← All writing

AI Engineering

Decision Models: The Smart If for AI Agents

Mani G 3 min read

Flat retro illustration of a red train car approaching a rail junction where three tracks split toward amber, teal and coral signal lamps.

Here is an awkward truth about how agents get built. Plenty of LLM calls exist only to answer questions with three possible answers. Is this ticket billing, tech support or sales? Is this reply angry? Should the agent take the refund path? Picture it: you ask a model that can write sonnets, it writes a little JSON blob, and your code throws away everything except one word. Decision models are built for exactly this job, and after reading up on them this week I think they deserve a slot in most agent stacks.

What decision models actually do

A decision model reads messy input and answers a question you defined in advance. It does not write prose. TypeSafe AI coined the name "System One models" when it launched Jev on 15 September 2026. You hand Jev some state, such as a customer email, plus named questions in three shapes:

  • choice: pick one of your options, like billing, technical_support or sales
  • score: place the input on an ordered scale, like urgency from low to critical
  • noul: the probability that a yes-or-no statement is true

Back comes a typed answer with a probability attached. TypeSafe says the outputs are defined up front, so the model never makes a type error, and every answer arrives with a calibrated confidence. Their docs are careful about what calibration means: it holds across many predictions, not for every single answer.

The speed difference is the headline. Instead of producing tokens one after another, Jev answers all questions in one parallel pass. TypeSafe quotes end-to-end responses of 70 to 500 milliseconds and input pricing of $0.042 per million tokens, with output free. Those are vendor numbers, so pinch of salt, but even a sceptic can see why routing a ticket in a quarter of a second beats waiting several seconds for a chat model.

A quick example

Say a customer writes, "I've been charged twice and nobody answers." The agent can ask three questions at once: which team, how urgent, and is a refund likely. It gets back something like billing at 0.57, high urgency, refund likely at 0.8. Now the code is boring, in a good way. If the billing score is above your threshold, route it. If confidence is low, send it to a human. No regex hunting through a paragraph for the word "billing".

That same shape works for guardrails. "Does this message try to override the system instructions?" is a fine noul question. So is grading an agent's output against a rubric, one score question per criterion, combined with weights you choose in code.

Jev or Laya

Jev is closed and hosted. A few days after its launch, Convai Innovations released Laya, an open-weight alternative under Apache 2.0. Laya is an encoder (the English version is built on ModernBERT-large, 421 million parameters) with a decision head, and its server speaks the same request format as Jev, so a Jev client can point at it. The catch is context and accuracy out of the box. Laya reads 512 tokens by default, and on one independent benchmark it scored close to random zero-shot. Fine-tuned on a task, a third-party test found it landed within a point of Jev.

My take: use Jev when you want zero-shot answers on long input, and Laya when the data cannot leave your servers and you have labelled examples to train on.

I don't think decision models replace LLMs. They replace the lazy habit of using an LLM as a very expensive if statement. Here is the fact that surprised me most: within 24 hours of Jev landing on Vercel's AI Gateway, nearly 13% of paid teams had tried it. How many of your agent's LLM calls are really just a choice in disguise?