The rise of System One models

This week, the AI community got all excited about a "new type" of AI model. One that doesn't output text, like the LLMs we are familiar with. In this article, we zoom in on what these models are, why they are not as new as some claim them to be and what the impact of these system one models is on practical applications in business.

What they are

First things first. The name "system one" model originates from neuroscience. Neuroscientists use the term system one model to describe the automatic, instinctive decisions (or snap judgments) the human brain makes; while system two thinking refers to more deliberate decision making. The terms system one and system two were popularized by Daniel Kahneman in the book “Thinking, Fast and Slow”.

A number of AI models are loosely modeled after these two types of human thinking. LLMs are usually described as system two models, because of their longer, "deliberate" process. That comes with real costs: generating an answer token by token is slow and expensive at scale, and since an LLM is essentially predicting the next plausible word, its confidence is rarely a true, calibrated probability. This can be risky when you want to use LLMs for automated decision making.

This is the gap that new "system one" models such as Jev (TypeSafe AI) and Laya (Convai Innovations) claim to fill: they take a structured question as input, and give a calibrated probability as output. Both give answers in under a second, at a fraction of the cost of a frontier LLM call.

Good to know this isn't a new kind of AI. Under the hood, these are classifiers, the same family that powered spam filters and fraud detection long before LLMs existed. What's new is the training method that calibrates their confidence, and the packaging as a ready-to-call component. "System one" is smart branding, borrowing credibility from Kahneman's dual-process theory.

Why the excitement?

Many decisions AI systems make today are simple. “Is this email spam?”, “does this ticket need escalating?” Yet many companies have been using a slow, expensive LLM call to make them anyway. A model that answers in milliseconds, for a fraction of the cost, is an easy sell at high volume. Even though the theory behind Laya was already published in a paper in March 2025, with the release of Jev this month, system one models dominate the discussion in AI communities.

Where do they still have weaknesses?

System one models are deliberately specific: they only pick from a short list of pre-set answers, and can't draft, summarize, plan, or hold a conversation. Accuracy on harder cases is still modest, so they're no substitute for judgment on anything ambiguous. Each one also needs training for the specific decision it makes, which is far less flexible than simply prompting an LLM. And the category of system one models is very recent, so it's still unproven at scale.

Practical applications

Picture a company using AI to handle incoming customer emails (a common setup by now).

With only an LLM: every email goes to a large language model, which reads it, works out what kind it is, decides who should handle it, and writes a reply. That works, but it's slower and pricier than it needs to be, since the model spends its expensive "thinking" on simple sorting it wasn't built for; and an occasional misjudgment, like the wrong urgency, can slip through unnoticed.

With a system one model added: the email first passes through a small, fast model that answers a few simple questions in a fraction of a second. “Is this spam?”, “how urgent is it?”, “which team should get it?”, each with a clear confidence score attached. Only then does it move on to the LLM, and only for the part that needs it: writing a natural, helpful reply. The customer notices nothing, but the sorting is now near-instant, cheaper, and easy to check, and the LLM does only what it's genuinely good at.

The same pattern applies anywhere a business leans on AI: checking if a question is safe before a chatbot answers, or picking which documents are relevant before an AI reads them. A cheap, fast decision happens first, and the AI's language skills are saved for where they're truly needed.

At Axveco, we believe these system one models are an important next step in making generative AI practically applicable, useful and trustworthy for organizations.