Jev: the AI that decides instead of talks, and what it means for customer contact

What happened this month

On 15 September TypeSafe AI, the lab founded by former OpenAI researcher Diogo Almeida, released a model called Jev. It arrived with 40 million dollars in seed funding and a claim that made developers sit up: this model does not write text. It answers a bounded question, in a fixed format, with a percentage attached saying how sure it is. And it does that in 70 to 500 milliseconds.

Vercel put Jev into its AI Gateway and called it the fastest-adopted model in that platform's history. TypeSafe calls it a 'System One' model, after the fast, intuitive mode of thinking in Daniel Kahneman's work. An ordinary language model is System Two: it reasons, writes, weighs things up. Jev does the other thing: look, decide, done.

This piece is not about the technology. It is about why this lands on precisely the decision that AI customer contact has been getting wrong for years.

What a System One model actually is

You give the model two things: the situation (an email, a call transcript, a log file) and a question with a limited set of possible answers. The model picks one and says how confident it is. Nothing more.

Jev takes three kinds of question. A choice: which of these five departments should have this? A score: how urgent is this on a scale of 1 to 10? And a yes or no: is this customer angry? The answer always fits the format you defined in advance. A sentence like 'I think it might be accounts, but it could also be' never comes out, because that is not a valid answer.

That is the whole idea. Software can read a percentage and act on a choice. Software cannot read a paragraph. An ordinary language model produces paragraphs, and that is where the friction has been for years: you ask a chatbot to decide something and get a story back that you then have to interpret.

Why 'cannot hallucinate' means something different from what you hope

TypeSafe says Jev cannot hallucinate, and that is true in one precise sense: the model cannot invent an answer outside your list. Ask 'which department' with five options and you will never get a sixth.

It can still pick the wrong one of the five. The difference is what happens next. A language model that is wrong sounds exactly as convincing as one that is right. Jev attaches a number. At 96% confidence you let the software act. At 61% you let a person look. You set that threshold, per decision, and that is worth more than the infallibility nobody can deliver.

TypeSafe also says, in its own launch post, that the speed and cost figures probably sit at the high end of what you will see in practice, and that the model may carry bias. An AI lab writing that into its own announcement is news in itself.

The hardest decision in AI customer contact is not the answer

Anyone who puts an AI assistant on the phone, in web chat or on WhatsApp thinks first about one question: can it answer well? After a year of real use, that is no longer the problem. The problem is the other question, the one that comes up hundreds of times a day: does this need a person now?

A caller with a deadline expiring. A customer asking the same thing for the third time. A message at eleven at night with the word 'urgent' in it. Each of those is a decision with a limited set of outcomes: handle it, transfer warm, log a callback request. It is exactly the kind of question Jev was built for, and exactly the kind of question you still have a language model write a paragraph about today.

In Heyloha you set, per contact and per time slot, whether the assistant transfers warm or cold, and what happens when nobody picks up. The assistant recognises urgency from your instructions. That works, but the underlying decision comes from the same model that is holding the conversation. System One models split that apart: one model talks, another decides, in half a second and with a confidence score you can set as a threshold.

To be clear: Heyloha does not run on Jev today. The model is two weeks old and in early access. But the direction is clear, and we are not the only ones watching it.

What it does to cost and speed

TypeSafe charges 42 dollars per billion input tokens and nothing for output. In its own words, too cheap to meter. Having an ordinary language model make a decision costs a few seconds and a few cents. Having a System One model make it costs half a second and almost nothing.

That sounds technical, but it changes what is feasible for a company with twenty staff. Having every incoming call assessed for urgency before anyone picks up was, until now, something for call centres with an IT department. If that assessment costs nothing and keeps nobody waiting, it becomes something you simply switch on.

The same goes for checking afterwards. A second model that tests every AI answer before it goes out (does this match the knowledge base, is there a price in here that does not exist) was too slow and too expensive to do by default. At 200 milliseconds a check, it no longer is.

What to do with this now

You do not need to pick a model. Three things you can do, whichever supplier you use.

Write your transfer rules down as decisions with a limited set of outcomes. Not 'transfer important customers', but: which situations, which outcomes, which threshold. If you cannot write it as a choice between five options, no model can decide it for you.

Decide, per decision, how sure the AI has to be. Booking an appointment can go ahead at 80%. Marking a complaint as resolved only at 98%, otherwise to a colleague. Those numbers are policy, not technology, and they belong with you.

Keep a person on the switch for anything with consequences. TypeSafe says it too: the model gives a probability, the application decides what happens with it. That is not a limitation of the technology, that is how it should be.

See it in practice

Want to see how an AI assistant already decides today when to transfer, recognises urgency or logs a callback request? Book a thirty-minute demo and we will walk through your transfer rules together. Or try Heyloha free for fourteen days on your own number and website.

Frequently asked questions about Jev and System One models