Definition
What is a decision model?
A decision model is an AI model built to answer one bounded question about a piece of input and return a typed result instead of free text: a choice from options you supply, a score on an ordered scale, or a probability for a yes-or-no question. Jev from TypeSafe AI and Clef and Clef Flash from Cloudflare are hosted examples.
What does a decision model do?
You send a decision model two things: the input to judge, such as a support message, and a typed question about it with instructions. It returns the answer in the shape the question asked for, and your code branches on that value directly. There is no prose to parse. The options are part of the request, so changing a routing policy means editing the question, not retraining a model. In the support-routing benchmark on this site, three decision models each chose one of five teams for 100 held-out messages. All three returned a valid label for every message, with median response times of 138 to 271 ms in the October 9, 2026 run. OpenAI’s Decisions API also returned a valid label every time, with a median of 111 ms.
What kinds of question can it answer?
The hosted models measured here accept the same three question types. Each returns a value your code can use without parsing text.
- Choice
- One option from a fixed set you supply. Example: Route a support request to billing, technical, account or sales, or ask for clarification.
- Probability
- The probability that a yes-or-no statement is true. Jev and Clef call this type a noul; OpenAI’s Decisions API calls it a predicate. Example: Is this message about an active outage affecting several customers?
- Score
- A position on an ordered scale you define. Example: Rate the customer impact from no impact to critical.
The support-routing recipe shows a complete choice question, and the playground has an example of each type.
How is it different from an LLM, a classifier or a rules engine?
All four can turn an input into a decision. They differ in what you give them, what comes back, and what it takes to change the set of possible answers.
| Approach | What you give it | What comes back | To change the answers |
|---|---|---|---|
| Decision model | The input to judge and a typed question with its options and instructions | One typed value, with a probability for each option | Edit the question |
| General LLM with structured output | A prompt and a JSON schema | Generated text constrained to the schema. A reasoning model spends output tokens on reasoning first | Edit the prompt or the schema |
| Trained classifier | The input only. The labels were fixed when it was trained | A label from the trained set, usually with a score | Collect labeled examples and retrain |
| Rules engine or DMN table | Structured fields | The outcome of the rules that matched, the same every time | Edit the rules |
Is this the same as DMN?
No. Decision Model and Notation (DMN) is a standard from the Object Management Group for writing business rules as decision tables and diagrams that a rules engine runs. People author a DMN decision model, and it gives the same output for the same structured input every time. The decision models on this site are machine-learning models that read unstructured input such as a customer message.
In decision analysis, a decision model can also mean a formal model of options and outcomes, such as a decision tree. That is a different thing as well. The approaches can work together: a model's typed answer can be one input to a rules table.
Which decision models are available?
These are the models in the catalog. Descriptions and access come from each provider's documentation, checked October 9, 2026. All of them have measured results here.
- Clef Flash
- Cloudflare · Public API. A 9B decision model for typed choices, probability checks, and scoring. Hosted on Workers AI. Measured on support routing.
- Clef
- Cloudflare · Public API. A 27B decision model with the same typed question interface. Evaluate it on the decisions your application actually makes. Measured on support routing.
- Jev
- TypeSafe AI · Public API. TypeSafe’s System One model. Turns application state and typed questions into structured decisions. Measured on support routing.
- Decisions API
- OpenAI · Public beta. OpenAI’s decisions endpoint, in public beta and served by GPT-6 Luna. Returns a probability, a choice or a score for text and image input. Measured on support routing.
When does a decision model fit?
A decision model suits a step where the possible answers are fixed, your code needs a typed value, and response time matters. It is not a replacement for a general LLM on every task. In the benchmark here, a structured-output LLM baseline had the highest operational success, 98% (93–99), with a median response time of 1,325 ms. The three decision models scored 92% (85–96) for Jev, 92% (85–96) for Clef, 87% (79–92) for Clef Flash, with median response times of 138 to 271 ms. OpenAI’s Decisions API scored 94% (88–97) with a median of 111 ms. The intervals describe uncertainty in each model's observed rate. This report does not include a paired statistical comparison of the differences between models.