About us Services Cases Pricing Blog Contact Get a quote →
RUEN
AI tutor: a tablet with a study programme, a textbook and pencils on a desk by the window
AI for business

OM-Digital expert opinion: where an AI tutor teaches and where a teacher is needed

26 October 2026  ·  ~5 min read  ·  OM-Digital

What studies of AI tutors show, how to build a learning programme on a neural network and where a teacher remains essential

Published on RBC Companies  ·  26 October 2026  ·  Image source: generated by the Gemini neural network

In June 2025 the journal Scientific Reports published an experiment by Harvard University. 194 physics students took some topics with an AI tutor and others in class using active learning. The median learning gain with the AI tutor was more than twice as high, and students spent an average of 49 minutes on a lesson versus a 60-minute class.

The World Bank obtained a similar result in Nigeria. In 2024 schoolchildren studied English after school with an AI assistant for six weeks, and the gain was about 0.3 standard deviations. The authors estimate this as almost two years of regular schooling. In Russia students have already made the choice themselves: according to a 2025 HSE University survey, about 87% of students at leading universities use AI in their studies.

An important detail: in both experiments AI was not simply an open chat. It was set up as a teacher and built into the learning process. At OM-Digital we have made a learning app in which the Claude neural network itself helps explain. Below is what from this experience applies to any learning programme, and the limits where a person is needed.

Short answer

An AI tutor helps people learn when it is set up as a teacher: it follows the programme, asks leading questions, relies on verified materials and is available at any time. The key here is the setup, not the model itself. A teacher is still needed for rhythm, motivation, control and everything related to the final assessment.

What turns a chat into a tutor

This is clearly seen in a study at a Turkish school published in PNAS in 2025. About a thousand high-school students solved maths problems with GPT-4: some with a regular chat, others with a version set up as a tutor. In practice sessions grades rose by 48% and 127% respectively. The same model, but the teacher setup gave a gain more than two and a half times higher.

The difference is in behaviour. A tutor first asks what is already clear, suggests the next step and gives the answer once the person has tried on their own. This way the knowledge stays with the learner, not in the chat.

How AI becomes a teacher

Through skills. In October 2025 Anthropic, the developer of Claude, introduced them as folders with instructions, scripts and materials that Claude loads when they are needed for a task. Put simply, it is a manual for the assistant: its role, the order of explanation, sources, what is allowed and what is not.

We work on this principle ourselves: each line of work in the company has its own set of such instructions, and the assistant loads the right one before a task. The approach for learning is the same. One and the same AI becomes a maths tutor for seventh grade, a spreadsheet mentor for an accountant or a coach for working with neural networks — only the manual changes. The programme can be delivered as an app, a bot or a section on a website.

What materials to build the programme on

A reliable foundation is what a person already learns from: a textbook or study guide approved at school or university, the teacher’s course notes, the company’s internal procedures. Then the assistant explains in the logic of the course rather than retelling random pages from the internet.

There are two conditions here. The first is the rights to the materials: textbooks are protected by copyright, so the programme is loaded with your own course notes, open publications or materials whose use is permitted. The second is checking: the teacher reviews the assistant’s answers to the course’s typical questions before launch and selectively afterwards.

If the assistant works with learners’ data, a third condition is added — the law. Under Article 9 of the Personal Data Law, consent to processing must be voluntary, specific and informed. So it is better not to load anything into learning scenarios that makes it possible to identify a particular child.

We have made an Android app for learning to work with Claude: inside it the neural network answers questions on the course, guides you step by step and helps word a request. The frame is set by an instruction: stay on the course topic and explain with examples from real tasks. The assistant answers automatically, at any time of day.

Where a person is indispensable

The authors of the Harvard study put it plainly: “An AI tutor should not replace in-person teaching: its job is to bring all students up to the level at which they get the most out of classroom sessions.” In the Nigerian pilot there were teachers next to the pupils too: they made sure the children stayed on task and understood the hints.

The state has enshrined the same logic. The rules for using AI at school that the Ministry of Education published on 26 September 2026 state that the final decision is always made by the teacher, pupils in grades 1–4 do not use AI on their own, and AI is not allowed in interim and final assessments.

This adds up to three zones for a person:

Where to start

With one topic and one course. Describe the goal, the order of topics, the materials and the rules of the assistant’s behaviour, for example “first a leading question, the finished answer only after an attempt”. Then test it on a dozen typical questions together with a teacher, and only then open it to learners.

Key takeaways

We will build an AI tutor for your course

Send us the programme and materials — we will set up a teaching assistant and test it on typical questions together with you.

AI assistants for your tasks

Related reading