The AI that decides instead of writes
System One Models for leaders, part 1 of 4
Most of the AI in a modern product is not writing anything. It is deciding. Is this email a refund request? How angry is this customer? Which team owns this ticket? Does this invoice match the purchase order? Each of these is a judgment a capable person makes in about a second, and most companies are now paying a large language model to make them, thousands or millions of times a day.
A new class of model, sold under the name System One models, is built for exactly those judgments and nothing else. It cannot write a sentence. That turns out to be the point. This post explains what these models are, why they exist, and the single question that tells you whether your organisation is overpaying for judgment.
The short answer
A System One model takes a piece of information (an email, a ticket, a record) and a set of fixed questions you define, and returns an answer to each question with a probability attached. It does not generate text, so there is nothing to parse and nothing to drift. The examples you will hear about are TypeSafe's Jev, a hosted service, and Convai's Laya, an open model you can run yourself.
The useful mental model: an LLM is a writer. A System One model is a switch with a confidence reading.
Where the name comes from, and a trap in it
The name borrows from Daniel Kahneman's work on two modes of thinking (Nobel lecture, 2002). System 1 is fast and intuitive: you read the anger in a face before you decide to look for it. System 2 is slow and deliberate: the mode you are in when you work out 17 × 24.
Be careful when you read about this in the press or in research. Academics already use "System 1" to mean ordinary chat models and "System 2" to mean reasoning models (Li, Zhang et al., 2025). Vendors use "System One model" for something much narrower: a model that never writes at all. When someone in a meeting says "System 1 AI", it is worth asking which one they mean.
Why it matters to the business
Today, when a team needs one of these judgments, it usually reaches for an LLM. It writes a prompt, asks for the answer in a structured format, and hopes the format holds in production the way it held in the demo. It works, but look at what it is: a writing machine used as a switch. That choice has three costs that show up on a leader's desk.
Cost. You pay for generated text even when you only wanted a yes or no. In one published comparison of AI "judges" grading agent runs, the System One model cost about $0.00035 per call against $0.028 for Claude Sonnet, and the whole experiment came to $0.34 against $28.17 (LangChain and the authors, Jev-as-a-Judge). Those are the authors' numbers on their own test, not a guarantee for yours, but the order of magnitude is the story.
Speed. The vendor's documentation says most calls take about 100 milliseconds (TypeSafe). The open model Laya reports about 33 milliseconds on a modest GPU (model card). A typical LLM call takes seconds. That difference decides whether a judgment can sit inside a live customer interaction or has to run in a batch overnight.
Consistency and auditability. This is the one I would weigh most. Ask an LLM the same question twice and you can get two answers, so you cannot tell whether your process improved or the AI simply changed its mind. In the same judge comparison, the System One model gave almost exactly the same score on every repeat run. A decision that is repeatable, comes with a probability, and can only be one of the options you defined is a decision you can log, test, threshold and explain to an auditor.
How it works, without the engineering
You give the model two things. The first is the situation: the email, the ticket, the record. The second is a short list of questions, each one of three kinds: pick one option from a list, place something on a scale, or answer yes or no.
Take a customer message: "I was charged twice for order #4411, please fix this." The questions might be: what does the customer want (refund, rebooking, information, other)? How upset are they (calm to very angry)? Is this a refund request (yes or no)? In an illustrative run, the model answers all three at once, each with a confidence: refund, at 93%; somewhere between concerned and very angry; yes, at 95%. Your systems then act on those answers by rule.
There is no magic inside. Laya, which publishes its design, is a compact text-understanding model with a small decision layer on top, a fraction of the size of a frontier LLM (model card). The research idea behind it, classifying text against labels you name on the spot, dates back at least to 2019 (Yin, Hay and Roth). What is new is packaging it so that any team can ask its own questions and get calibrated probabilities back.
The division of labour that makes it work
The most useful thing to take from this post is a way of looking at any AI-powered process. Every one of them splits three ways.
Rules and software handle what can be written down exactly: arithmetic, lookups, policy, the order of steps. System One handles the narrow, one-second judgments over messy human input, and returns an answer with a confidence. System Two, which is an LLM, a reasoning model or a person, handles writing, explaining, multi-step reasoning, and every case where System One was not confident.
The software stays in charge. It asks for judgments, acts on the confident ones and escalates the rest. None of this is a new idea in how well-run operations work: we have always written down what we can and sent the ambiguous cases to someone senior. What changes is that the first-line judgment is now fast, cheap and measurable.
The test to carry into your next review: could a knowledgeable person answer this in about a second? If yes, it is a System One job. If no, it needs to be broken down or sent to a human or an LLM.
Where it fits, and where it does not
It fits routing, triage, classification, compliance checks, quality checks on AI output, and any step where the possible answers are known in advance. It does not fit anything that requires producing new text, like a reply, a summary or a report, or reasoning across many steps. Most real processes need both, which is why the question is never "which model do we use" but "which steps go where".
Claims where you should be careful
"Cannot hallucinate" is the claim you will hear most, and it is only half true. It means the format cannot go wrong: you will always get one of your options. The model can still pick the wrong option, and pick it confidently. The headline accuracy figures for Jev come from the vendor's own benchmark, and Laya's own documentation says its base version performs close to chance until it is tuned on your data. I found no independent benchmark comparing these models with LLMs on the same tasks. Ask for results on your own examples before you believe anyone's numbers, including the ones in this post.
Questions to ask your team
- Where are we calling an LLM just to get a label, a score or a yes/no back? How many times a day, and at what cost?
- For those steps, what happens today when the AI is unsure? Do we even know when it is unsure?
- Could we show an auditor why a specific decision was made last Tuesday?
- Do we have past human decisions on these cases that we could test a new model against?
- Can this data leave our environment, or would we need an open model we host ourselves?
The next post looks at five patterns from real projects where this approach paid off, and what they have in common. Before then, it is worth asking yourself one thing: if you listed every AI call your business made yesterday, what share of them were really asking for a decision, not a piece of writing?
Member discussion