# Frequently asked questions

Plain-English answers about classifier models, System 1 AI, and this directory.

## What is a classifier model?

A classifier is an AI model that sorts things into categories. You give it an input, like an email, a review, or a support ticket, and it returns one answer from a set you defined in advance.

A spam filter is the classic example. It reads an email and answers one question: spam or not spam. It never writes you a paragraph. It just picks.

That narrow job is the point. Because the answer can only be one of a few options, it is fast, cheap, and easy to check.

## How is a classifier different from a chatbot like ChatGPT or Claude?

A chatbot (a large language model, or LLM) generates text. Ask it anything and it writes an answer word by word. That makes it flexible, but also slower, more expensive, and harder to verify, because the answer can be anything.

A classifier answers a fixed question with a fixed set of possible answers. Think of the difference between an essay question and a multiple-choice question. A chatbot writes the essay; a classifier fills in the bubble.

Many real systems use both: the LLM writes, the classifier decides, and ordinary code acts on the decision.

## What is System 1 thinking?

System 1 and System 2 are terms the psychologist Daniel Kahneman popularized in his 2011 book *Thinking, Fast and Slow*.

- **System 1** is fast, automatic, and intuitive. You use it to recognize a friend's face, notice that someone sounds angry, or know that 2 + 2 = 4 without working it out.
- **System 2** is slow, deliberate, and effortful. You use it to fill out a tax form, compare two mortgages, or solve 17 × 24 in your head.

Most of your day runs on System 1. You only call in System 2 when a problem needs careful step-by-step reasoning.

## What is a "System One" AI model?

It borrows Kahneman's idea. A System One model makes quick, single-step judgments: is this message urgent, which department should get this ticket, how risky is this transaction on a 1 to 5 scale.

Chatbots that write long answers or reason step by step are closer to System 2. They are better at hard, multi-step problems, but they cost more and take longer.

The practical idea is to match the tool to the job. Most business decisions that repeat all day are System 1 decisions, so they can go to a fast decision model instead of a slow, expensive writer.

## What is Jev?

Jev is a System One decision model from TypeSafe AI. It reads text and picks from options you provide. It does not write free text.

It answers three kinds of questions: a yes/no probability, a choice from a list of options, or a score on a scale you describe. TypeSafe serves it through an API. Their launch announcement claims it is 20 to 200 times faster and 40 to 400 times cheaper than earlier models for these decisions; those are TypeSafe's figures, not measurements from this site.

Learn more at [madewithjev.com](https://madewithjev.com). This directory lists tasks that fit Jev or any similar classifier.

## What is a "typed decision"?

A typed decision is a question whose answer has a fixed shape, so a program can use it directly without reading prose. Every use case on this site is one of three types:

- **Yes/no:** "Does this email contain a refund request?"
- **Choice:** "Which team should handle this ticket: billing, technical, sales, or other?"
- **Score:** "How urgent is this message, from 1 (can wait) to 5 (emergency)?"

Because the answer is always one of the allowed values, your code can route, block, flag, or count it with no parsing and no surprises.

## Why use a classifier instead of just asking ChatGPT or Claude?

You can ask a chatbot to answer "yes" or "no", and for a handful of items that works fine. The case for a dedicated classifier grows with volume:

- **Speed:** decisions come back in a fraction of a second, which matters inside a live product or an agent loop.
- **Cost:** you pay for a short decision, not a page of generated text. At thousands of items a day the difference adds up.
- **Reliability:** the answer is always one of your options. No rambling, no reformatting, no "Sure! Here's my analysis...".
- **Confidence:** many classifiers return a probability, so you can automate the confident calls and send the uncertain ones to a person.

## When is a classifier the wrong tool?

Skip the classifier when:

- **You need written output.** Drafting emails, summaries, or code is a job for an LLM.
- **Plain rules already work.** If a regular expression, a lookup table, or an "if amount > 500" check gives the right answer, use that. It is free and exact.
- **The problem needs many reasoning steps.** Multi-step math or planning is System 2 work.
- **The input is raw images or audio** and the model only reads text.

These are marked grade X in the directory.

## What do the grades S, A, B, C, and X mean?

Each use case is graded on how well it fits a classifier and how much it is worth automating:

- **S:** top tier. A clean decision that repeats often and has a major payoff.
- **A:** a clean decision that is concrete, repeats routinely, and is worth automating.
- **B:** a clean decision with either high volume or high value, but not both.
- **C:** a clean decision with low volume or low value, or too vague to act on yet.
- **X:** not a classifier job (see the question above), or an idea that should not be built.
- **?:** not enough information to grade.

See the [grade pages](/grades/) to browse by grade.

## What is the difference between "seen" and "idea" use cases?

**Seen** means someone has publicly built, shown, or described this use case, and the page links to the source where available. **Idea** means it was proposed but we have not found it in the wild yet. Ideas are a good place to look for an unclaimed project.

## What does the "human-review" flag mean?

It marks decisions where a wrong answer could cause real harm: medical triage, fraud, legal privilege, safety reports, and similar. For these, use the classifier to sort and prioritize, and keep a person in the loop for the final call. A good pattern is to automate only high-confidence answers and send everything else to a human queue.

## What is a confidence score?

Many classifiers return a probability along with the answer, for example "urgent: 0.94". Higher means the model is more certain.

You choose a threshold. Above it, act automatically; below it, ask a person. Note that confidence is not the same as correctness. A model can be confidently wrong, so check its answers on real examples before you trust a threshold.

## How do I start building one of these?

1. **Pick one use case** that happens often in your work. The S and A grades are the best starting points.
2. **Write the question and the allowed answers** in plain language, as if briefing a new employee. Be specific about edge cases.
3. **Collect 30 to 100 real examples** and label the right answer yourself.
4. **Run the classifier on those examples** and count how often it agrees with you.
5. **Set a confidence threshold**, automate above it, and send the rest to a person.
6. **Keep spot-checking** after launch, because real inputs drift over time.

Each use case page has a "Build this with a classifier" box with a starter question.

## Do I need to know how to code?

Some. Calling a classifier API takes a few lines of code, and acting on the answer (routing a ticket, tagging a lead) usually happens in code (often with an LLM coding assistant or agent) or in an automation tool (such as Zapier, Make, or n8n). If you can describe the decision clearly, a coding assistant can usually write the glue for you. The hard part is choosing a good question, which is what this directory helps with.

## Can I submit my own use case?

Yes. Use the [submit form](/submit/). Every submission is reviewed before it appears, and approved ones are added in a future version of the directory. Good submissions describe a decision that repeats, name the possible answers, and explain why it matters.

## Can AI agents use this site?

Yes. Agents can search the directory through a JSON API, an MCP server, or an A2A endpoint, read any page as Markdown, and submit new use cases by solving a short proof-of-work puzzle instead of a CAPTCHA. See the [agent guide](/agents/).

## Who made this directory?

Jev AI Use Cases is an independent directory created by [Clark Mackey](https://cakewebsites.com). It is not affiliated with or endorsed by TypeSafe AI, the maker of Jev. It grew out of the Jev Use Case Atlas, a research sweep of public builds, posts, and videos about decision models. You are welcome to link to or cite any page; each use case page has a ready-made citation.