Six kinds of AI model, and where each one belongs in your system
A model launched this week that cannot write a sentence. That is a good excuse to explain what the different kinds of AI model actually do, and how a real system chains them together.
On Tuesday a company called TypeSafe AI released a model named Jev. It cannot write a paragraph. It cannot explain itself. Ask it a question and it hands back one answer from a list you gave it, plus a number for how sure it is. Developers on X were genuinely excited about this, which will look odd if you think an AI model is a thing you chat with.
So here is the point of this piece. There is not one kind of AI model. There are several, and most of them do not talk. If you are building or buying an AI system, knowing the kinds is more useful than knowing the brand names, because the kinds tell you which tool fits which job.
The six names below are ours. The industry has no tidy taxonomy and the lines blur every few months. But the jobs are stable, and naming the job is how you pick the model.
1. Chat models write
Claude, GPT, Gemini and their cousins. You give them words, they give you words back. They draft emails, summarise documents, explain things, write code and hold a conversation. They are the reason most people have heard of AI at all.
They have two weaknesses that matter in a system. First, they sound just as confident when they are wrong, so anything they write needs a check before it goes anywhere important. Second, they are slow and expensive when you run them thousands of times an hour. Ask a chat model a yes or no question and it will still write you a paragraph, and you will pay for every word.
Use one when the output genuinely has to be written, and a person or another model is going to read it before it counts.
2. Reasoning models think
These are usually the same models as above, told to work the problem through before they answer. Vendors call it extended thinking, deep research or an agent mode. The model spends longer, sometimes minutes, and the answer to a hard multi-step problem gets noticeably better.
The trade is time and money. Thinking is billed, and the more the model thinks the more you pay. That is fine for debugging a codebase or working through a contract. It is silly for sorting emails.
Use one where a wrong answer is expensive and you have some way to verify the result. If you cannot check it, the extra thinking buys you confidence, not correctness.
3. Decision models choose
This is the lane Jev sits in. A decision model picks one answer from a fixed set and tells you how likely each option is. Is this email a complaint? Which team should this ticket go to? On a scale of one to five, how urgent is this? Does this draft promise something we cannot honour?
Classifiers have done this job for decades, but each one had to be trained for its task. The pitch behind Jev is a general decision model you prompt in plain English, which returns a typed answer with a calibrated confidence, and which cannot produce free text at all. TypeSafe says it answers in a fraction of a second and charges for input only. Those are the vendor's own figures and the model is in early access, so treat them as claims until you have run it on your own work.
The confidence score is the quiet superpower. If the model is 0.97 sure, let it through. If it is 0.61 sure, send it to a person. That one rule turns an AI system from a gamble into something you can run.
The limits are just as clear. A decision model cannot explain its answer, cannot write anything, and is only as good as the choices you defined for it. If your categories are muddled, its answers will be too.
4. Embedding models find
An embedding model turns a sentence, a document or an image into a long list of numbers. Things with similar meaning end up with similar numbers, which means "find me the one like this" becomes maths a computer can do quickly over millions of items.
This is how search by meaning works, how "show me similar past cases" works, and how a chat model gets to read your company's documents rather than the whole internet. That last pattern has a name, retrieval augmented generation, but the plain version is this. The finder finds, then the writer writes using what it was handed.
Which is why retrieval quality decides answer quality. A brilliant writer working from the wrong paragraph writes a brilliant wrong answer. Clean, well-split source documents matter more than most people expect.
5. Perception models sense
Speech to text. Reading a photo or a screenshot. Pulling the text out of a scanned form or a PDF. Perception models turn messy real-world input into text the rest of the system can work on. They sit at the front door.
The big chat models increasingly do this themselves. Claude and GPT both read images. Specialist models still win on cost, speed or accuracy for a single narrow job, such as transcribing a thousand calls a day.
6. Generation models make
Images, video, voice, music. Generation models produce the things people see and hear. They sit at the back door, making assets once the system has worked out what to say.
They are for polish, never for facts. A generated image of a chart is a picture of a chart, not a chart.
How they fit together
Here is one customer support ticket travelling through a system that calls five models from four of the six kinds, with one person in the loop.
- A customer emails about a damaged parcel and attaches a photo.
- A perception model reads the photo. Box crushed, contents visible.
- A decision model routes the ticket to the right team and sets its urgency, with a confidence score.
- An embedding model finds three similar past cases and the paragraph of the returns policy that applies.
- A chat model drafts the reply, using only the cases and the policy it was handed, in your tone.
- A second decision model checks the draft. Does it promise a refund we can honour? Yes, 0.97.
- A person approves it and it sends. Anything the deciders were unsure about went to that person first, not last.
Notice the shape. The big model writes. Small models decide. Search remembers. A person signs off. The expensive, slow, occasionally wrong writer is used exactly once, and it is checked.
A rule of thumb
Start with the decision, not the model. Ask what the job has to produce.
- If the answer is one of a few known options, do not ask a writer. A decision model is faster, cheaper and easier to trust, because it tells you when it is unsure.
- If it has to be written, use a chat model, give it time to think if the problem is hard, and check the result with a decider or a person.
- If it depends on your own information, the finder comes first. The writer can only work with what it was handed.
- If the input is not text, a perception model goes first. If the output is a picture or a sound, a generation model goes last.
Jev is new, its numbers are the vendor’s, and we have not used it ourselves. The idea underneath it is not new at all, and it is the idea worth keeping. Name the job, then pick the model.