Explained simply · Free to read · 3 min read

Decision model explained simply: the AI that doesn't write but chooses

A decision model is an AI that does not produce text but chooses between given answers and states how confident it is. The idea became known in September 2026 with Jev from TypeSafe AI, and OpenAI, Microsoft, Amazon, Cloudflare and Perplexity now offer such models too. They are much faster and cheaper than chatbots and usually sit invisibly inside software, for example when sorting requests or checking AI agents (as of October 2026).

Published 15 October 2026

In short

Decision model: A decision model is an AI that does not produce text but, for a question with fixed answer options, outputs a probability for each option. This makes it very fast and cheap, but it can only choose between the given answers.^1,2,3,5; Example: An online shop receives thousands of customer messages every day. A decision model assigns each one to the returns, payment or shipping team in a fraction of a second. If it is less than 80 percent sure, an employee reads the message.; Also called: System One model · Entscheidungsmodell · Jev alternative · Decision AI

What a decision model does differently

A language model such as ChatGPT writes its answer word by word. That is flexible but slow and expensive when all you need is a short decision. A decision model works differently: it receives a situation, such as a customer message, and a question with fixed answer options, such as “Which team is responsible: billing, sales or technical support?”. As a result, it outputs a probability for each option.

There are three typical question types: a choice among fixed options, a rating on a scale and a yes-or-no question for which the model states the probability that the statement is true. Because the model does not have to formulate anything, a single computation pass is enough. The provider TypeSafe AI gives response times of 70 to 500 milliseconds for Jev.

TypeSafe calls such models System One models. The name recalls the fast, intuitive thinking known in psychology as System 1, as opposed to slow, deliberate thinking.

What it is used for

Decision models are worthwhile wherever very many small, recurring decisions arise. Typical examples are sorting support requests, pre-sorting job applications, spotting suspicious payments or rating security alerts.

They play a growing role in AI agents. Before an agent does something, a decision model can check in a fraction of a second whether the next step is permitted and sensible. The software company Workday also uses Jev to assemble the screens of its AI assistant Sana in real time from ready-made building blocks.

The probability is particularly useful: companies can decide that only very confident cases are handled automatically and everything uncertain goes to a person. That requires the model to be well calibrated, though. If it says 80 percent, it should be right in about 80 percent of cases.

Which decision models exist

Jev from TypeSafe AI came onto the market on 15 September 2026 as the first well-known model of this kind and is far ahead in usage on the OpenRouter platform. Many competitors have appeared since.

Available as online services are, among others, OpenAI's Decisions interface based on GPT-6 Luna, Microsoft-Decision-1 and Inception's Mercury Decide. Open for download are, for example, Amazon's Strands Decider 2B, Perplexity's Decider V1.1, H2O-Lightning-4B, Kev and Cloudflare's Clef models. The open models can run on your own servers, so data does not leave the company.

Prices are low: for the hosted offerings, Forbes gives 2 to 10 US cents per million input tokens, and outputs cost nothing. A detailed comparison is in our overview of the new decision models.

Where the limits are

A decision model can only choose what it is offered. It is unsuitable for open questions, explanations or new ideas. That is why many developers combine both: the decision model handles the clear-cut cases, and when it is unsure, a conventional language model takes over.

Decision models make mistakes too. TypeSafe writes in its own documentation that Jev struggles with counting, arithmetic and dates, tends to favour the first answer listed and does not automatically recognise instructions smuggled into data as an attack. For decisions about people, money or security, it should therefore always be checked how often the model is wrong and who notices the mistakes.

For you as a user: you usually do not operate a decision model yourself. It works in the background of programs and services you use, and at most you will notice it because they respond faster.

Sources

Created with AI assistance and editorially approved before publication.