Teacher assistant: lesson planning & differentiation

Teachers spend many unpaid hours on preparation, and classes are more heterogeneous than ever: mixed reading levels, new-to-the-language pupils, learning support and high achievers in one room. An LLM-based teacher assistant drafts lesson plans against the curriculum and produces the same task in several versions (simpler language, scaffolds, extension). The teacher reviews, adapts and owns the result; the tool works on materials, not on judgements about individual pupils.

Problem & value

• Today: lesson preparation is manual: searching textbooks and portals, adapting worksheets by hand, creating separate versions for weaker and stronger pupils. Differentiation is widely wanted but often skipped for lack of time.
• AI value: a first draft in minutes (structure, timing, methods, materials), automatic leveling of texts and tasks, translation and simplified-language variants, ideas for scaffolding. Time saved is shifted to feedback and relationships.
• Evidence: the OECD Teaching and Learning International Survey (TALIS 2024) reports on teachers' use of AI and on workload; the UK Department for Education has published guidance and evaluation material on generative AI in education, including lesson planning use [1][2]. An EEF-funded randomised trial (NFER, 10 weeks, Year 3–6 teachers in England, published Oct 2026) found that Oak's Aila cut lesson and resource preparation by about 49 minutes a week with no evidence of lower resource quality, although it supported only a small share of lessons [9]. Treat vendor time-saving claims as unproven until measured in your own school. Quality of generic output is mixed: plans often look plausible but are didactically shallow without teacher edits.

Variants

• Primary / secondary school teacher: lesson sequences, tiered worksheets, class-specific scaffolds.
• Vocational training (VET) and adult education: competence-oriented learning situations (e.g. German Lernfelder), practice-related tasks.
• Inclusion & language support: simplified German (Leichte/Einfache Sprache), DaZ materials, multilingual glossaries, support for dyslexia-friendly formats.
• Higher education: seminar plans and tutorial exercises (see EDU-001).
• Substitute-teacher and cover plans; parent communication drafts.
• Private / home-schooling parents: simple prompts with a general assistant; same caveats.

How it works

1. teacher request (topic, grade, time, goals, class profile - anonymous)
2. retrieval over curriculum standards + own/OER materials (RAG)
3. draft lesson plan (phases, methods, timing, materials, assessment idea)
4. differentiation engine (levels, language, scaffolds, extensions)
5. format/export (docx, slides, LMS, H5P)
6. TEACHER REVIEW & edit
7. use in class
8. feedback
9. saved in school library

• Curriculum grounding: curriculum documents (e.g. German Lehrpläne/Bildungsstandards) are chunked and embedded so each plan cites the competence it addresses; this reduces invented standards.
• Differentiation: one source task is rewritten along dimensions the teacher selects (reading level, language level, scaffolding, open vs. guided task); the same learning goal is kept.
• Human-in-the-loop: the teacher approves every output; fact checking of content (dates, formulas, sources) is mandatory. Class profiles are described in aggregate ("3 pupils reading at A2"), not by name.

Data preparation

The main inputs are the curriculum or syllabus, the school-internal curriculum, textbooks and OER, and teachers' own earlier lesson plans. They typically live in PDFs from the ministry, on school file shares, in learning platforms (e.g. Moodle, Teams) and in teachers' private folders. Typical gaps: curricula only as scans or long PDFs, outdated plan versions, duplicated worksheets, no tags for grade, topic or level, and copyrighted textbook pages that may not be uploaded freely.

Preparation steps: collect the current curriculum and internal standards in one folder, convert scanned pages with OCR (see PREP-006), drop outdated versions, name files by grade, subject and topic (PREP-010), and curate a small set of good example plans as templates (PREP-016). Class profiles must be written without names, as a pseudonymised description (PREP-022). Check licences before indexing textbook content.

Tools: Paperless-ngx or a simple OCR pipeline for scans, a shared folder with consistent naming, and the built-in knowledge upload of the chosen assistant. Rough share of effort: low, about 10-20 percent; a first usable setup takes one to two days, since the assistant works without any private data too. Also see DOC-014 for turning documents into clean text.

Implementation paths

Which path for whom

• public: B - schools and authorities should use a state or sovereign EU-hosted platform with API access, because pupil data, procurement rules and sovereignty favour it over a consumer product.
• smb: A - private schools, training providers and tutoring firms get a mature, EU-hostable teacher assistant at low per-seat prices; the process is standard.
• developer: B - build on EU-region model APIs and the curriculum as retrieval source; self-hosting rarely pays for this workload.
• private: A - a hosted assistant (or a local app such as LM Studio for sensitive notes) covers individual lesson preparation.

Path A - Buy (recommended for private schools, training providers and individual teachers)

• School-specific assistants: fobizz (DE, EU-hosted AI tools for teachers), Telekom/Schulcloud and state-provided services such as BaWü's telli and similar Länder offerings (check your state), Aila by Oak National Academy (UK), Khanmigo teacher tools (Khan Academy), MagicSchool, Brisk Teaching, Diffit (leveled texts), Eduaide.
• Suites: Gemini in Google Classroom / Workspace for Education, Microsoft 365 Copilot and Microsoft Learning Zone / Teach module, Canva for Education Magic Write.
• Check: EU hosting or DPA with SCCs, no training on inputs, age restrictions, curriculum coverage for your country, export to your LMS (Moodle, itslearning, Schulportal).
• Choose A when: a school or authority wants low-effort rollout with a vetted DPA.

Path B - Build on cloud APIs

• Models: Mistral (EU provider), Claude, Gemini, GPT via EU regions (Azure OpenAI EU Data Zone, Vertex AI EU, Bedrock EU); structured output for lesson-plan JSON schemas.
• Stack: RAG over curriculum and OER (embeddings + vector DB), SSO via school IdP, LTI integration into the LMS, PII filter on prompts, logging without pupil data.
• Choose B when: a ministry, school authority or EdTech company serves many schools and wants own curriculum alignment or a branded tool.

Path C - Self-host (open source)

• Models: open-weight Mistral Small / Mixtral class, Qwen3, Llama, Gemma, served with vLLM or Ollama; embeddings such as BGE-M3 or multilingual-e5.
• UI/stack: Open WebUI or LibreChat, Qdrant/pgvector, Docling for parsing textbooks and PDFs.
• Hardware: HW2 serves a school or a small district with a ~30B quantised model; HW3 for a regional service. German-language quality of small models for simplified language needs testing.
• Choose C when: a municipality or school authority wants data sovereignty.

Path D - Custom training

Rarely needed. Fine-tuning on high-quality, rights-cleared lesson plans or leveled texts may help for a specific language (e.g. Easy-to-Read) or readability control; most gains come from RAG and prompt templates instead.

State of the art (2026)

• Curriculum-grounded generation with citation of standards is the main quality driver; generic chat prompts give generic plans.
• Readability control: LLMs can rewrite texts to target CEFR or grade levels, but the actual level is often off; verify with readability metrics or classifiers and a teacher read-through.
• Multimodal: photo of a textbook page or handwritten task → differentiated version via VLM.
• Agentic suites link planning, worksheets, slides and LMS posting (Gemini in Classroom, Copilot, MagicSchool).
• Limits: hallucinated facts and sources, culturally or age-inappropriate examples, bias in generated names and scenarios, "one-size" didactics, copyright issues when copying textbook content, over-reliance reducing teachers' own pedagogical reasoning. Differentiation by an LLM is not a diagnosis of pupils' needs.

Shared resources & combinations

• Same stack (llm-general + RAG + guardrails): EDU-001 course material, EDU-002 quiz items, EDU-016 study tools, KNOW-002 enterprise RAG; one self-hosted LLM serves several of them.
• Downstream: EDU-003 feedback on pupil work, EDU-005 adaptive paths, LANG-009 translation of materials, EDU-017 screening is a separate high-risk-adjacent use case.
• Related planning: OPT-013 timetabling.

Effort, cost & KPIs

• Path A: days; often free for teachers or licensed per school/state.
• Path B: 2-4 months for a production service with curriculum RAG and LMS integration.
• Path C: 2-4 months plus operations by the school authority's IT.
• KPIs: teacher-reported preparation time (surveyed, not assumed), share of drafts used with edits, teacher quality rating, factual-error rate in spot checks, adoption rate, number of differentiated versions produced per unit.

People & change

Involve management early for the AI policy and budget (GUIDE-007), the head of department or school principal as process owner, and the teachers as daily users from the start. IT clarifies accounts and the platform; the data protection officer checks the AVV and rules for pupil data, and a DPIA is needed as soon as pupil data or assessments enter the tool (GUIDE-005). Hiring or evaluation of people is not involved.

Works council: depends. The assistant plans lessons and does not assess staff, but a product that logs individual usage is suited to monitor behaviour, so co-determination under section 87(1) no. 6 BetrVG applies in companies with a works council; section 90(1) no. 3 requires timely information about AI-supported work processes. In public schools the staff council (Personalrat) under the state staff representation law is the counterpart.

AI literacy: the AI Act Art. 4 asks providers and deployers to support the AI literacy of staff, so offer short hands-on sessions on prompting, checking facts and bias (GUIDE-006). Start with a pilot group of two to five volunteer teachers, name champions, and collect their feedback (GUIDE-008, GUIDE-009). Typical concerns: loss of teaching quality, extra workload, deskilling and copyright. Address them by stating that the teacher always reviews and decides, and that no pupil names go into prompts.

Risks, regulation & compliance

• EU AI Act: typically limited/minimal risk: generating teaching materials is not an Annex III use. Annex III point 3 (education) is high-risk for admission, evaluating learning outcomes, assessing the appropriate level of education and monitoring prohibited behaviour in tests. The boundary is crossed if the assistant evaluates individual pupils, assigns them to levels or feeds grading and progression decisions; then see EDU-003/EDU-005. Art. 50 transparency applies to AI-generated content published as such; Art. 4 AI literacy duty applies to schools and teachers. Art. 5 prohibition: emotion recognition of pupils in educational institutions is banned; do not add affect detection. Annex III obligations apply from 2 Dec 2027 under the Digital Omnibus on AI (Reg. (EU) 2026/1744).
• GDPR: do not enter pupil names, grades, diagnoses or special-needs data into non-vetted tools; data on learning disabilities is special-category-adjacent and children merit enhanced protection (Recital 38). Use school-approved tools, DPA, DPIA where pupil data is processed; German state school laws and data protection authorities issue specific guidance.
• Copyright: respect textbook publisher rights; prefer OER and own material.
• Mitigations: teacher approval, anonymous class profiles, source citations, fact-check checklist, bias review, transparency to pupils/parents about AI-assisted material.

Getting started

1. Pick one subject team and 2-3 recurring tasks (tiered worksheet, leveled text, lesson skeleton).
2. Select a vetted tool (Path A) with DPA and clarify a school rule on what data may be entered.
3. Build prompt templates with curriculum references and your differentiation dimensions.
4. Run a 6-week pilot, collect teacher time and quality feedback, fact-check samples.
5. Share good prompts and plans in a school library and train staff (AI literacy, Art. 4).

Sources

1. OECD - TALIS 2024 results - https://www.oecd.org/en/about/programmes/talis.html
2. UK Department for Education - Generative AI in education - https://www.gov.uk/government/publications/generative-artificial-intelligence-in-education
3. Regulation (EU) 2024/1689 (AI Act), Annex III and Art. 5 - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
4. UNESCO - Guidance for generative AI in education and research - https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research
5. 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
6. Oak National Academy - Aila lesson assistant - https://labs.thenational.academy/
7. Khan Academy - Khanmigo for teachers - https://www.khanmigo.ai/teachers
8. fobizz - AI tools for teachers - https://fobizz.com/
9. Education Endowment Foundation — Lesson planning using AI lesson assistant Aila (Teacher Choices trial, NFER evaluation, Oct 2026) — https://educationendowmentfoundation.org.uk/projects-and-evaluation/projects/aila-teacher-choices-trial