Jev is a new type of AI model designed not to generate language, but to make fast, structured decisions. Juan César Jover Sanz-Pastor, AIOps Engineer in Sopra Steria Spain, looks at where specialised models could fit into the emerging AI stack.
Generative AI has taught us to expect machines to talk. We ask a question and get an answer. The same models can draft reports, write code, summarise documents and hold a reasonably convincing conversation.
But much of the software behind everyday services has no need for conversation. It needs to make small choices, quickly. A call reaching a customer service centre has to go to billing, technical support or cancellations. A bank has to decide whether an unusual transaction should pass, be reviewed or be stopped. None of these tasks requires an eloquent answer. The software needs a choice it can act on.
Jev is designed for that kind of work. Developed by TypeSafe AI, founded by Diogo Almedia, it is the first model in a category the company calls System One Models. Rather than producing prose, Jev receives information, evaluates it, and returns a structured decision that another piece of software can use.
Predictable, not infallible
Companies have been automating decisions for decades. Rule-based systems and statistical classifiers work well when the input is clean and the possible cases are understood. Human language, however, is rarely clean. People make mistakes, omit details, use ambiguous expressions and describe the same problem in different ways.
Large language models cope much better with that variation, but they were built to generate language. Modern LLM platforms have become considerably better at structured output, so this is hardly an unsolved problem. There is, however, something slightly roundabout about using a model designed to generate text when the application only needs to select one of four departments.
Jev approaches the task more directly. A developer can provide a customer message and define the available answers in advance: billing, technical support, cancellations or other. Jev evaluates the message and chooses within that set.
The restriction applies to the format of the answer, not to its correctness. Jev may still send a billing enquiry to technical support. What it cannot do is return an unexpected paragraph when the application is waiting for a department. Jev has not made AI deterministic. It has made the possible outputs predictable.
It can choose from a predefined list, place something on a scale or return a probability for a yes-or-no question. Confidence information can accompany the result, so uncertain cases can be treated differently. A routine request might continue automatically, while an ambiguous one could be sent to a larger model or a person.
Confidence remains an estimate, not a guarantee. It has to be tested against real examples from the environment where the model will operate. The useful part is that this uncertainty becomes easier for developers to manage.
The economics of small decisions
Speed and cost are the other parts of the argument. A generative model produces its answer token by token. Jev does not write an explanation or generate a block of JSON to be parsed and checked. It returns the result directly.
TypeSafe reports response times in the tens or hundreds of milliseconds for its intended workloads, although those figures come from the provider and real performance will depend on the input, infrastructure and network. Its pricing follows the same logic: the company charges for the information sent to Jev, while the structured output has no additional cost.
Saving a fraction of a cent on one decision is unremarkable. Across several million, it becomes more meaningful, especially if many would otherwise have gone to a much larger model.
A traffic controller for AI
AI agents provide a useful example. An agent makes many small choices while completing a task: selecting a tool, retrying a failed step, assessing risk, discarding irrelevant context or deciding whether a problem warrants a more capable model.
Large generative models handle many of these choices because they are versatile and already part of the system. They can do the job, but that does not make them the best component for every decision. Putting an expensive frontier model in charge of routine routing is a little like asking a neurosurgeon to sit at reception and direct visitors to the correct floor. They will probably get where they need to go. It is still an odd use of the neurosurgeon.
This is the strongest case for Jev, in my view. Its value may lie in deciding when GPT, Claude, Gemini or another generative model is actually needed.
A more divided AI
The industry has tended to treat the large language model as the centre of everything. It writes, classifies, plans, calls tools, checks results and chooses the next step. There are good reasons for this. One capable general-purpose model is easier to experiment with than half a dozen specialised components.
At scale, though, that convenience becomes expensive. It also concentrates generation, reasoning, routing and execution in the same component. Jev suggests a more divided architecture: fast models handling narrow, repetitive choices, while larger models are reserved for complex reasoning, coding, explanation, open-ended analysis or cases where the possible answers cannot be defined in advance.
Whether Jev itself becomes widely adopted is an open question. It is new, and companies will need to test its confidence estimates on their own data. Adding another model and provider also introduces complexity and another potential point of failure. Lower inference costs do not automatically mean a simpler system.
Even with those reservations, the underlying idea makes sense. AI systems may work better as a collection of specialised components rather than relying on a single, increasingly powerful model for every step.
In that arrangement, Jev would not be the brain of the system. Its role would be more modest, although perhaps just as important: deciding which brain should handle the next problem