Most AI interactions end with a paragraph of text. You ask your favourite LLM a question or to help you with a task. You chat about it. It shows you how its thinking. You marvel at how clever it all seems.
That can be useful when you need ideas, a first draft or a better way to explain something. For that type of work, the loose, conversational style ends up working against you.
It is slow and can get expensive.
When software needs to make a limited decision and pass the result into a controlled workflow, we dont want complexity, slow responses or unnecessary cost. This is something we do A LOT so we have been experimenting with cheaper models, output formats (json) and much prompt tuning to try and get as standard an approach as possible.
So when the announcement for Jev came out last week, it immediately got our attention.
Jev is TypeSafe AI’s first System One model. Instead of generating an answer in prose, it takes a defined set of information and returns a structured decision with a probability. You choose the question and the permitted answers. The result might be a category, a score or a yes-or-no judgement that another system can inspect.
There is a big difference between asking AI to ‘tell us what this means’ and asking it to choose from a clear, limited set of options.
And it can be easily checked.
The benefit is a decision you can use
A normal language model is good at producing text. It can explain, summarise and suggest. It can also make a confident case for an answer that should have been treated with more caution.
Jev is built for a different job. It returns typed answers and probabilities rather than a generated explanation. That gives software something more practical to work with.
If a workflow needs to know which team should handle a request, whether a message is asking for a refund or how strongly it matches a defined condition, the answer can be constrained before the model sees the question. Code can then combine that answer with normal rules, records and checks.
That does not make a decision automatically right. TypeSafe is clear that calibration is measured across groups of predictions, not guaranteed for an individual answer. The probability is useful because it gives a workflow a sensible place to pause. A high-confidence, low-risk case may take the standard route. A weak or uncertain result can go straight to review.
Fast judgement still needs slower judgement
TypeSafe described Jev as their first System One Model. That term triggered a memory so i grabbed my copy of Daniel Kahneman’s Thinking, Fast and Slow – he describes System 1 as fast and intuitive, and System 2 as slower and more deliberate.
TypeSafe’s ‘System One’ label describes a technical design for fast, structured decisions. It is not a claim that a model thinks in the same way as a person. The parallel is still helpful: some work benefits from a quick first judgement, while other work deserves time, context and a proper human check. (or a System 2 frontier LLM model from Claude or OpenAI)
The real power comes when you use both.
A business can use a fast, limited decision to reduce repetitive sorting. It should still decide where a person must step in, what the model is allowed to influence and what must never happen without a proper review.
Where that can help
There are plenty of small decisions that slow down a process because someone has to make the same first assessment over and over.
A request arrives. It needs a category. It may need a priority. It might need another piece of information before anyone can act. None of that requires a long essay. It requires a clear choice, a record of how certain the system is and a sensible route when the answer is unclear.
That can make work more consistent. It can reduce unnecessary hand-offs. It also gives the person reviewing the case a shorter, clearer starting point rather than an vague AI-generated reply.
The controls matter as much as the model. Defined answer choices, confidence thresholds, deterministic checks and a clear review route are what turn a quick judgement into something a business can use responsibly.
How we are trying it out
We are exploring Jev in a deliberately narrow internal test to see whether it can help with the early classification of support requests.
The aim is not to replace the person handling the work or to build a customer-facing chatbot. We are testing whether a structured first pass can help identify the likely category of a request, show when key information is missing and route uncertain cases for review.
It does not make website changes, close tickets or take action on its own.
We think is the right place to start. A useful AI tool should make a small, measurable part of the process easier to handle, while leaving accountability with the people who own the service.
If you need ongoing help keeping your website looked after, from real experts, our Website Care Plans explain how we handle maintenance, monitoring and support. You can also see how we work on business websites before deciding whether GorillaHub is the right fit.
