Today's LLMs are remarkably capable. But an enterprise automation system does not always need a well-written sentence.
Sometimes it needs to decide which team should receive a request, whether a transaction needs human review, or whether an agent's work can be approved. What matters then is not a long explanation, but a clear decision and a probability.
Until now, even a simple decision often required a large LLM to generate prose, format it as JSON, and let another program parse the result. TypeSafe AI's Jev begins with a different question.
Should a model that answers people and a model that helps software make decisions be designed in the same way?
Instead of generating strings, Jev returns predefined choices and probabilities. Its full architecture and training process have not been disclosed. Yet products and implementations aimed at the same problem began appearing quickly after Jev arrived.
Teams did not need to know its exact internal design to understand the problem it was solving. The important work was not reproducing the same model. It was defining the capability and building it with technology that already existed.
The same perspective applies to enterprise data and ontologies.
Walk into a company and begin by connecting every dataset, defining every entity, and building a complete ontology, and a question will probably surface a few months later.
What can we actually do with this?
The problem is not that ontologies are unnecessary. The problem is the order in which they are built.
Suppose the goal is an agent that detects an incident, identifies affected customers, traces the cause, and asks the right owner to respond. The work the agent must perform becomes visible first. It needs to connect the incident to a service, find affected customers and recent deployments, and identify the team that owns the service.
Only then do the necessary entities and relationships become clear. Incident, Service, Customer, Deployment, and Team now have a reason to exist. The ontology is no longer a vast dictionary describing every piece of enterprise data. It becomes the system of meaning an agent needs in order to do its job.
Data follows the same order. The first question is not whether every dataset is ready. It is what information the agent needs to make a particular decision. Existing data can be connected. Missing data can be sourced externally, extracted from logs, structured with an LLM, entered by people, or created through the agent's own work.
Good data does not always have to exist before the system is built. A useful system can produce better data as people use it.
What Jev reveals is less about a particular model architecture than this order of problem solving. Define the outcome first. Break it into the decisions and actions required. Then work backward to the models, data, and systems of meaning the job demands.
Models, data, and ontologies become materials for solving a problem rather than goals in themselves.
That may be one of the most important capabilities in the AI era: not building everything from scratch, but defining the desired state precisely, assembling what already exists, and creating only what is missing to reach a real business outcome.