How much of an AI product should you script?

I've been building an AI tutor for learning French. One of the things I want it to handle well is the moment someone gets stuck.

Should it give a hint? Explain the answer? Let them have another go? That is part of the appeal of building with AI: the next step can depend on what just happened.

But there is a catch. If the tutor shows someone the answer, it cannot then count their correct response as proof they knew it. Otherwise, it could move them on to a harder lesson on the strength of its own help.

I had tried to manage this by putting rules around when the tutor could reveal an answer. Working through one of those rules made me reconsider how much of the teaching I should be deciding in advance.

Why make someone try twice?

During an assessment, the default was to wait for two incorrect attempts before revealing the answer. After the first miss, the tutor could offer a clue and ask again.

There was a reasonable idea behind that. Sometimes a small prompt is enough. Jump straight to the answer, and you lose the chance to find out whether someone could work it out.

But “another attempt might help” had become a restriction: while the tutor stayed on that assessment path, it had to wait.

The test scenario behind the change made the effect concrete. It included a reply with an explanation and a corrected example after a first wrong answer. The rule would strip out the answer before the reply reached the learner.

An explanation was right there in the reply. The product could still send someone back for another attempt.

That did not prove the explanation was the best next step. It showed that I had already ruled it out.

Let it help, and remember that it helped

The requirement I cared about was that an answer supplied by the tutor should never count as independent work. That did not require the tutor to withhold the answer until a second miss.

I changed the assessment rule so it could show the current answer after the first graded miss. It could still choose a clue or ask again. If it revealed the answer, the product marked the exercise as practice with help. A later correct response would not count as unaided success, and answers to future exercises stayed hidden.

The tutor now had another option. The product still knew what kind of result it was looking at.

This is a question I would bring to other AI products: are we asking the model to follow a sequence because that sequence matters, or because we have not worked out how else to check its work?

A workflow can quietly make that decision for us. Once a step is written down and tested, following it starts to look like success. It takes a closer look at the experience to ask whether the step is helping.

Would you make the customer start over?

Imagine a support assistant working through a troubleshooting checklist. The customer says they have already tried the first three steps.

Making them repeat those steps might keep the workflow tidy. It could also make the conversation maddening.

Skipping ahead seems sensible, but the product still needs a reason to close the case. A suggested fix is not a confirmed resolution. The customer saying it worked could provide that confirmation.

The conversation can take a different route if the product keeps track of what has been tried, what has only been suggested, and what has actually worked.

Before giving an AI more choice, I would ask:

• What problem was this step supposed to prevent?
• If we let the AI take another route, how will we check the result?
• What would tell us that the experience got better?

Those questions also leave room for an answer of “keep the step.” If a refund needs approval, approval has to come before the money goes out. Recording what happened afterward would not make that acceptable.

The part I still need to learn

For the tutor, I checked that the explanation and current answer could get through after a wrong attempt, that a revealed answer stayed marked as practice with help, and that future answers remained protected.

That gives me room to test the teaching choice. I still need to learn when a clue is enough, when an explanation helps, and when another attempt is worth asking for.

This is what interests me as I keep building: how much better could the lesson become if the tutor had more room to respond? I can explore that while keeping an honest account of what the learner did independently.

Letting it choose is the start of that experiment. Whether the lesson improves will decide if the change was worth making.