Skip to content
techguy:~$ — Tenjen Sherpa, home
← ~/blog#ai#it-support#automation

Jev: an AI model that doesn't chat

TypeSafe AI's Jev skips text generation and returns typed decisions with confidence scores. Here's what it is, and why it caught my eye from an IT support point of view.

Tenjen Sherpa4 min read
$ cat toc.md +
  1. 01What Jev actually is
  2. 02The numbers (as claimed by TypeSafe)
  3. 03Why an IT support person cares
  4. 04The catch
  5. 05My takeaway

Most AI news this year has been about chatbots getting bigger and more talkative. So a model that can’t hold a conversation at all stood out to me. It’s called Jev, it comes from a San Francisco start-up called TypeSafe AI, and it went into limited early access on 15 September 2026.

I haven’t used it myself yet, since access is by waitlist. But the idea behind it maps surprisingly well onto the kind of work I do in IT support, so I wanted to write down what it is and where I think it could fit.

What Jev actually is

Jev isn’t a large language model (LLM) in the usual sense. You don’t get a paragraph of text back. Instead:

  • You define the possible answers up front. For example: a fixed list of categories, a numeric score, or a true/false check.
  • Jev picks from those options and returns a typed value (the answer in a format software can use directly).
  • Every answer comes with a probability and a confidence score, so your code can tell how sure the model is.

TypeSafe calls this a “System One model”, a nod to the fast, intuitive System 1 thinking described by psychologist Daniel Kahneman. The pitch is that a lot of real-world AI work isn’t writing essays; it’s making quick, repeatable decisions inside other software.

The numbers (as claimed by TypeSafe)

These figures come from the company’s own launch post, so treat them as marketing claims until independent tests appear:

Response time
70–500ms
per decision
Speed
40–200×
faster than frontier LLMs on structured tasks
Price
$0.042
per million input tokens (USD); output is free

Company-reported figures · TypeSafe AI launch post, Sept 2026

Because Jev only chooses between answers you’ve defined, TypeSafe also says it can’t produce the “hallucinated” free text that LLMs sometimes do. That makes sense by design, although the model can still pick the wrong option; that’s what the confidence score is for.

Why an IT support person cares

A huge amount of help desk work is classification. Before anyone fixes anything, someone has to answer questions like:

  • Is this ticket about hardware, software, network, access or security?
  • How urgent is it: is one person affected, or a whole site?
  • Does it look like a phishing report that should go straight to the security team?
  • Is it a duplicate of a ticket we’re already working on?

Today that’s usually done by a person reading the queue, or by keyword rules that break the moment someone writes “the internet is broken” instead of “network outage”. A general-purpose LLM can do it, but it’s slower, more expensive per ticket, and it answers in free text you then have to parse.

A decision model like Jev is aimed squarely at this gap. Here’s the kind of setup I’d like to try. It’s my own pseudo-code to show the idea, not Jev’s actual API:

# Pseudo-code: auto-triage for a help desk queue (e.g. GLPI)
decision: ticket_category
input: ticket.subject + ticket.description
options: [hardware, software, network, account_access, security, other]

decision: priority
input: ticket.description + affected_users
options: [P1_critical, P2_high, P3_normal, P4_low]

rules:
  - if confidence >= 0.85: apply category and priority automatically
  - if confidence < 0.85: leave for a technician to review
  - always: log the model's answer and confidence for auditing

The important part is the confidence threshold. Anything the model is unsure about still goes to a human. That’s how I’d want AI in a support team: it takes the obvious, repetitive sorting off the queue, and people handle the judgement calls.

In my homelab I already run GLPI as a ticketing system, so this is exactly the kind of experiment I’d like to try once I can get access.

The catch

A few things I’d want to see before trusting it in production:

  • Independent benchmarks. The speed and cost figures are the company’s own. I’d like to see how accurate it is on real, messy tickets.
  • Setup effort. You have to design the options and inputs yourself. That’s more upfront work than typing a prompt into a chatbot, although it’s also what makes the output predictable.
  • It’s not a chatbot replacement. Jev won’t write a reply to the user or explain a fix. It decides; something else (or someone else) acts.
  • Data handling. Tickets often contain personal details and sometimes passwords people shouldn’t have pasted. Any AI service in the loop needs the same privacy and security review as any other vendor.

My takeaway

Jev is an interesting sign that AI is splitting into different tools for different jobs: big conversational models for writing and explaining, and fast, cheap, decision-only models for the plumbing inside business systems. For IT support, where so much time goes into sorting and routing, that second category could be the more useful one.

I’ll write a follow-up if I get early access and can test it against a GLPI queue in my homelab.

Sources

← All postsSee my projects