that govern those escalations are business decisions as much as technical ones, because they set how much risk the organisation accepts in exchange for speed and cost. Where those thresholds sit, and who owns them, belongs in the design of the workflow (rather than being left as a setting someone adjusts later).
For leadership teams it is also a reminder that capability increasingly comes from how models are combined. A system that sends every step to one general-purpose model will often be slower, costlier and harder to audit than one that matches each step to the right kind of model.
Misconceptions
Decision models don't replace language models. They can only choose between answers someone has already defined, so the quality of the questions and options matters as much as the model itself.
They aren't a wholly new idea either. Classification models have existed for decades, and some practitioners argue the label is fresh packaging. What has changed is that these models read natural language and context with something close to frontier-level understanding while keeping a classifier's speed and price, and they plug directly into agent workflows.
The early accuracy figures deserve caution. Most benchmarks so far have been published by the vendors themselves, so testing on an organisation's own cases, including whether the confidence scores reflect real-world accuracy, should come before relying on one.