Jev is an artificial intelligence model developed by TypeSafe AI to perform classification and evaluation tasks within software applications.
Unlike the language models used in chatbots, it does not generate text or converse with users. It receives a set of information and directly returns:
- a yes-or-no answer;
- a score;
- a choice from several predefined options.
Each result is accompanied by a probability. The software can therefore use the answer to proceed automatically or, when the model is uncertain, request human intervention.
A model that differs from LLMs
Language models are well suited to explaining, summarising and producing content. In many business processes, however, there is no need to generate text: the task is to classify a request, assign a priority or choose which action to take.
Jev was designed for this second type of task. TypeSafe AI defines it as a System One Model: a model specialised in quick, narrowly scoped decisions.
The system receives a “state”, made up of text or structured data, and evaluates it according to criteria specified by the application. The result can be used directly by the software, without having to interpret a discursive response.
What benefits does it offer?
The first benefit is speed. Because it does not have to generate a response word by word, Jev can complete an evaluation quickly. Multiple questions can also be processed in parallel using the same data.
The second benefit is cost. The model processes the input but does not produce text tokens as output. According to tests published by TypeSafe AI on its own decision-making workflows, Jev can be up to 193 times faster and 444 times less expensive than the language models used in the comparison. These figures, however, are claims made by the vendor and refer to specific types of tasks.
Another important aspect is uncertainty management. Each decision is associated with a probability, so the application can set a threshold above which it proceeds automatically and send less clear-cut cases to a person.
Jev can still make mistakes. The probability does not certify that an answer is correct, but it makes it possible to design the process with the level of uncertainty in mind.
Some possible use cases
Customer support
Jev can classify tickets and emails, identify the appropriate department, estimate the severity of a problem and recognise specific requests, such as a refund request.
Document analysis
It can check whether certain requirements are met, assign a document to a category or give it a completeness score. Uncertain cases can be forwarded to a human operator.
Oversight of AI agents
Before an agent performs an action, Jev can assess its level of risk. Routine operations can be authorised automatically, while more sensitive ones may require human confirmation.
Sales management
Requests received through the website or recorded in the CRM can be classified by type, priority and level of interest, helping the sales team organise its activities.
Operational monitoring
Reports and data from different systems can be evaluated to distinguish routine events from situations that require intervention.
Will Jev replace language models?
Jev is not intended to replace LLMs, but to perform a different task.
A language model remains better suited when the goal is to understand a complex request, interact with a person or produce content. A model such as Jev can instead step in when information needs to be transformed into a choice that software can act on.
The two approaches can therefore be integrated into the same process: the language model understands and communicates, while the decision model classifies, assigns scores and helps govern automations.