What is a decision model?

Take a support message that says, “I was charged twice.” A decision model could return a billing category, a refund-request probability, and an urgency rating. Those are inputs to a workflow, not the workflow’s final action.

A person or a generative model can write the reply. Your policy and integration decide whether to authorize a refund. Classification alone doesn't authorize a refund.

A decision model can be asked for a defined choice, score, or yes-or-no probability. The surrounding software can check that the result has an expected shape before it routes work or combines it with other rules.

That structure doesn't make the classification correct. Generative models can also return structured output. The useful distinction is whether a model and workflow focus on focused, inspectable decisions rather than a free-form reply alone.

Jev is TypeSafe AI’s hosted System One model. Its documentation describes text and JSON inputs and typed choices, ratings, and yes-or-no probabilities. Customers shape its behavior with the context, instructions, criteria, and workflow in a request rather than per-account fine-tuning.

Laya is Convai Innovations’ Apache-2.0 open-weight model family. Its documentation describes an encoder with a decision head, local inference, and domain fine-tuning tooling. The family includes English, multilingual, and typed-decision checkpoints, so choose and evaluate the specific checkpoint that fits your work.

System One is a vendor term inspired by a psychology metaphor. It doesn't show that a model thinks like a person.

A provider confidence field isn't automatically a calibrated probability that the answer is correct for your task. Evaluate it on held-out, representative examples and examine the mistakes that matter before using it to route work automatically.

The same applies when the input, model version, questions, or workflow changes. Keep human review where the consequences require it, and use feedback to test a proposed improvement rather than assuming the underlying model learns from every correction.