“Decision models” are a hot new topic in the world of AI, because they are smart and fast at the same time. But there’s a catch: You can only use them for a very specific kind of task.
One example: evaluating comments on a website. A decision model might be asked three questions: Is it spam? Which topic does it belong to? How angry is the writer? It answers “not spam, 0.97”, “tech, 0.88” and “frustrated, 0.71”. It does that in fractions of a second, much faster than a Large Language Model like Claude. But as you can see: It only delivers the results of decisions. Hence the name. This is perfect for some jobs and useless for others.
In this article I’ll show you what these models are, what they are good for and what their limitations are.
Decision models compared to ChatGPT, Claude etc.
You already know chatbots like ChatGPT, Claude and Gemini. They are Large Language Models (LLMs). They process natural language from an input and can deliver an output in natural language as well. This makes them enormously flexible in what they can do.
A decision model is an AI that does not write. Instead, you give it a piece of information and a fixed list of possible answers. It rates every answer on the list and tells you which one fits best and how sure it is. The “how sure” part is a number between 0 and 1. If the answer is “0.97”, it means the model is almost certain.
The term took off in September 2026, when the U.S. startup TypeSafe AI released a model called Jev. Within four weeks, OpenAI, Amazon, Cloudflare and others followed with their own versions. See the tag “Decision Models” for the latest news in this category.
Decision models answer three kinds of questions
Most decision models work with three types of questions:
- Yes or no: “Is this true?” The model returns a probability. Example: Is this comment spam?
- Choice: “Which of these options fits best?” You define the options, the model picks one. Example: Does the comment belong to sports, politics, tech or something else?
- Score: “Where on this scale?” You describe a scale, and the model places the input on it. Example: Is the writer calm, frustrated or very angry?
You can ask several questions in one request. That is what happened in the comment example above.
Are they actually new?
Decision models are not a completely new idea. They are basically classifiers, the same kind of software that has sorted out spam in your email inbox for many years.
What is new is the breadth of their abilities. Older classifiers were usually built for a single task. Jev and its competitors handle thousands of different tasks without special training.
Not to be confused with …
“Decision model” is a loose label from the developer community, not an official term. TypeSafe calls its models “System One Models”, after the psychologist Daniel Kahneman’s idea of fast, intuitive thinking. Chatbots like Claude would be the slow, deliberate “System 2”.
Also, don’t mix it up with the “decision models” of business software, known as DMN. Those are hand-drawn diagrams of business rules. They have nothing to do with AI.
The key difference: writing versus choosing
Here is how these differences play out in practice:
- Output: A chatbot answers in free text. That makes it flexible, because it can summarize or explain something or draft an email for you. It also means it can invent facts (the infamous “hallucinations”) or ignore the format you asked for. A decision model can only return one of the answers you defined, so it can never go off script.
- Speed and cost: A decision model reads the input once and does not have to write a reply. It’s a streamlined process. Because of this, vendors report response times between tens and hundreds of milliseconds. And they are cheap, too: An independent test put the cost of Jev at about 4 cents for 1,000 decisions.
- Confidence: A decision model is built to say how sure it is. A chatbot tends to sound just as convincing when it is wrong.
- Explanation: A chatbot can explain its answer. A decision model gives you numbers only. Both can be wrong, but in different ways, as we will cover below.
How LLMs and decision models can work together
Because of these differences, LLMs and decision models are partners, not rivals. It makes sense to combine both. One example: The decision model does the routine sorting. If it is confident, the result is used right away. If it is unsure, the item goes to a chatbot or a human for a closer look.
In the comment example above, obvious spam would be deleted automatically and clear cases filed under their topic. Only the unclear or very angry comments would land on someone’s desk, and an LLM chatbot could help draft the reply.
What are decision models used for?
Decision models work well when the same kind of judgment has to be made over and over, especially on messy input that follows no clear pattern.
Use cases for content teams
For content teams, suitable jobs could look like this:
- Tag an archive: Sort thousands of articles by topic, or by whether they are still relevant.
- Route incoming messages: Send pitches, press releases and reader mail to the right person.
- Spot what needs a human: Flag problem comments, newsletter replies or leads that deserve a personal answer.
- Pre-filter for chatbots: Let only the unclear cases reach a more expensive LLM, for example for rewriting.
- Check images before publishing: Some models also accept images. They could scan screenshots for visible passwords or internal web addresses.
One limit: For a one-off job, such as sorting a single spreadsheet of 200 comments, uploading the file to ChatGPT or Claude works fine. Decision models make sense when the same judgment keeps coming back and should run automatically.
How you can try one
Most people will not use a decision model directly. The easier way is an automation tool. The platform n8n, for example, has an official integration with Jev. A workflow can ask the model to pick a category and then branch depending on how confident it is, sending unsure cases down a different path.
TypeSafe also offers a playground in its developer console, where you can try out questions in the browser.
Where are their limits?
One important question is left: How good are these decision models? Early evidence is encouraging that they are useful and reliable for sorting tasks. An independent academic evaluation tested Jev on 37 datasets and found accurate answers and confidence numbers that matched reality fairly well. But the field is young, and many figures still come from the vendors themselves.
Beyond that, there are some limits to keep in mind when thinking about potential use cases for them.
It cannot write or reason
Because a decision model doesn’t give you a written reply, there’s no explanation for its answers. It is also weak at tasks that need reasoning, such as arithmetic, comparing dates or multi-step logic.
“Can’t hallucinate” is only half true
A decision model can never return an option you did not define. That removes one kind of error. But it can still pick the wrong option, and do it with high confidence. TypeSafe concedes that its claim of 0% hallucinations follows from the output format, not from measurement. The confidence numbers also vary: One audit found them reliable on one test set but overconfident on another.
Hidden judgments
A neat number can hide a value judgment. Developer Simon Willison asked Jev to rate Bay Area cities as “good cities”. The tidy scores mirrored socio-economic patterns in the training data. Since you get no explanation, you cannot ask why.
Language
TypeSafe says accuracy is best in English. Small tests in eight languages, including German, looked good. If you publish in German or another language, test a model on your own content before you rely on it.
Who offers them?
Jev comes from TypeSafe AI, a startup led by Diogo Almeida, who previously worked at OpenAI. Within four weeks, OpenAI launched a Decisions API, which is still in public beta. Cloudflare released its Clef models, Amazon released Strands Decider, and a team from Stanford and NVIDIA released CLM-8B. Several of these come as free downloads that developers can run on their own hardware.
If you want to keep up to date on important topics like this one, consider subscribing to my Smart Content Report newsletter.
Hype or here to stay?
I’ve been writing about tech for several decades by now. I’m used to hype that goes nowhere (Virtual Reality), hype that makes a comeback years later (mobile web) and hype that turns out to be truly revolutionary (ChatGPT).
Where do decision models belong?
If I were a decision model, I might reply: “hype, 0.61”.
Because I’m not a decision model, I can also tell you how I arrived at this number.
Signs of hype
- Money: TypeSafe’s valuation rose from about $200 million to $7.5 billion in under four weeks. The startup raised $870 million in a funding round led by Andreessen Horowitz. These numbers are staggering.
- Vague adoption claims: Andreessen Horowitz says 25% of the Fortune 500 use Jev. No customers are named. What does “use Jev” mean in reality? Unclear.
- Clones: Many imitations appeared within days. These vendors choose their own headline numbers that are also impossible to verify. The hype train has left the station and everyone’s running to get on board. We have seen this before.
Why it may last anyway
But it’s not all hype.
Independent tests show gains in speed and cost on sorting tasks. OpenAI, Amazon and Cloudflare shipped their own models within weeks, which big companies rarely do for a passing fad. And automation tools like n8n now treat a “decision” as a standard step. It feels like this category is here to stay in some form.
A comparison helps: The personal AI assistant OpenClaw went viral in developer circles in early 2026. The excitement faded, only to return as consumer products a few months later like Meta’s Muse.
Decision models will probably follow a similar curve. The first wave of excitement cools and clones disappear. What stays is the idea.
In the end, there’s real substance. There are realistic use cases. But some of the headlines and claims around these new tools have been exaggerated.
For content professionals I see it like this: You might never choose a decision model consciously. Instead, you will use a CMS, a newsletter tool or a moderation service that uses one as part of its feature set. And that will be helpful for you.
That is not exciting. But it doesn’t have to be.
Sources
- Introducing System One Models & Jev – TypeSafe AI Blog
- Evaluating and Benchmarking the System One Model Jev – arXiv
- AIM-Decision: Jev vs Kev vs LLMs – AIMultiple
- ASSAY-001 — TypeSafe Jev: calibration and type safety, independently verified – JourdanLabs (on GitHub)
- Jev AI Review (2026): Fast, Cheap, Literal – Preuve AI
- Jev: Hype vs. Reality – Antirez Critique & 2-Hour Clone Reveal Calibration Moat – BestHub
- It’s Easy to Dismiss Jev as Just a Classifier – Sebastian Raschka
- TypeSafe AI’s Jev Brings Confidence-Based Workflow Routing to n8n – Scalevise
- Amazon releases its own Jev clone as decision models flood the web – TechCrunch
- Introducing Clef: our open-source decision models, and new RL fine-tuning platform – Cloudflare Blog
- OpenAI Decisions API: How to Use, Pricing, Public Beta (October 6, 2026), 3 Question Types, and Differences from Jev – SmartScope
- Stanford and NVIDIA release CLM-8B, 9 times faster than Jev – Pasquale Pillitteri
- Jev creator TypeSafe closes $870M round at $7.5B valuation – SiliconANGLE
- OpenClaw 2.0 Is Here, and the AI Agent That Started a Craze Is Growing Up – Kingy AI
Stay up to date
AI for content creation: the latest tools, tips and trends. Every two weeks in your inbox:
