Pupils, students and employees get stuck outside lesson hours, and one-to-one human tutoring does not scale or is unaffordable. An AI tutor answers questions, gives graded hints instead of final answers, explains in several ways and adapts to the learner. Done badly it becomes an answer machine that undermines learning; done well (pedagogy-constrained, curriculum-grounded, teacher-supervised) it extends support to every learner at any hour.
Problem & value
- Today: homework help comes from parents, paid tutors, forums or - increasingly - general chatbots that simply give the solution. Teachers cannot answer 25-300 individual questions per day.
- AI value: 24/7 individual support, patience, multilingual explanations, low threshold for asking "stupid" questions, and aggregated insight for teachers on where the class struggles.
- Evidence is mixed and design-dependent. Controlled studies of tutoring-style AI report learning gains when the tutor is built around pedagogy (hints, guardrails against giving answers), while unrestricted chatbot access can improve practice performance yet hurt later unaided performance [4][5]. The central design question is therefore how the tutor helps, not whether it answers correctly.
Variants
- School (K-12) homework helper: age-appropriate, curriculum-aligned, teacher-configured, strong child-safety filters; deployed via the school with parental/legal basis.
- University course tutor: RAG over lecture notes and papers, office-hours overflow, code and maths tutoring; instructor controls what the tutor may reveal.
- Corporate learning coach: on-the-job Q&A over internal training and policies (see EDU-006, EDU-010).
- Private use: consumer learning modes in general assistants, photo-of-the-task apps, language practice (see EDU-011) and own-notes study tools (EDU-016).
How it works
- learner question (text / voice / photo of task)
- input safety + age/role check (PII filter, jailbreak & self-harm routing)
- task understanding (OCR/VLM for photos; map to curriculum topic)
- retrieval of course material (RAG over teacher-approved corpus)
- tutor policy (system prompt + rules: hint ladder, no final answer unless allowed)
- LLM/reasoning model generates guidance with citations
- verification (solver / code execution / answer key check for maths & code)
- response (text/voice) + optional practice item
- logging, teacher dashboard, escalation to human on distress or repeated failure
- Pedagogy layer: hint ladder (clarifying question → nudge → worked analogous example → step check), retrieval practice, asking learners to explain their reasoning. This is configured by teachers, not left to the base model.
- Grounding: RAG with citations to the course corpus reduces hallucination; for maths/code a tool (symbolic solver, sandboxed code run, reference solution) verifies steps.
- Human-in-the-loop: teachers set scope and rules, review aggregated logs, and receive escalations; summative grading stays with humans (see EDU-003, EDU-008).
Data preparation
A tutor is only as good as the curriculum content behind it: textbooks, scripts, exercises with worked solutions, glossaries and teaching guidelines. These sit in the LMS, on file shares, in publisher PDFs and on paper. Typical gaps: no worked solutions, scans without text, mixed levels and languages, outdated or contradictory versions, licence limits on publisher material and scattered know-how of experienced teachers.
Preparation steps: curate one authoritative knowledge base per subject and level (see PREP-013); rewrite key documents for retrieval where needed (see PREP-014); build a glossary of terms and abbreviations (see PREP-015); OCR scans (see PREP-006); collect typical learner questions with good answers as test set (see PREP-018); define feedback loops to fix wrong answers (see PREP-017, PREP-021). Parse sources with DOC-014 and check licences.
Effort: medium. Expect 25 to 30 percent of the total effort, mostly content curation and the test questions. See GUIDE-003.
Implementation paths
Which path for whom
- SMB: Path A, because EU-hostable tutoring and training assistants exist at SMB prices and cover standard courses.
- Enterprise: Path B, because the tutor must be grounded in own training content and integrate with LMS and single sign-on.
- Developer: Path B, because building a tutoring product on EU-region APIs with RAG over the curriculum is faster than hosting models.
- Public: Path C on sovereign infrastructure, because minors' data and school rules call for German or EU hosting and control.
- Private: Path A, because certified or well-known learning apps work out of the box; local apps are an option for privacy-minded learners.
Path A - Buy (recommended for most schools and SMB training providers)
- Education-specific tutors: Khanmigo (Khan Academy), Duolingo Max (languages), Google Gemini for Education / Guided Learning, OpenAI ChatGPT Edu with study mode, Anthropic Claude for Education (learning mode), Microsoft Copilot with Learning Zone/Teams for Education.
- EU / DACH vendors: fobizz (DE, teacher tools and school-safe chatbot access), telli (DE, school chat platform for German states), Knowunity and Simpleclub (DE, study platforms with AI tutors), Sana Labs (SE, corporate learning), Century Tech (UK, adaptive learning).
Path B - Build on cloud APIs
- Models: Claude, GPT, Gemini (EU regions via Vertex AI, Azure OpenAI, AWS Bedrock Frankfurt/ Paris) or Mistral (EU provider, La Plateforme) for the tutoring dialogue; a reasoning model for maths; a small model for safety classification and routing.
- Stack: RAG (embeddings + vector DB such as pgvector, Qdrant, Weaviate), LangChain/LlamaIndex or a plain service, LTI 1.3 integration into Moodle/Canvas, SSO via school identity (e.g. Shibboleth, Entra ID, VIDIS in Germany).
- Choose B when: you need your own curriculum, branding or LMS integration but no on-prem constraint.
Path C - Self-host (open source)
- Models: open-weight LLMs such as Qwen3, Llama 3.x/4, Mistral Small/Medium open releases, gpt-oss, Gemma 3; served with vLLM, Ollama or llama.cpp; embeddings with BGE-M3 or multilingual-e5; Whisper for speech input; Open WebUI as front end with Moodle/LTI bridge.
- Hardware: HW2 for a ~8-30B quantised model for a school or course; HW3 for a ~70-120B model or hundreds of concurrent users; a shared school-district server is the typical setup.
Path D - Custom training
Fine-tune only for tutoring style (hint ladders, age-appropriate German) with a few thousand teacher-reviewed dialogues; rarely needed before prompt and RAG design are exhausted.
State of the art (2026)
- Pedagogy-constrained modes are now standard in major assistants (study/learning/guided modes that ask questions rather than answer) and in dedicated tutors; they are system-prompt and policy layers on frontier models rather than separate models [1][2][3].
- Reasoning models solve most school and undergraduate maths and science tasks, but step-level errors, wrong "confident" explanations and sycophancy toward wrong learner answers persist; verification by tools and answer keys is still advisable.
- Research findings: tutoring-style AI shows promise in controlled studies, while uncontrolled answer-giving access can cause over-reliance [4][5]. Learner motivation and metacognition, not model quality, are the binding constraints.
Shared resources & combinations
- Same stack (llm-general + vector-db + embedding-text + guardrails): EDU-001 content generation, EDU-016 study tools, EDU-009 teacher assistant, DOC-014 document parsing for RAG, and enterprise knowledge assistants - one RAG platform and one self-hosted LLM server (HW3) can serve all of them.
- Natural combinations: EDU-003 feedback on submissions, EDU-005 adaptive paths (tutor interactions feed the learner model), EDU-007 learning analytics (aggregated, non-punitive), EDU-011 language practice, EDU-010 role-play training.
Effort, cost & KPIs
- Path A: days to weeks; licences per learner or school. Path B: 2-4 months; token cost scales with usage.
- Path C: 3-6 months incl. hardware and ops staff; no per-token cost but fixed GPU and staff cost.
- KPIs: learning gain vs. control group or prior cohort, unaided performance after the tutor period, share of hints vs. final answers given, teacher time saved, usage by learner group (equity), escalation rate, hallucination rate in sampled audits, learner and teacher satisfaction.
People & change
Involve management for the pedagogical and financial decision, teachers or the training lead as content owners and reviewers, IT for SSO and hosting, the DPO from the start (learners, partly minors), legal for child protection and copyright, and the learners or customers via a pilot. IT security checks prompt injection and misuse.
Works council: depends. If the tutor serves employee training, usage logs per person are objectively suited to monitor performance, so §87(1) no. 6 BetrVG applies; inform the council in any case about AI in work processes (§90(1) no. 3). For pupils and private learners no works council is involved; schools involve the Personalrat when staff are affected.
Data protection: conversations with learners are personal data, a DPIA is likely for minors or at scale, an AVV is required, and no vendor training on learner data should be agreed. Under the AI Act, transparency applies (Art. 50): learners must know they talk to an AI.
AI literacy: support staff and learners with short training on limits, hallucinations and not outsourcing thinking (AI Act Art. 4). Start with one class or course as pilot, name an AI champion, and keep teachers in control with dashboards and escalation to a human. Typical concerns are cheating, wrong answers and loss of the teacher role; address them with guardrails and clear rules. See GUIDE-005, GUIDE-006 and GUIDE-008.
Risks, regulation & compliance
- EU AI Act: classification depends. A tutor that only explains and hints is typically limited risk (chatbot transparency, Art. 50: tell users they talk to AI) plus AI literacy duty (Art. 4). Systems that evaluate learning outcomes, steer the learning process, determine access or assign learners to institutions, or monitor prohibited behaviour during tests fall under Annex III point 3 (education) and are high-risk, with obligations for providers and deployers [6]. Adaptive features that grade or decide progression therefore need careful scoping.
- Prohibited practice (Art. 5): inferring learners' emotions in education institutions (e.g. from webcam or voice to detect frustration or attention) is banned except for medical or safety reasons. Do not add affect-sensing to a tutor [6].
- GDPR: learner data, often of children, is personal data; DPIA is usually required; legal basis (public task or consent, parental consent for young children), data minimisation, no training on pupil chats, EU hosting or valid transfer mechanism, retention limits. German states publish school-specific guidance; consult the data protection officer [7]. assessment design.
- Mitigations: teacher-configured scope, hint-first policy, citations, tool verification, logging with access limits, human escalation, periodic audits, AI-use disclosure to learners and parents.
Getting started
- Define the learning goal and the rule per task: what help is allowed, what must stay unaided.
- Pilot with one subject and 1-2 classes using a Path A education tool or an LMS-embedded assistant; involve teachers in setting hint rules.
- Run the DPIA, DPA and age/consent check before any pupil data flows.
- Build a small gold set of 50-100 real learner questions to test correctness, hint behaviour and jailbreak resistance before launch.
- Measure unaided learning outcomes against a comparison group after 6-8 weeks; then decide on scale-up or a Path B/C build.
Sources
- OpenAI - Introducing study mode - https://openai.com/index/chatgpt-study-mode/
- Anthropic - Claude for Education (learning mode) - https://www.anthropic.com/education
- Khan Academy - Khanmigo - https://www.khanmigo.ai/
- Bastani et al. - Generative AI Can Harm Learning (PNAS 2025) - https://www.pnas.org/doi/10.1073/pnas.2422633122
- Kestin et al. - AI tutoring outperforms in-class active learning (Scientific Reports 2025) - https://www.nature.com/articles/s41598-025-97652-6
- Regulation (EU) 2024/1689 (AI Act), Art. 5, Art. 50, Annex III - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- European Commission - Ethical guidelines on the use of AI and data in teaching and learning for educators - https://education.ec.europa.eu/news/ethical-guidelines-on-the-use-of-artificial-intelligence-and-data-in-teaching-and-learning-for-educators
- Moodle - AI subsystem documentation - https://docs.moodle.org/en/AI_subsystem