Most of the AI tools you have tried write. You ask a question, and you get paragraphs back: an email draft, a summary, an answer to a customer.
But a lot of what happens in a business is not writing. It is deciding. Is this ticket urgent? Which team gets it? Does this invoice look right? Is this lead worth a call today? These are small questions with a short list of possible answers, asked hundreds of times a week.
In September 2026 a company called TypeSafe AI released Jev, a model built for exactly that kind of question. It shows where AI is heading: different models for different jobs.
#What Jev is, in plain English
Jev does not write. Your software sends it a structured question and the options it is allowed to choose from. Jev returns one of those options and a confidence score from 0 to 1. It cannot reply with a sentence or invent a fourth option.
Its main question types are:
- Choice: pick one option from a list you provide (“urgent”, “normal”, “low”).
- Score: rate something against a rubric you define.
- Yes or no: return a probability that something is true.
Because it only answers in that fixed shape, it is fast and cheap to run. In TypeSafe’s own tests it answers in well under a second and costs a small fraction of what a large writing model costs for the same question. Those speed and cost multiples come from the vendor; there is no independent benchmark yet, so treat them as a promising claim, not a guarantee.
#Where it fits in a business
The useful question is “which of our tasks are really decisions?” Some common ones:
- Support tickets. Sort every incoming ticket by topic, urgency and team the moment it arrives, so nothing urgent waits in a general inbox.
- Invoices and expenses. Flag the one that does not look like the others (a duplicate, an unusual amount, a supplier you have never paid) for a person to check.
- Leads. Score each new enquiry against what a good customer looks like for you, so the strongest ones get a call first.
- Checking other AI. Before a chatbot’s answer reaches a customer, ask a second question: does this answer stay within our policy? Yes or no.
The last one is easy to overlook. A writing model and a decision model work well as a pair: one drafts, the other checks, and nothing goes out on the first model’s word alone.
#The honest limits
Jev is a tool for a specific shape of work, and its own makers are clear about where it falls short.
- It does not explain itself. You get an answer and a confidence score, not a reason. If you need to show a customer or an auditor why a decision was made, the reason has to come from your rules or from a person.
- It needs well-defined options. If you cannot list the possible answers, it is the wrong tool. TypeSafe’s guidance suggests including an “other” option so it is not forced into a bad fit.
- It has known weak spots. TypeSafe’s documentation lists them: it is not a calculator, it can miscount, it reads dates as text rather than as quantities, it can stumble on double negatives, and it tends to favor the first option in a list.
- There is no public accuracy benchmark. TypeSafe chose not to publish results against standard tests. You find out how accurate it is by testing it on your own examples.
#The smart setup: use the confidence score
The confidence score is what makes a decision model safe to use. TypeSafe’s own guidance is three bands:
- High confidence: act automatically. The ticket gets routed, the lead gets its score.
- Medium confidence: proceed with care. Flag it for review or ask for more information.
- Low confidence: do not act. Send it to a person, or to a larger model that can reason through the hard case.
Where you draw those lines is a business decision: a misrouted ticket costs minutes, a wrongly approved refund costs money. Start cautious, run it on a few hundred of your own past cases, and compare its answers with what your team actually did before you let it act on anything.
#Knowing which AI goes where
The real lesson of Jev is not about one model. It is that “use AI” is no longer one decision. Some tasks need a model that writes, some need one that decides, some need plain rules in code, and some need a person. Putting each task in the right place is where most of the savings and the safety come from.
That is the work we do with clients: map the tasks, pick the right tool for each one, and build the handoffs between them. If you want to see which of your tasks are really decisions, book a free 30-minute call with one of our engineers.
Sources: TypeSafe AI documentation (docs.typesafe.ai: models, confidence and Jev 1.13 limitations pages); DataNorth, “Jev: what it is and why this new model matters”; Zapier, “What is Jev?”; Requesty, “TypeSafe Jev explained”. Speed and cost claims are TypeSafe’s own figures.
