An email arrives. The business’s AI workflow sends it to a large language model with a simple instruction: invoice, complaint or sales pitch? The answer takes several seconds, every call is billed, and it sometimes arrives in a form the next piece of software cannot read: a sentence instead of a word, a category that does not exist.
None of this is a malfunction. It is the normal behaviour of a tool built to write, being asked to sort. Before automating further, the question deserves to be asked: is this the right tool for that gesture?
System 1, System 2: the distinction automation forgot
In Thinking, Fast and Slow, the psychologist Daniel Kahneman describes two modes of thought. The first is fast, automatic, almost effortless: recognising a face, sensing that a sentence rings false. The second is slow, deliberate, costly: working out a multiplication, weighing two options, building an argument.
Today’s large language models, especially when they reason step by step before answering, belong to the second mode. They excel when an answer has to be built: summarising a file, drafting a letter, explaining a discrepancy.
Yet in software, most decisions belong to the first. Is this ticket about billing or technical support? Is this document a purchase order or a quote? Nobody drafts the answer to these questions: it is recognised. And a process asks hundreds of them a day.
By adopting generative AI, automation imported a System 2 tool for System 1 gestures. It works, most of the time. But it is a default, not a design choice.
What the release of Jev says about the market
On 15 September 2026, the US company TypeSafe AI unveiled Jev, which it describes as the first “System One Model”. Its founder, Diogo Almeida, worked at OpenAI on the training methods that led to ChatGPT.
Jev does not produce text. It receives the state of a program and questions whose form is fixed in advance (pick from options, place on a scale, answer yes or no), and returns for each a typed answer with a probability. According to the vendor, it would answer in 70 to 500 milliseconds, 40 to 200 times faster than frontier models, for $0.042 per million input tokens, with output not billed.
The product matters less than the signal. A researcher who helped make conversation the dominant form of AI considers it the wrong form for automation. The market is starting to tell the tool that writes from the tool that decides.
What we do not know yet
A great deal. The announced performance was measured by TypeSafe on its own evaluations, and had not been independently reproduced at the time of writing. The vendor asserts that an answer outside the expected format is impossible; that is its claim, not a finding. Access is through a waiting list. The model returns probabilities without a rationale, which raises the question of explaining a decision after the fact, as DataCamp points out.
Finally, TypeSafe states that its service currently runs from the West Coast of the United States. For a European business routing its operational flows through this kind of model, that is no minor point. We devote a separate article to it.
Four gestures in an AI workflow that call for a reflex
In the processes we observe, the steps handed to AI usually come down to four gestures. None of them calls for writing.
Sorting
Sorting means placing an incoming item in a known category: it is a classification. A firm receives emails every morning that must reach the right file; a small manufacturer, delivery notes, invoices and certificates of conformity. Sorting does not require understanding the whole document, only recognising what it is.
Routing
Routing means deciding where an item goes once you know what it is: to which person, which team, which tool, and sometimes which model. A customer request goes to the sales rep for the area, to support or to accounts. In a more elaborate chain, routing also picks the next treatment: a simple question gets a standard answer, a complex one goes to a slower model or to a human.
Scoring
Scoring means placing an item on a scale. Scoring an incoming prospect, the priority of a ticket, the risk level of a refund request: in each case, what is wanted is a position on a short scale, not a commentary. A small business receiving many quote requests does not need an analysis of each, but to know which ones to handle today. The score is useful if it is fast, stable from one day to the next, and comes with its degree of confidence.
Checking
Checking means verifying what someone else produced, and more and more often what another model produced. Does the answer drafted by an assistant follow the expected format? Does it contain personal data that should not go out? This safeguard runs on every output: it must be fast, cheap, and say when it is not sure. Handing it to the same kind of model as the one being checked is like asking a writer to proofread their own text.
In all four cases, the need is the same: an answer chosen from a closed list, obtained quickly, with a usable level of confidence.
Why the LLM makes a poor switchman
Large language models do very well what they were designed for. But three of their characteristics work against them as soon as they are asked to route.
Slowness and unit cost. An LLM produces its answer word by word. To choose between three categories, it generates text, sometimes a full chain of reasoning, before reaching the useful word. On a step called thousands of times a day, that detour is paid for in seconds and in invoices.
Confidence without calibration. An LLM answers with the same assurance whether it is sure or guessing. The tone says nothing about reliability. It is the most discreet form of AI hallucination: not a spectacular invention, but a plausible, wrong category presented as certain, which nobody will check because nothing flags it.
Format. The answer is text, which the software has to parse to extract the decision. One comma too many, a slightly reworded category, and the next step fails or, worse, gets it wrong without saying so.
These flaws can be mitigated. Most providers now offer structured outputs that constrain the format, and the model can be asked to state its confidence, with no guarantee that the figure faithfully measures it. That improves the tool. It does not change the register: a decision that only needs a reflex is still being drafted.
The real issue: probability and threshold
Once this sorting is done, the essential remains to be decided, and it is not a question of tools. The value of an automated decision lies not in the decision itself, but in what is done with its level of confidence.
The principle is simple. Each decision comes with a probability. Above a confidence threshold set in advance, the software acts alone: the email is filed, the ticket routed, the request scored. Below it, the decision goes elsewhere, to a person or to a slower model that will take time to reason. The reflex handles the routine; deliberation is kept for the cases that deserve it. It is a precise way of putting the human in the loop: where they are useful, not everywhere.
This threshold is not a technical setting. It expresses an organisational question: how many errors does the business accept on this step, and who catches them? Sorting incoming mail tolerates a few errors, which an assistant will fix in passing. A refund decision, far fewer. The same model, with the same accuracy, can be perfectly suited to the first step and dangerous on the second.
It is also what makes an enterprise AI agent trustworthy. An agent chains many switches before producing anything visible. If they are made without a threshold, its errors compound in silence. An agent is never more reliable than its switches.
None of this is settled in code. It is designed beforehand, with the people who know the process, like any transformation that affects the organisation more than the tool.
Reviewing your AI workflow in three questions
Before changing tools, three questions are enough to review an existing process.
- Which steps produce a closed answer: a category, a destination, a score, a yes or a no?
- For each, what does an error cost, and who catches it today?
- What level of confidence justifies acting without human intervention?
These three questions hold whatever tool is chosen: a decision model like Jev, a small model run locally, or a plain business rule. The answer is often the rule. The practical guide to Jev says so itself: what a program can compute exactly, deterministically, should not be asked of a model.
Let us go back to the email from the start. In a redesigned workflow, it is first sorted by a tool built to sort, in a fraction of a second. If the category is certain, it joins its file without anyone seeing it. If not, it goes to someone who will decide, and that hesitation becomes information: it shows where the process is ambiguous. The large language model only steps in where writing is needed: replying to the customer, summarising the complaint.
Automating means first deciding who decides. One question remains, rarely asked early enough: who sees all these decisions go by?
Common questions
What is an AI workflow?
An AI workflow is an automated process in which some steps are handed to an artificial intelligence model: sorting messages, routing requests, scoring files, drafting replies. Most of these steps are closed decisions; only a few genuinely require producing text.
Do you need an LLM to automate a classification?
Not necessarily. A classification expects an answer chosen from a closed list, quickly, with a level of confidence. A large language model can do it, but slowly and without reliable calibration. A decision model, a small local model or a business rule is often a better fit, depending on volume and the cost of an error.
What is a System One Model?
It is the name TypeSafe AI gave, in September 2026, to a category of models that do not generate text: they receive the state of a program and typed questions, and return structured decisions with probabilities. The name refers to Daniel Kahneman’s System 1, fast and intuitive thinking.
Sources: TypeSafe AI, introducing System One Models and Jev (15 September 2026); The Register (16 September 2026); DataCamp; DEV Community, practical guide; The Rundown AI; AI News. The performance figures cited are those declared by the vendor, not independently reproduced as of September 2026.