Say a retail chain wants to train new employees in handling customer complaints. The course includes exercises where generative AI plays the customer. The course participant might meet a customer who is irritated about a damaged item and write how they would handle the situation.
When the answer comes in, the course needs to do two things. It should give the course participant feedback on how it went (A). It should also assess the answer and decide what happens next (B), for example a new attempt, a hint, a harder customer, or a notice to a supervisor.
A is a conversation, and that is where generative AI is strong. B is a decision. In many solutions today, the same AI handles both parts. If you let the language model also make the decision, the same answer can be assessed differently from one run to the next, and you cannot tell when the assessment is uncertain. If the decisions are instead fixed in advance, the assessment stays consistent and can be followed up.
Two different kinds of work
A language model, often called an LLM, is the type of AI behind ChatGPT and Claude. It writes text word by word, and that is why it is good at explaining. It can play a customer, describe what worked in an answer, and formulate feedback that the course participant understands and can learn from.
Choosing the next step sets different requirements. The course developer has already decided which paths exist in the course, and the decision must always land on one of them. Two course participants who answer the same way should end up in the same place. The course also needs to know how certain the assessment is, so that uncertain cases can be sent to a human.
A language model can be instructed to reply in a fixed format, but it still writes its way to the answer. If you ask it to state how certain it is, that number is also text, and it tends to be too high. The language model is simply built to write, not to weigh options against each other.
What a bounded decision is
By a bounded decision we mean that the course developer formulates the question and the possible answers in advance, and that the AI may only choose among them. The question can be asked in a few different ways:
- Choose an option: which of the five most common mistakes does the answer resemble most?
- Place the answer on a scale: how well did the course participant meet the customer, from 1 to 4?
- Answer yes or no: did the course participant follow the company’s return policy?
Every answer comes with a probability. On the return-policy question the answer might be “yes, 0.85”. What happens next is governed by rules the course developer has set. If the value is above 0.8 the course participant moves on, below 0.4 they get a hint, and everything in between goes to a supervisor. The model therefore does not set the thresholds itself.
A model that only makes decisions
On 15 September 2026 the company TypeSafe launched a model called Jev. Jev writes no text. It receives a state, for example the course participant’s answer, and one or more questions of the type described above. The answers come back at once, each with a probability.
According to TypeSafe, Jev responds in between 0.07 and 0.5 seconds and costs a fraction of what a language model costs per question. The company also states that the model is trained so that the probabilities can be trusted. If it says 0.8, it should be right roughly eight times out of ten. Because Jev does not write freely, it cannot answer with anything outside the options the course developer has decided, but it can of course choose the wrong one among them.
Jev is not the only way to solve this. The same split can be built with a language model that is required to reply in a fixed format, or with a smaller model trained on answers that have already been assessed. Within a week of the launch, several open alternatives appeared that try to do the same thing, but according to the first comparisons it is hard to get their probabilities to line up. What is new with Jev is that the decision is so fast and costs so little that it can be made after every answer.
Four examples
The examples below are sketches, not finished solutions. In all four, the language model handles the dialogue with the course participant, while the decision model decides what happens next.
1. Customer complaints in a store
We return to the exercise from the beginning. A customer comes up to the till with a jacket they bought two weeks ago. The zip has broken, the receipt is gone, and the customer is irritated: “I’ve barely used it. This is not okay.”
The language model plays the customer and becomes calmer or more irritated depending on how the course participant responds. At the same time the decision model scores the answer on a scale from 1 to 4 that the course developer has written in advance. A one might be that the course participant immediately says “without a receipt we can’t do anything”. A four might be that they first show understanding of the customer’s irritation, then offer to look up the purchase via the card the customer paid with, and finally explain what happens with the jacket.
If the answer lands on 1 or 2, the exercise is paused. The language model explains what was missing, for example that the course participant never acknowledged that the customer was irritated, and the course participant gets to try again. If it is 3 or 4, the exercise moves on to a harder customer, who wants a refund for a jumper they have used for three months.
2. Which mistake did the course participant make?
The language model describes a case and asks how the course participant would handle it. The decision model decides which of the five most common mistakes the answer resembles most, or whether it resembles none of them. The language model then explains that particular mistake with an example from the company’s own routines, and the next exercise trains that step.
3. When a supervisor needs to look
Here the course asks the decision model whether a supervisor needs to see the answer. If the probability of yes is higher than the threshold the course developer has set, the answer is sent to a supervisor, and the course does not move on automatically. The language model writes a short summary so that the supervisor understands why the answer was flagged without having to read the whole conversation.
4. When the answer must follow a routine
In a course on fire safety, the course participant can ask questions about the company’s routine. Someone might ask whether it is okay to run back in and fetch their jacket once the alarm has gone off. The language model answers in its own words based on the routine, and that is usually easier to understand than the routine text itself.
The risk is that the language model adds something that is not there, for example “if it only takes a minute”. Therefore the decision model gets a question before the answer is shown: does the answer say the same thing as the routine, without adding or removing anything? If the answer is yes with high probability, the language model’s explanation is shown. Otherwise the course participant reads the routine exactly as it is written. When safety is at stake, it is better that the course participant reads the routine word for word than gets an answer that is easier to understand but wrong.
Why speed matters
An assessment that takes several seconds and costs money every time is something you rarely do, usually in a test at the end of the course. If it takes a fraction of a second and costs almost nothing, it can be done after every answer, and then the course can adapt while the course participant is working.
Speed also affects how the exercise feels. The assessment happens between the course participant’s answer and the customer’s next line. If it takes a few seconds, the wait is noticeable every time, the conversation feels sluggish, and it becomes harder to inhabit the situation.
The name Jev comes from the economist William Stanley Jevons, who described how the use of something often increases when it becomes cheaper. TypeSafe argues that the same applies to decisions: when they become cheap enough, they can be used in places where it previously did not pay off.
What the course developer still needs to do
The technology does not remove the pedagogical work. The questions and scales must be clear enough that a colleague could assess the answers with the same rubric. The thresholds need to be tested before the course goes live, by having the model assess answers that a human has already assessed. Where the assessment has consequences, for example for a certificate or an authorisation, a human should review a sample. Course participants should also be told that their answers are assessed by an AI and how that works.
We think the split between dialogue and decision is one of the more interesting areas in digital learning right now, and we look forward to testing it in different training programmes.
Further reading
- Introducing System One Models & Jev. TypeSafe’s own presentation of Jev, with their measurements.
- JevBench. An independent comparison of Jev, common language models and open alternatives, with accuracy, price, response time and how well the probabilities hold up.
- Jev: System One Models, Use Cases, Open Source and Benchmarks. A walkthrough of the open alternatives and what sets them apart, including licences.