Blog · 2026-09-22

One AI explains, another decides

A conversation and a bounded decision, side by side against a dark background.

Suppose a retail chain wants to train its new employees in handling customer complaints through a course. In the exercises, generative AI plays the customer. The course participant meets a customer who is irritated about a broken jacket and writes 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. It should also assess the answer and decide what happens next: a new attempt, a hint, a harder customer, or a notice to a supervisor.

The feedback is a conversation. The next step is a decision. In many solutions today, the same language model does both, and that is where the problems begin.

When the language model also gets to decide

A language model, often called an LLM, is the type of AI behind ChatGPT and Claude. It writes text word by word, and that makes it good at playing a customer and explaining what worked in an answer. Ask it to decide the next step as well, and three things happen:

1. The assessment varies from one time to the next. Two course participants answer almost the same way. One gets “good, you acknowledged that the customer was irritated” and moves on. The other gets “you should have asked for the receipt” and a new attempt. Even the exact same answer can be assessed differently if you submit it twice.

2. The model does not know when it is uncertain. Ask it how certain it is, and that number is just more text it writes, and it is almost always high. The course therefore cannot separate the clear cases from the ones that should go to a supervisor.

3. The model does not stay on the course’s paths. The course developer has decided that the answer should land on one of four steps. The language model invents a fifth, justifies itself in three paragraphs, or answers 3.5. It can be pushed into a fixed format, but it still writes its way to the answer instead of weighing the options against each other.

On top of that, every assessment takes a few seconds, costs money and makes the experience sluggish. The result is often that the assessment is only done at a few points in the course instead of in every step.

Split the work

The solution is to let the language model keep talking and to place the decision with a model that may only choose among options the course developer has decided in advance. We call this a bounded decision. The question can be asked in a few different ways:

a. Choose an option: which of the five most common mistakes does the answer resemble most?
b. Place the answer on a scale: how well did the course participant meet the customer, from 1 to 4?
c. 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, and two identical answers end up in the same place.

The same exercise, two models

Back to the example with the jacket. The customer bought it two weeks ago. The zip has broken and the receipt is gone, but the customer stands their ground: “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 places the answer on the course developer’s scale from 1 to 4. A one is the course participant immediately saying “without a receipt we can’t do anything”. A four is first showing understanding of the customer’s irritation, then offering to look up the purchase via the card the customer paid with, and finally explaining what happens with the jacket.

If the answer lands on 1 or 2, the exercise is paused, the language model explains what was missing, and the course participant gets to try again. If it is 3 or 4, the next customer arrives, wanting a refund for a jumper they have used for three months. If the decision model is uncertain, the answer goes to a supervisor, with a short summary written by the language model. The difference from letting the language model decide is that everyone who answers the same way gets the same path, and that the uncertain cases are actually caught.

When the answer must follow a routine

Let us take one more example. In a course on fire safety, the course participant can ask questions about the company’s routine, for example 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.

With only a language model, there is a risk that it adds something that is not there, such as “if it only takes a minute”. Nobody notices until someone does exactly that. With two models, 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 explanation is shown. Otherwise the course participant reads the routine exactly as it is written. That is better than an answer that is easier to understand but wrong.

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, together with one or more questions of the type described above. The answers come back at once, each with a probability. Because it computes all the answers in a single pass instead of writing word by word, it is both fast and cheap.

According to TypeSafe, Jev responds in between 0.07 and 0.5 seconds, costs a fraction of what a language model costs per question, and is trained so that the probabilities can be trusted: if it says 0.8, it should be right roughly eight times out of ten. 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 do this. The split itself can be built with a language model that is forced to answer with one of the allowed options, or with a smaller model trained on answers that have already been assessed. That solves the answer landing on one of the course’s paths. What is harder to get from a language model is a probability that can be trusted, and that is what the course needs to send the right cases to a supervisor. Within a week of the launch, several open alternatives to Jev appeared, but according to the first comparisons it is precisely the probabilities that are hard to get right. Add to this that the decision is so fast and costs so little that it can be made after every answer, between the course participant’s line and the customer’s next one. If it takes several seconds, the wait is clearly noticeable every time, and the dialogue feels sluggish and less authentic.

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.