What is a judgement model?

Imagine dozens of customer enquiries, invoices or notifications arriving every day that need sorting before anyone can work on them. At its core, that sorting work is about choosing: which category does this fall into, who should it go to, how urgent is it? A newer category of AI models exists for exactly those steps: judgement models, also called beslissingsmodellen in Dutch.

A judgement model receives a fixed set of possible answers in advance, such as a list of categories, a scale or a yes/no. The model chooses from those options and returns a confidence score with each choice: a number indicating how certain it is of its answer. By design, it produces nothing beyond that choice. The answer is always a valid value that your software can process directly.

How it differs from a large language model

A large language model such as ChatGPT or Claude generates free text, allowing it to write, summarise and reason. That makes it powerful but also relatively slow and expensive, and software has to interpret the answer again. A judgement model cannot do any of that writing, which is precisely why it suits decisions in the middle of a work process: it is designed to respond in milliseconds and cost a fraction of a language model call per decision.

TypeSafe AI, the maker of Jev, calls these models System One Models, after Daniel Kahneman’s distinction between fast, intuitive thinking (System 1) and slow, deliberate thinking (System 2). According to the company, Jev is the first model in that class.

What it is used for

Typical applications include classification (which category does this document fall into), routing (which department or system should this request go to), scoring (how urgent is this notification) and assessing the output of other AI models, so you can automatically check whether an AI answer is good enough to use.

In practice, the confidence score is the most valuable part. You set a threshold: confident cases are processed automatically, while uncertain cases go to an employee. That keeps a person responsible for precisely the cases that need one.

Here is a concrete example. An accounting firm receives purchase invoices, each of which must be assigned to a general ledger account. A judgement model receives the invoice text and the list of permitted accounts, and chooses one with a confidence score. Invoices above the threshold are prepared for processing; the rest go into a review queue for the bookkeeper. The bookkeeper then focuses only on the cases that deserve attention.

In setups with AI agents, judgement models often work alongside large language models: the judgement model sorts and routes at the start, and the language model handles the reading and writing afterwards. You can read what that division of responsibilities looks like in a complete process in our guide to AI agents for SMEs.

Hosted or self-hosted

Like language models, judgement models come in two forms. Jev is a hosted API: you send your text to TypeSafe’s servers, in this case in the United States. That is the fastest way to get started, but it does mean your data leaves your environment. For sensitive data, you need to make that decision deliberately and put data processing agreements in place.

OpenAI has also entered this market: at Dev Day (29 September 2026), it announced the Decisions API, built on GPT-6 Luna. It can assess images as well as text, but is currently only available in closed preview and does not yet publish a confidence score.

The open-source Laya runs on your own infrastructure, so nothing leaves your environment. The cost is work: the model first needs to be fine-tuned for your own area of application before it is accurate enough to be useful. The right option depends on the sensitivity of your data and how much evaluation and maintenance work you can take on yourself.

Limitations

A judgement model does not formulate answers or give explanations, and generally only reads text. TypeSafe’s claim of “0% hallucinations” means the answer is guaranteed to come from your options; that claim says nothing about whether the choice is factually correct. Current models are also often trained primarily on English, so testing on your own material is especially important for Dutch-language applications.

For the sorting work described at the start, the principle still holds: automate what is certain, and let an employee review uncertain cases.