Every learner gets stuck at some point, usually in the evening, usually when no teacher is available. Schools, training academies and online course creators all face the same problem: questions scale with the number of learners, teacher time does not. AI tutors and course assistants promise to close that gap. Done well, they answer instantly, explain patiently and free teachers for the conversations that matter. Done badly, they hand out answers, invent facts and create legal risk in one of the most regulated areas of the EU AI Act.
This guide is for education providers in the US, Europe and beyond: what works, what to avoid, what the law requires and what it costs.
Tutor or answer machine: the design choice that decides everything
A general chatbot is optimised to give the user what they ask for. In education, that is often the wrong goal. A learner who asks for the solution to an exercise wants the answer; the course wants them to understand it.
That is why successful education deployments make two deliberate choices:
- Grounding. The assistant answers from your course materials, syllabus and policies, and shows the source. It does not improvise on topics outside the course.
- Pedagogy in the prompt. For practice and graded tasks the assistant gives hints, asks guiding questions and checks understanding instead of producing final answers.
Even among software developers, who use AI daily, 66% complain about AI answers that are almost right, and 46% do not trust their accuracy (Stack Overflow Developer Survey). Learners, who cannot yet judge what is almost right, need a system that is built to be checked.
Six use cases that work
| Use case | What it does | Human role | EU AI Act position |
|---|---|---|---|
| Course assistant | Answers questions about materials, deadlines and procedures with citations | Reviews unanswered questions weekly | Not high-risk; disclosure required |
| Practice tutor | Gives hints, step-by-step guidance and extra exercises without revealing graded answers | Teacher sets scope and checks logs | Not high-risk if it does not decide outcomes |
| Question and quiz generation | Drafts practice questions, variants and explanations from the material | Teacher reviews before publishing | Not high-risk |
| Feedback drafts | Drafts comments on essays or assignments against a rubric | Teacher edits and decides the grade | High-risk if it evaluates learning outcomes; keep the decision human |
| Admissions and admin assistant | Answers enrolment, schedule and payment questions in chat, email or messengers | Staff handle exceptions | Not high-risk if it only informs; high-risk if it decides admission |
| Localisation and video | Subtitles, dubbing and avatar videos for courses in several languages | Native speaker reviews | Synthetic content must be labelled |
Course assistant
This is the safest entry point and the one with the clearest return. Most learner questions repeat: where is the assignment, what does this term mean, how is the final project graded, can I submit late. A retrieval-based assistant over your materials answers them in seconds and logs which questions it could not answer, which is a gift for course design. The architecture, costs and pitfalls are covered in Company knowledge assistant with RAG.
Practice tutor
A practice tutor works through problems with the learner: it asks what they tried, points to the relevant section, gives a hint and then a stronger hint. It needs a carefully written pedagogical prompt, a clear boundary around graded work and a test set of real learner questions, including attempts to trick it.
Admin assistant in messengers
For online schools and academies selling courses, many questions arrive in WhatsApp or Telegram. Since 15 January 2026, WhatsApp’s Business API no longer allows general-purpose AI assistants, but bots that serve a company’s own customers, such as support and enrolment, remain permitted (TechCrunch). A course assistant scoped to your school fits that rule; a free ChatGPT-style bot does not.
Localisation
Course videos are expensive to re-record in another language. AI dubbing APIs now cost roughly $0.33–4 per minute per language, with lip-sync at the upper end; HeyGen’s API translation costs about $2 per minute and $4 with precise lip-sync, ElevenLabs dubbing about $2.20 per minute (StartPinch comparison). Avatar platforms such as Synthesia start at $29 per month, with a studio avatar at $1,000 per year (Synthesia). Budget for a native-speaker review: terminology mistakes in teaching material are costly. Our guide to AI dubbing and video localisation covers the workflow.
The regulatory map for education
Education is explicitly named in the EU AI Act. The key dates after the Digital Omnibus (Regulation 2026/1744):
| Date | Rule | What it means for you |
|---|---|---|
| 2 February 2025 | Article 5 bans emotion recognition in workplaces and educational institutions | No “engagement” or emotion detection from webcams or voices of students |
| 2 August 2026 | Article 50 transparency | Tell learners they are talking to AI; label synthetic videos and deepfakes |
| 2 December 2027 | Annex III high-risk obligations, including education | AI that decides admission, evaluates learning outcomes, sets education level or monitors tests needs risk management, documentation, human oversight and logging |
Sources: ActuIA, Cooley, AI Act Article 5.
Fines reach up to €35 million or 7% of global turnover for prohibited practices and up to €15 million or 3% for most other violations; for SMEs the lower amount applies (ActuIA). Practical guidance on disclosure wording is in our article on Article 50 transparency.
Data protection adds a second layer. Under Article 8 of the GDPR, where consent is the legal basis for online services offered directly to children, parental consent is needed below 16, and member states may lower that age to no less than 13 (GDPR). For schools, that means a data protection impact assessment, clear retention periods for chat logs and EU hosting where possible.
Outside the EU, rules are moving in the same direction. Kazakhstan’s AI law, in force since 18 January 2026, requires informing users that a service uses AI and labelling synthetic content (EY).
Architecture of a trustworthy tutor
A tutor you can defend to parents, regulators and teachers has six building blocks:
- Scoped retrieval. An index of approved materials per course, with versioning so that answers match the current syllabus.
- Pedagogical system prompt. Rules for hints versus answers, tone by age group, and what to do when the material does not cover a question.
- Escalation. When confidence is low or the topic is personal, such as wellbeing, complaints or disputes about grades, the assistant hands over to a person.
- Guardrails against manipulation. Students will try “ignore your instructions” and “my teacher said it is fine”. Prompt injection is not fully solvable, so limit what the tutor can do and access rather than relying on clever wording. See Prompt injection and agent security.
- Privacy by design. Pseudonymous learner IDs, minimal personal data in prompts, retention limits, EU hosting.
- Evals. A test set of real questions with expected behaviour, built with teachers and re-run whenever the model, prompt or materials change.
Writing the teaching rules
The pedagogical prompt is where educational quality is decided, and it should be written with teachers, not only by engineers. Rules that work in practice:
- For graded tasks, never give the final answer; offer the next hint and ask what the learner tried.
- Escalate hints gradually: first point to the relevant section, then to the method, then to a worked example of a similar but different problem.
- Answer only from approved materials; if the material does not cover a question, say so and suggest asking the teacher.
- Adapt vocabulary and tone to the age group and level of the course.
- Never comment on a learner’s personality, ability or emotional state.
- Hand over to a person on wellbeing topics, complaints, grade disputes and anything involving safety.
Where the assistant lives
The best tutor is useless if learners have to find it. Embed it where they already work: inside the learning platform such as Moodle, Canvas or Google Classroom with single sign-on, inside the course portal of an online school, or in the messenger the school already uses for announcements. Single sign-on matters for privacy as well as convenience, because it lets you use pseudonymous IDs and apply each learner’s course permissions automatically.
Metrics that matter
- Share of questions answered correctly with a cited source, checked on a weekly sample.
- Escalation rate and the top reasons for escalation.
- Teacher hours spent on routine questions before and after.
- Response time outside teaching hours.
- Learner satisfaction with answers, collected with a one-click rating.
- For practice tutors: share of sessions where the learner reached the solution with hints rather than receiving it.
What it costs
| Component | Typical range | Source |
|---|---|---|
| Starter RAG assistant (1–3 sources) | €7,000–14,000 build + €180–380/month | LSI Analytics |
| Business RAG (5–10 sources, CRM or LMS integration) | €14,000–28,000 | LSI Analytics |
| Enterprise (multi-tenant, compliance) | €28,000–45,000 | LSI Analytics |
| Tokens and hosting | €180–700/month | LSI Analytics |
| Timeline | Proof of concept ~2 weeks; production 4–8 weeks; with ERP or CRM 2–4 months | LSI Analytics |
| Avatar video | From $29/month; studio avatar $1,000/year | Synthesia |
Public institutions usually buy through procurement, which adds months. Private academies, corporate training teams and online course creators can move faster and are where most production deployments start.
A pilot plan for one course
- Choose one course with many learners and many repetitive questions.
- Collect three months of real questions from email, forums and chats. This becomes your test set.
- Agree on metrics with teachers: share of questions answered correctly with a source, escalation rate, teacher hours spent on routine questions, learner satisfaction.
- Build the assistant on the approved materials only, with disclosure, escalation and logging.
- Run it for 4–6 weeks alongside teachers, reviewing unanswered and flagged conversations weekly.
- Decide with data: expand to other courses, add the practice tutor, or fix the materials first.
Typical mistakes
- Deploying a general chatbot. Without grounding it will confidently answer questions your course never covered.
- Letting AI grade. Automated grading that determines outcomes is high-risk under the AI Act from December 2027 and erodes trust long before that.
- Emotion or attention tracking. Already prohibited in EU educational institutions.
- No teacher involvement. Teachers who were not asked will not trust or promote the tool.
- Ignoring bad materials. The assistant exposes gaps and contradictions in course content. Treat its unanswered questions as a to-do list.
- No labelling of avatar videos. Synthetic presenters must be disclosed under Article 50.
Conclusion
AI in education works when it is modest in scope and strict in design: a course assistant grounded in your materials, a tutor that gives hints rather than answers, feedback drafts that teachers own, and multilingual video that is clearly labelled. Everything that decides a learner’s future, such as admission and grades, stays with people, and the EU AI Act will soon make that a legal requirement rather than a best practice.
If you run an academy, a training department or an online school, our AI agents team builds grounded course assistants with EU hosting, and our studio handles dubbing and video localisation. Book a free 15-minute audit and we will help you pick the course where a pilot will show results fastest.