Decision models: a new category of AI for your workflows
A decision model is an AI that writes nothing: it picks an answer from the ones you defined and gives its probability. It sorts, routes and checks faster and for less than an LLM.
4 min read
The SynAI team
Key points
- A decision model is an AI that does not generate text: it answers with a choice, a score or a yes or no, with a probability.
- TypeSafe AI released the first decision model, Jev, on September 15, 2026; OpenAI, Microsoft, Perplexity and Liquid AI followed.
- In a workflow, a decision model handles routing: above a probability threshold, the process continues on its own; below it, a person reviews the case.
What is a decision model?
It is an AI model that chooses from answers defined in advance and tells you how certain it is, without writing any text.
TypeSafe AI released Jev on September 15, 2026. OpenAI introduced its Decisions API on September 29 and opened it to everyone on October 22. Microsoft, Perplexity and Liquid AI followed; Liquid AI also publishes open-weight models that you can run on your own servers.
Are decision models a new idea?
No: the idea goes back at least to the perceptron of 1958. What is new is understanding an instruction in plain language, like an LLM, with no task-specific training.
| Period | What decides | Its limit |
|---|---|---|
| 1958 | Frank Rosenblatt's perceptron classifies an input | Two categories only |
| 1989 | The FICO score helps lenders decide | Structured data |
| 2018 | Google's BERT classifies text | Training for each task |
| 2022 to 2025 | LLMs classify from a simple instruction | Slow and costly for making a call |
Period
1958
- What decides
- Frank Rosenblatt's perceptron classifies an input
- Its limit
- Two categories only
1989
- What decides
- The FICO score helps lenders decide
- Its limit
- Structured data
2018
- What decides
- Google's BERT classifies text
- Its limit
- Training for each task
2022 to 2025
- What decides
- LLMs classify from a simple instruction
- Its limit
- Slow and costly for making a call
How does a decision model work?
It receives some context, such as an email, and questions whose answers are set in advance. It returns one of three formats:
- A choice. One option from your list, with the probability of each: "Billing, delivery or product?"
- A score. A position on your scale: "How urgent, from 1 to 4?"
- A yes or no. The probability of yes, between 0 and 1: "Is the customer asking for a refund?" 0.92.
Decision model or LLM: which one should you use?
An LLM writes; a decision model chooses. According to OpenAI, its Decisions API answers up to 10 times faster than GPT-6 Luna used as an LLM, even though it runs on that same model. TypeSafe reports 70 to 500 milliseconds per answer for Jev.
| LLM | Decision model | |
|---|---|---|
| What it returns | Free text | One of the planned answers, with its probability |
| Input price | $0.10 (GPT-6 Luna) | $0.10 (Decisions API), $0.042 (Jev) |
| Output price | $0.50 (GPT-6 Luna) | Free |
| Strengths | Writing, explaining, planning | Sorting, routing, qualifying, checking |
What it returns
- LLM
- Free text
- Decision model
- One of the planned answers, with its probability
Input price
- LLM
- $0.10 (GPT-6 Luna)
- Decision model
- $0.10 (Decisions API), $0.042 (Jev)
Output price
- LLM
- $0.50 (GPT-6 Luna)
- Decision model
- Free
Strengths
- LLM
- Writing, explaining, planning
- Decision model
- Sorting, routing, qualifying, checking
Prices in US dollars per million tokens, as of November 1, 2026.
A decision model replaces neither the LLM, which writes the reply to the customer, nor your rules: detecting a refund request does not mean granting it.
How do you add a decision model to a workflow?
At SynAI, we give it a single job, routing:
- Code checks the exact facts: order number, amount, date.
- The decision model assigns a category and a probability.
- Above a probability threshold set in advance, your rules allow the action or not; below it, a person takes the file.
- An LLM writes the reply, if there is one.
We use it for four tasks:
- Sorting requests. Each email gets a category and an urgency level, then goes to the right team.
- Classifying documents. An invoice or a contract is recognized before any data is extracted.
- Choosing the right tool. A request follows an existing process or goes to a more powerful LLM.
- Checking AI agents. The model checks that the agent followed the procedure, for example that it asked for the missing information.
Before going live, we test each decision on real examples, ambiguous cases included. We set thresholds based on the cost of an error, and we plan a fallback in case of an outage.
Frequently asked questions
- Does a decision model work in French?
- Yes, but not necessarily as well as in English. According to TypeSafe, English is Jev's primary training language: test it on your own content before you rely on it.
- Are automated decisions allowed in Quebec?
- Yes, with obligations. If a decision rests solely on automated processing of personal information, the person must be informed and be able to reach an employee who can review it (Law 25, s. 12.1). This is not legal advice.
Sources
- 1TypeSafe AI · Introducing System One Models & Jev
- 2TypeSafe AI · Documentation: Models
- 3OpenAI · DevDay 2026 recap
- 4OpenAI · Build more responsive apps with the Decisions API
- 5OpenAI · API Pricing
- 6Microsoft · Introducing Microsoft-Decision-1
- 7Perplexity · Decisions API
- 8Liquid AI · Open d1: Edge decision models for text, vision, and audio
- 9Cornell Chronicle · Professor's perceptron paved the way for AI
- 10FICO · FICO Score Fact Sheet
- 11Google Research · Open Sourcing BERT
- 12LégisQuébec · Act respecting the protection of personal information in the private sector
