Every apprentice in Germany must keep a training record, and the trainer must review it. In practice it is written late, in one-line phrases, on Friday afternoon, and nobody compares it with the training framework plan until the exam admission is at stake. An AI step lets apprentices speak or type a few raw notes, produces a clean draft in their own words, assigns each activity to the framework plan and shows the trainer which learning objectives are still open. The apprentice remains the author; the trainer remains the reviewer.
Problem & value
- Today: apprentices must keep an Ausbildungsnachweis, in writing or electronically (§13 sentence 2 no. 7 BBiG). A complete record is a condition for admission to the final exam (§43 (1) no. 2 BBiG), and trainers are expected to review and sign it regularly, often weekly [1][2]. In small craft firms this is squeezed into the end of a long working day, entries are formulaic, and the comparison with the Ausbildungsrahmenplan is manual work that many firms skip, so gaps show up only shortly before the exam [3].
- AI value: lower writing barrier (especially for apprentices with weak German or writing difficulties), better entries, automatic coverage check against the framework plan, early warning on missing topics (for example a trade where no commissioning work has been done in 18 months), less chasing by the trainer.
- Scale: 3 h per period of apprentice paperwork were identified for a SHK craft business in the shadow analysis; the use case applies to all craft and industrial trainers. Savings figures are an estimate for the writing and checking effort, not measured results; vendor claims are unproven until measured in your own firm.
- Honest caveat: the total time is small per apprentice, so the value lies as much in training quality and exam readiness as in hours saved.
Variants
- Craft business (Handwerk): mobile app, voice or photo entry on the building site, trade-specific framework plan (e.g. SHK, electrical, joiner).
- Industrial / Mittelstand trainer: department rotation plan, several trainers per apprentice, link to the company training plan.
- Training provider / inter-company training centre (ÜBA): records from courses, shared view with the vocational school.
- Vocational school: teachers see the entries and can align lessons (see EDU-009).
- Not a variant: grading the apprentice or rating performance from the reports. That moves into high-risk assessment (see Risks).
How it works
- apprentice input (voice note / short text / photo of the work)
- speech-to-text (ASR) and photo captioning (VLM) in German
- draft entry in the apprentice's own words (LLM, trade vocabulary, no invented tasks)
- APPRENTICE REVIEW & edit (apprentice stays author, confirms the text)
- matching of activities to framework-plan positions (embeddings + LLM, citing the position)
- coverage and gap view per learning objective for the trainer
- TRAINER REVIEW & sign-off (digital confirmation, weekly)
- export in the format the examination board accepts
- Drafting: ASR (e.g. Whisper-class) plus an LLM that only rephrases what was said or shown. The prompt forbids adding tasks that were not mentioned. Photos help recall but must not be the sole source.
- Curriculum check: the framework plan of the trade is chunked and embedded; each activity is matched to the closest positions and the trainer sees the reference. Low-confidence matches are shown as suggestions, not as facts.
- Human in the loop: the apprentice confirms every entry and the trainer approves it. The tool does not decide whether the training goals are met.
Data preparation
Which data: the official Ausbildungsrahmenplan and the school's Rahmenlehrplan for the trade, the company's own training plan (betrieblicher Ausbildungsplan), earlier reports, and the apprentices' daily notes.
Where it lives: framework plans as PDFs on the BIBB or chamber websites, the company training plan in Word or Excel on a file share, existing reports on paper, as Word/PDF or in a digital Berichtsheft app.
Typical gaps: the company training plan does not exist or is a copy of the framework plan; framework plans are PDFs with tables that parse badly; old reports are handwritten; reports are very short and generic; photos show colleagues or customers.
Preparation steps: 1. Collect the current framework plan and company training plan in machine-readable form (see DOC-014, PREP-006). 2. Break the company training plan into learning objectives with identifiers and the planned department or station. 3. Agree rules for photos (no faces, customer premises only with consent, no personal data of customers). 4. Pick 10-20 real, anonymised entries as examples of good style for prompts (PREP-016). 5. Define who may see what: apprentice, trainer, school teacher, chamber (PREP-023).
Tools: the app of the chosen provider, a spreadsheet for the training plan, Paperless-ngx or OCR for old paper reports. Rough share of effort: low, about 10-20 percent; the main work is a usable company training plan.
Implementation paths
Which path for whom
- SMB - A: a ready digital Berichtsheft app with an AI mapping module is cheap or free for apprentices and needs no own IT; the process is standard.
- Developer - B: product builders and chamber/association tools use EU-region ASR and LLM APIs plus the framework plans as retrieval source.
- Public - B: chambers, training centres and vocational schools need integration with their own platforms and governance; APIs with EU regions fit better than self-hosting unless sovereignty rules require it.
Path A - Buy (recommended)
- Digital Berichtsheft apps: Zubido (general app and trade-specific versions; AI maps entries to framework plan positions and shows gaps to the trainer, plus an optional writing aid; free for apprentices, trainers and school teachers, companies pay per tier) [3]. BLok (state-funded online Berichtsheft; free for apprentices and teachers, companies free only if their chamber or association provides it, otherwise one-off plus per-apprentice fees), and trade-specific apps for about 10 trades (e.g. electrical trades, painters, joiners, metalworkers, roofers) listed by the ZDH, some with a target/actual comparison against the training regulation but without AI [4][5].
- Check: where the AI runs and whether data is used for model training, data processing agreement (AVV), export in a format the examination board accepts, signature procedure, whether the chamber or trade association already provides an app.
- Choose A when: you have fewer than a few dozen apprentices and a standard trade.
Path B - Build on cloud APIs
- Pipeline: EU-region speech-to-text, a general LLM with structured output (JSON with entry text, matched positions, confidence), embeddings over the framework plan, storage in your own tenant.
- Integration: Microsoft 365 (Forms/Teams or Power Automate for intake), existing apprentice app via export, HR system for master data.
- Choose B when: you are an association, training centre or software vendor serving many firms, or need a custom plan structure.
Path C - Self-host (open source)
- Stack: faster-whisper or whisper.cpp for German speech, an open 8-30B instruct model via vLLM or Ollama, bge-m3 embeddings with pgvector, n8n for workflow. HW2 is enough for dozens of apprentices.
- Choose C when: data sovereignty is mandatory (public bodies, sensitive customer sites). German workshop jargon and dialect need testing with small ASR models.
Path D - Custom training
Rarely needed. A fine-tuned classifier for matching to one trade's framework plan could pay off only for a large association with thousands of labelled, trainer-approved entries.
State of the art (2026)
- Production, but young: AI matching of entries to the Ausbildungsrahmenplan and AI writing aids exist in at least one commercial German product; many other trade apps still offer only manual assignment or target/actual lists [3][4]. Hence maturity early-adopter, not mainstream. Independent evaluations of accuracy were not found.
- Speech to text for German is good in quiet settings; noisy workshops, dialect and trade terms cause errors.
- Matching works well for broad learning areas and poorly for fine-grained positions; expect trainer corrections.
- Limits and failure modes: the LLM may embellish ("polishing" into tasks never done), apprentices may stop reflecting and just click approve, photos may show third parties, and a shiny coverage dashboard may suggest training quality that the actual work does not deliver. The report is also a reflection and learning tool: letting AI write it fully removes the educational point [4].
Shared resources & combinations
- Same stack (asr + llm-general + RAG): SPEECH-001 meeting notes, SPEECH-005 dictation for field service, KNOW-002 RAG over company documents; one HW2 server can serve all of them.
- Upstream/downstream: EDU-001 and EDU-006 generate learning material for the gaps found; EDU-009 helps vocational school teachers align lessons; EDU-003 gives feedback on written exam tasks; EDU-014 covers recognition of prior learning.
- Documents: DOC-005 transcribes handwritten legacy reports; LANG-003 helps with training certificates and letters; AUTO-008 turns forms or notes into structured rows if no app is used.
- Biggest combined effect: the same voice-note pipeline can feed the Berichtsheft, the job report (Arbeitsbericht) and the time sheet from one spoken entry on the building site.
Effort, cost & KPIs
- Path A: days to weeks; apprentice use often free, firm pays tiered licence per apprentice (check provider prices).
- Path B: 1-2 months for a pilot; API cost per entry is minimal.
- Path C: 1-3 months plus one GPU workstation and operations.
- KPIs: share of weeks with a timely entry, trainer time per review, coverage of learning objectives at the half-way point, share of AI matches corrected by trainers, number of gaps detected early, apprentice satisfaction, exam admission without delay.
People & change
Involve early: the management or training manager (policy, budget), the trainers (Ausbilder) as process owners, apprentices as daily users, and the data protection officer. Apprentices are often minors: involve parents where required and use clear consent-free legal grounds (employment relationship, Art. 6 GDPR) plus transparency (GUIDE-005, GUIDE-010).
Works council: depends. A system that records who writes and approves what and when is suited to monitor behaviour (§87(1) no. 6 BetrVG); §90(1) no. 3 requires timely information on AI-supported work processes. The Jugend- und Auszubildendenvertretung (JAV) should be heard. A short agreement excluding performance rating from report data usually settles it (GUIDE-004).
AI literacy: Art. 4 AI Act applies; give apprentices and trainers a short session on checking drafts, truthful entries and what the tool must not do (GUIDE-006).
Communication and pilot: present it as writing help, not control. Start with one trainer and two to four apprentices for a quarter, compare entry quality and time (GUIDE-008, GUIDE-009). Typical concerns: loss of learning, being monitored by the boss, faked entries. Answer: apprentice edits and confirms, no performance scoring, trainer reviews content.
Risks, regulation & compliance
- Training law: §13 BBiG (duty to keep the record), §43 BBiG (exam admission), the examination board decides the accepted form of the digital record; a written notice to the chamber when switching from paper to digital is advised, not a confirmed legal duty [1][2][5].
- EU AI Act: depends. Drafting reports and showing coverage is limited/minimal risk. If the output is used to evaluate learning outcomes, steer the training path or decide on exam admission, Annex III point 3 (education and vocational training) applies and the use case becomes high-risk. Keep the tool descriptive; do not score apprentices; no emotion inference (Art. 5). Check current dates of the Digital Omnibus (as of mid-2026 the agreed date for Annex III high-risk obligations is 2 December 2027; verify final adoption) [6].
- GDPR: apprentice data (often minors), training progress, photos with third parties and customer sites. AVV with the vendor, EU hosting, no training on inputs, deletion after the training and retention period; DPIA when data is combined with performance evaluation.
- Authenticity: a record that is largely AI-written but signed as the apprentice's own may be a problem; define a rule that content comes from the apprentice's input and is confirmed by them.
- Mitigations: apprentice approval step, trainer sign-off, citations of framework positions, audit log, photo rules, export check with the chamber.
Getting started
- Ask your chamber (IHK/HWK) which format the exam board accepts and whether it already provides a digital Berichtsheft.
- Write or update the company training plan against the framework plan of the trade.
- Test two Path A apps with real entries of two apprentices; check the AI matching on one month of past entries.
- Sign the AVV, clarify photo rules and inform the JAV/works council.
- Review KPIs after a quarter: timeliness, trainer time, gaps found.
Sources
- Gesetze im Internet - Berufsbildungsgesetz §13 (duties of apprentices, written or electronic Ausbildungsnachweis) - https://www.gesetze-im-internet.de/bbig_2005/__13.html
- IHK Gera / Ostthüringen - Ausbildungsnachweis (Berichtsheft), exam admission §43 BBiG, weekly review - https://www.ihk.de/gera/ausbildung-weiterbildung/ausbildung/fuer-auszubildende2/ausbildungsnachweise-2976312
- Handwerksblatt - Berichtsheft von Zubido (AI matching to the Ausbildungsrahmenplan, data protection, costs) - https://www.handwerksblatt.de/betriebsfuehrung/so-wissen-ausbilder-die-das-berichtsheft-von-zubido-nutzen-wo-ihre-auszubildenden
- ZDH - Berichtsheft im Handwerk (apps by trade, Soll-Ist comparison) - https://www.zdh.de/ueber-uns/fachbereich-berufliche-bildung/ausbildung/ausbildungsqualitaet-im-handwerk/berichtsheft-im-handwerk/
- Deutsche Handwerks Zeitung - Digitales Berichtsheft: what trainers and apprentices need to know (exam board rules, signature, BLok) - https://www.deutsche-handwerks-zeitung.de/digitales-berichtsheft-was-ausbilder-und-azubis-wissen-muessen-145933/
- Regulation (EU) 2024/1689 (AI Act), Annex III and Art. 5 - https://eur-lex.europa.eu/eli/reg/2024/1689/oj