AI Tutor for Course Providers That Guides Without Giving Answers

Students get stuck most often when they're studying alone at home, outside lesson hours. A human tutor can't answer every question at any time, and students who stay stuck too long tend to give up or look for a ready-made answer online. An AI tutor can help fill that gap. But for a course provider, an AI tutor that simply answers everything does more harm than good: students copy the answer without understanding the method. This article looks at how an AI tutor can be designed to genuinely guide.

Why general chatbots don't suit students

General AI chatbots are built to help as fast as possible, so when a student pastes in a problem, they usually give the full solution straight away. For an adult at work, that's helpful. For a student who is learning, it removes the most important part of the process.

• Students see the final answer before they've had a chance to think.
• The method may differ from the one taught in class, which confuses students.
• The explanation can be wrong, and students don't yet know enough to notice.
• The course provider never learns which part of the material is actually causing trouble.

Staged hints, not instant answers

The heart of a good AI tutor is a hint ladder. Help comes in small steps, and each step only opens if the student still needs it. At Numekara, our own education business that is live today, the order is roughly this:

1. A concept explanation from the lesson module related to the problem.
2. A similar example with different numbers, so the student can follow the pattern.
3. The official hint written for that problem, given only when the student asks for deeper help.
4. The official worked solution, opened only when the problem's policy allows it, for example after the student submits an answer.

The order is set by the type of source, not by the AI model's judgement, so teachers can review and adjust it. Students still have to think at every step, while help remains available when they're truly stuck.

A ladder like this is also easy to explain to parents. They don't need to worry that their child is just copying answers from an AI, because the final answer isn't given before the student has made an effort. For a course provider, that's a clear selling point compared with leaving students to use a general chatbot unsupervised.

Guardrails that keep the answer key safe

Instructions to the AI model aren't enough on their own, because a model can slip and state the final answer. So the Numekara tutor checks every draft reply before a student sees it. Drafts that contain the answer key, including the same number written in a different format, are rejected. On top of that, the tutor may only cite sources that are allowed at that stage. If the official solution isn't open to the student yet, it isn't passed to the model at all.

Guardrails like these matter for course providers, especially for practice work and mock exams. Without them, practice scores stop reflecting what students actually understand.

Grounded in your official material

An AI tutor should answer from your own material: modules, example problems, hints, and solutions written by your teachers. That way the tutor uses the same terms and methods taught in class. Numekara, for example, has 543 lesson modules and a question bank of 9,754 static questions, and its tutor answers from published material and official solutions rather than from the model's general knowledge.

The tutor is available in practice sessions, lesson pages, and coding classes. If the AI service is having problems, students still get help from the official material and worked solutions.

Keeping costs under control

A course provider needs to know what it will spend. Numekara's AI tutor has daily quotas and spending caps, uses a cheaper model by default, and only calls a stronger model when a student genuinely needs it. For ordinary hint requests, the tutor first serves hints from official material without calling an AI model at all.

Teachers still matter

An AI tutor doesn't replace teachers. It helps students past small sticking points outside lesson hours, so face-to-face time can go to more valuable work, like tackling hard concepts or practising exam strategy. Students' question history can also show teachers which topics need re-explaining. If you store that history, treat it as students' personal data and look into your obligations under UU No. 27 Tahun 2022 tentang Pelindungan Data Pribadi (Indonesia's Personal Data Protection Law), ideally with a professional's help.

Next steps

1. Pick one subject or programme whose material is most complete.
2. Give every problem an answer key, staged hints, and a written solution.
3. Decide when the full solution may be shown, for example after the student answers.
4. Pilot it with a small group of students and have teachers review the tutor's replies.
5. Set quotas and spending caps before opening it to everyone.