Jev AI helps software choose from a fixed list. But what happens when an email is both a receipt and a request for a reply? A permitted answer can still send that message to the wrong place.
That is the useful question behind Jev, TypeSafe’s model for structured decisions. You give it information and defined questions. It returns values software can use, rather than a written conversation. TypeSafe’s introduction explains that interface.
This guide uses a fictional inbox to make the difference clear. Then there is a short worksheet you can copy to plan your own sorting job. I have not connected Jev to an inbox or measured its accuracy.
Watch the Jev AI explanation
What Jev returns instead of a chat reply
Jev’s Choice question selects from options you supply. Its response includes the choice, probabilities for the options and a confidence score. Those are pieces of data the surrounding app can use to decide its next step.
For an inbox, the options could be Needs reply, Receipt and Newsletter. The model’s job is to choose a label. Reading your email account, moving messages and sending replies would require separate app behavior and permissions.
TypeSafe calls Jev a System One model. The name describes its approach to structured decisions. It does not mean the model understands your priorities without being told what they are.
Three fictional emails, one rule that matters
Meet Sam. He has time to answer one email before work and wants to find the message that needs a reply. All three messages below are made up. The labels are intended answers under our rule, not observed Jev results.
The rule is simple. A request for a reply takes priority over a receipt. Otherwise, a payment record belongs in Receipt, and a regular update belongs in Newsletter.
- A payment receipt. “Your payment has been received. Here’s your receipt.” The intended label is Receipt. There is no request for Sam to answer.
- A gardening update. “This week’s gardening tips.” The intended label is Newsletter. Sam can read it later without keeping someone waiting.
- An order receipt with a deadline. “Please confirm your delivery address by five today so we can send your order.” The intended label is Needs reply. The request wins, even though the subject says receipt.
The third email is the test worth thinking about. Looking for the word “receipt” would miss the action Sam needs to take. Real messages have an annoying habit of doing two jobs at once.
Now imagine a result of Receipt with low confidence. That result is invented for this explanation. In our example, the app keeps uncertain messages visible for Sam to review. He sees the address request, changes the label and writes the reply himself.
Three separate decisions happened. The model chose a label. The app applied a review rule. Sam decided what to do with the email. Keeping those responsibilities clear matters more than making the sorting animation look clever.
Confidence helps with review, but it is not proof
A confidence score gives the app a signal to work with. According to TypeSafe’s confidence documentation, the score summarizes the distribution of probabilities across the possible answers.
When the probabilities are spread across several options, the choice is less clear. The app could flag the result, ask for more information or send it to a person. TypeSafe describes those as patterns the surrounding code can implement.
My recommendation for Sam’s example is to keep important messages visible even when the model reports confidence. Otherwise, a confident mistake could skip the very review meant to protect the deadline.
There is no universal threshold I can honestly give you for a real inbox. You would need examples from that inbox, known answers and a way to measure which mistakes slip through. This fictional story provides none of those measurements.
Jev and ChatGPT serve different parts of the job
A chat tool is useful when you want a draft reply, an explanation or a conversation that changes direction. A fixed-choice interface fits a narrower request, such as selecting one of three labels for software to use.
Chat models can also produce structured decisions. TypeSafe’s launch explanation acknowledges structured values among their possible outputs. Jev specializes in that kind of interface rather than generating prose.
For Sam, choosing Needs reply and writing a polite address confirmation are different jobs. A model that fits the first job does not automatically fit the second.
TypeSafe also makes speed and cost claims based on its evaluations. Those claims do not tell us how Jev would perform on your emails. This article includes no hands-on speed, cost or accuracy comparison.
A decision worksheet you can copy
Start by defining the job before choosing a model. Copy the six prompts below into a note and fill in your answers. You do not need an account or code to spot an unclear rule.
- What single decision needs an answer?
My decision is ______.
Sam’s example asks which tray an email belongs in. - Which answers are allowed?
My options are ______.
Include a plan for an input that fits none of them. It could go to review rather than being forced into a misleading category. - Which rule wins when categories overlap?
When both ______ and ______ apply, choose ______.
For Sam, a reply request takes priority over a receipt. - What would a wrong answer cost?
The mistake I most need to catch is ______.
Missing a delivery deadline is different from filing a gardening update in the wrong tray. - What must stay visible for a person?
A person reviews ______ before ______.
Include the important cases that need review regardless of the model’s confidence. - Who may take the final action?
The app may ______. A person must approve ______.
Suggesting a label does not need to grant permission to send a reply.
Try the rules on a few tricky examples you can judge yourself. An overlapping category, a missing detail and a request with a deadline will often tell you more about your instructions than a perfectly tidy sample.
For the next step, my guide to AI agent boundaries looks at where an assistant’s task and permissions should end. The ToolGrad explainer covers how models can be trained from working tool calls.
The point I would carry forward is modest. A fixed list makes an answer easier for software to handle. Whether it is the right answer still depends on the question, the information and the checks around it.
Product documentation checked September 22, 2026. Sam, the messages and the low-confidence result are fictional. The worksheet is editorial guidance, not a tested Jev integration.
