Skip to content
devianai
devianai

This is where the detour starts. You pick the direction.

Let's talk
All articles

6 min read

Not all AI has to chat

There's a new kind of artificial intelligence that doesn't write anything: it decides. It's called Jev. Here's what it is and why it might be more useful to a business than to a tech enthusiast.

JevTypeSafe AI

In short

Jev doesn't write text: it answers with a yes/no, a choice or a score. It's designed for the small decisions that repeat every day.

Name
Jev
Made by
TypeSafe AI, a San Francisco lab
Cost
Reportedly, 42 dollars per billion input tokens; the answer is free
Access
Early access, with a waiting list

Close your eyes and think "artificial intelligence".

Almost certainly a chat came to mind. A white window, you type a question and a long, polite, often well-written answer arrives. In recent years, AI for everyone has become exactly that: something you talk to.

Except that, if you think about it, most of the work in a company isn't talking. Or writing, for that matter.

It's deciding. Lots of small decisions, always the same ones, all day long.

Is this email a complaint or just a question? Is there something odd about this order? Does this message go to the warehouse, to accounts or to customer care? Is this guest just curious, or about to leave us a one-star review?

They're two-second decisions. The problem is there are two hundred of them a day.

What Jev is, in simple terms

In September 2026 a San Francisco lab, TypeSafe AI, presented a model called Jev. Behind it is a former OpenAI researcher who contributed to one of the techniques underlying ChatGPT.

The curious thing is that Jev goes in the opposite direction. It doesn't write text. No summaries, no emails, no small talk.

You give it some information, for example a message or the details of an order, and ask it questions with a precise shape. It answers with a number or a choice. The questions come in three types:

  • Yes or no. "Is this message urgent?" And it tells you how likely the answer is to be yes.
  • A choice between options you decide. "What is it about: check-in, cleaning, breakdown or other?"
  • A score. "From 1 to 5, how unhappy is this customer?"
  • Yes or no

    Is this message urgent?

    Yes, very likely

  • A choice

    What is it about: check-in, cleaning, breakdown or other?

    breakdown

  • A score

    From 1 to 5, how unhappy is this customer?

    A number from 1 to 5

Three questions, three example answers: no sentences, just things the software can use straight away.

It doesn't sound like much. And that's exactly the point.

Not every problem needs a chat.

Why something that does less can be more useful

Anyone who's tried using a chat inside business software knows it: chats are great at talking, but they talk even when they shouldn't.

You tell it "answer only warehouse, accounts or support" and nine times out of ten it does. The tenth, it replies with a kind and slightly hesitant paragraph. Lovely, but your management software has no use for it: it wanted one word.

Or it invents an option that wasn't there.

Jev doesn't do that. If there are three options, it answers with one of those three. Always. And for people who build software that's worth a lot, because the answer can be used straight away, without having to interpret it.

A chat answers like this

"I'd say this message is mainly about support, although it might concern the warehouse too. If you like, I can explain in more detail…"

The software has to interpret a paragraph.

Jev answers like this

Department:
support
Confidence:
high

The software uses the answer straight away.

Same question, two answers. One has to be interpreted, the other is ready to use.

Then there's speed. According to TypeSafe, Jev answers in 70 to 500 milliseconds, less than half a second. The first customers report significant results: Vercel, which builds infrastructure for websites, uses it to check commands before running them; Deel, which handles contracts and payroll for companies worldwide, talks about much lower costs than before.

And it's cheap: reportedly, the announced price is 42 dollars per billion input "tokens" (tokens are, roughly, bits of words), while the answer is free.

An example from real life

Take Marta. Marta doesn't exist, but we know quite a few Martas.

She manages ten short-term rental flats on the coast. In summer she gets dozens of messages a day from platforms, WhatsApp and email. Most are trivial questions: the wifi password, where to park, whether you can arrive an hour early. Every now and then, though, the serious one arrives: "There's no hot water". "The door won't open".

Marta's problem isn't replying. It's noticing in time which message is the serious one, while she's driving from one flat to the next.

With a tool like Jev, every message goes through three questions, all at once, in half a second. Then the software does the rest, with simple rules decided by Marta.

A message arrives

Three questions, in half a second:

  • Is it urgent?
  • What is it about?
  • How unhappy is the guest?
  • Urgent breakdown: alert the technician

  • Wifi question: ready-made answer

  • Unhappy guest: top of the list

Three questions, three roads. The rules are set by the people doing the work, not by the machine.

No robot answering on her behalf with strange sentences. Just a system that sorts the mail the way a good assistant would on their first day, and leaves Marta the things that matter.

The same goes for an online shop receiving orders, a practice receiving emails from a hundred clients, a workshop receiving quote requests. Anywhere someone spends the day doing "this one here, that one there".

Before getting excited

No tool is magic, so here are the limits.

  • It doesn't write. If you also need the reply to send, someone else writes it: a person, a ready-made text or another AI.
  • It doesn't reason at length. It's good for quick, clear decisions, not for problems with many steps.
  • It can be wrong. Jev also tells you how confident it is in the answer, but being confident doesn't mean being right. For costly decisions, a person stays in the middle.
  • It's just come out. Access is through a waiting list. The figures on speed and cost come from the company and its first customers: promising, but we haven't verified them first-hand yet.

In practice, for your business

  • For one week, note down every time someone asks "and where does this go?".
  • If the questions are always the same, a machine can help there: with Jev, with something else or even without AI.
  • For costly decisions, keep a person in the middle.

Our detour

We won't tell you to rush and try it tomorrow. It's not a tool to use on its own: it has to be connected to the software you already use, and that's work to do calmly.

But we like the idea behind it, because it's what we've always said: not every problem needs a chat. Sometimes you need a system that sorts, alerts and puts things in order. Quiet, fast, boring. In the good sense.

An exercise that costs nothing: for one week, keep a sheet of paper next to your keyboard. Every time you or someone in the company stops in front of an email, an order or a message and asks "and where does this go?", make a tick and write down the question.

At the end of the week, look at the sheet. If the questions are always the same and there are lots of ticks, there's work there that a machine can take off your hands. With Jev, with something else, or even without AI.

And if you like, bring that sheet to us. We'll start from there.

Got an idea?

Tell us in a couple of lines: we'll start from there.