Credit transfer, course equivalence & recognition of prior learning

Recognition offices and examination boards compare foreign or earlier coursework with their own curriculum, module by module, in dozens of languages and formats. AI reads the transcripts and module descriptions, aligns learning outcomes, ECTS workload and EQF level with the home modules, retrieves comparable past decisions and drafts a justified proposal. The examiner decides. The gain is mainly turnaround time and consistency; the legal responsibility stays human.

Problem & value

• Today: staff or professors read PDFs of foreign catalogues, translate them ad hoc, compare content by hand and write a decision letter. Quality varies by examiner, precedents live in individual heads or mailboxes, and students wait weeks for answers that affect their study plan, funding or visa.
• Legal frame pushes volume and burden of proof: under the Erasmus Charter, credits achieved abroad as agreed in the learning agreement are to be recognised automatically; declining requires a justification and an alternative form of recognition [1]. The Lisbon Recognition Convention requires recognition of foreign qualifications and study periods unless a substantial difference can be demonstrated [2]. Student organisations report that recognition is still not applied consistently [3]. This puts the burden on the institution to show why something is not equivalent, which is exactly where structured, evidence-based comparison helps.
• AI value: multilingual extraction, outcome-level comparison instead of title matching, precedent retrieval, consistent justifications, queue prioritisation. No reliable public figure on time savings exists; measure your own baseline.
• Distinct from neighbours: KNOW-012 matches people to skills; EDU-015 converts grades for admission. Here the object compared is curriculum against curriculum.

Variants

• Erasmus+ / exchange learning agreements: pre-mobility check of proposed courses, post-mobility conversion of the transcript, increasingly via Erasmus Without Paper (EWP) / Online Learning Agreement.
• Transfer students and change of institution: domestic and international credit evaluation for entry into a later semester.
• Foreign qualifications (Lisbon Convention cases): ENIC-NARIC style comparison of whole programmes.
• Recognition of prior learning (RPL): work experience, vocational qualifications, bootcamps and micro-credentials mapped to modules (exemptions or credit for non-formal learning).

How it works

1. intake (upload / EMREX / EWP / Europass credential / email)
2. OCR + layout parsing + language detection (transcripts, catalogues)
3. normalisation: module record {title, ECTS or local credits, level, grade, outcomes, hours, assessment}
4. credit & grade-scale conversion (ECTS credits, EQF/NQF level, ECTS grading table)
5. candidate retrieval: embeddings + reranker over home modules and past decisions
6. outcome-level comparison by LLM (covered / partially / not covered), workload and level check
7. rule layer: examination regulation limits (max credits, mandatory modules, expiry of old credits)
8. proposal + justification with quoted evidence from both module descriptions
9. examiner review (accept / change / reject; reason codes)
10. decision letter
11. record system

• Extraction: a vision-language model or document parser returns schema-constrained JSON; credits and grades are validated numerically (sum of credits, scale bounds).
• Matching: hybrid retrieval over home module descriptions; an LLM judges coverage per intended learning outcome and must cite the passage on both sides. Scores are evidence summaries, not decisions.
• Rules stay deterministic: regulation limits and mandatory-module logic are coded rules, not LLM judgement.
• Human-in-the-loop: every negative or partial outcome and every low-confidence match goes to the examiner; high-confidence, precedent-backed positives may be fast-tracked, but a named person approves the decision.

Data preparation

The data is module descriptions and handbooks (often PDF, Word or web pages), transcripts and certificates submitted by applicants (scans, foreign languages), earlier recognition decisions, and examination regulations. These live in the campus-management system, a DMS or the registrar's file shares, partly on paper. Typical gaps: outdated or contradictory module descriptions, scanned certificates of mixed quality, no structured credit or learning-outcome fields, decisions stored as free-text letters, and unclear access rights to student files.

Preparation steps: - Inventory sources and owners, see PREP-001 and PREP-005 for exports from campus-management systems. - Scan and OCR paper files, see PREP-006, and parse PDFs with layout-aware tools (DOC-014). - Build a curated catalogue of current module descriptions with metadata (credits, level, outcomes, year), see PREP-010 and PREP-013. - Harvest past recognition decisions as labelled examples and a gold test set with the officers, see PREP-020 and PREP-018. - Pseudonymise student data for development, see PREP-022.

Rough share of effort: 40-50 percent, because the catalogue quality decides retrieval quality. Plan this first; the models are the smaller part.

Implementation paths

Which path for whom

• Public (universities, recognition offices): Path B on EU-region or sovereign German cloud APIs, because the process must plug into the campus-management system and module catalogues, and no product fits national recognition rules.
• Enterprise (private universities, training providers): Path B, because equivalence rules, catalogues and student records are institution-specific and need integration.
• Developer: Path B, because managed APIs plus own retrieval and rules engine are fastest to build; move to Path C only if data residency demands it.

Path A — Buy (off-the-shelf)

• Transfer evaluation platforms: CollegeSource TES (US-centred), Ellucian and Workday Student / Oracle Student modules with transfer-articulation rules; Anthology and Unit4 campus systems.
• EU / DACH campus systems with recognition workflows (EU): CAMPUSonline (TU Graz, AT), HISinOne (HIS eG, DE), Pluto / Moveon (QS), Mobility-Online (EU). Mostly workflow and form handling; AI matching is limited or an add-on.
• Credential infrastructure: EMREX and the Europass Digital Credentials Infrastructure deliver structured transcripts, which removes OCR for participating institutions [4].
• Choose A when: you mainly need workflow, tracking and articulation tables; AI comparison of free-text syllabi is still rarely a mature packaged feature in 2026.

Path B — Build on cloud APIs (recommended)

• Extraction: Azure AI Document Intelligence or Google Document AI (EU regions), Mistral OCR (EU provider), or a multimodal LLM with JSON schema.
• Matching: multilingual embeddings (e.g. Mistral, Cohere, or OpenAI embedding APIs with EU data residency) plus a reranker; a frontier LLM (Claude, Gemini, GPT, Mistral Large) for outcome-level comparison, accessed in an EU region (e.g. Bedrock / Vertex / Azure OpenAI EU).
• Orchestration: the campus system's API or n8n / Camunda; EWP and Online Learning Agreement integration via the institution's existing mobility software.
• Choose B when: you want quality fast, have a data processing agreement with an EU-hosted provider and a small developer team.

Path C — Self-host (open source)

• Parsing: Docling, PaddleOCR, MinerU for transcripts and handbooks.
• Retrieval: BGE-M3 or multilingual-e5 embeddings, bge-reranker, pgvector / Qdrant / OpenSearch.
• Reasoning: open-weight models such as Qwen3, Mistral / Mistral Small, Llama or gpt-oss via vLLM; at least HW3 for 70B-class quality on multilingual outcome comparison, HW2 is enough for a 20-30B model on easy cases.
• When: public universities with strict data-protection requirements for applicant files.

Path D — Custom training

Fine-tune an embedding or reranker model on the institution's own past equivalence decisions (accepted vs. rejected pairs, a few thousand needed) and distil a small classifier for triage. Rarely required at first; precedents are better used as retrieval examples. Beware of baking historic bias (e.g. against certain countries) into the model.

State of the art (2026)

• Multilingual LLMs compare module descriptions well when both sides are text; the weak points are thin or marketing-style syllabi, missing outcome statements, implicit prerequisites and grading cultures that differ from the ECTS scale.
• Retrieval plus evidence-citing justification is the pattern that examiners accept; unexplained similarity scores are not.
• Structured credentials (EMREX, Europass/European Learning Model, Open Badges-style micro-credentials) are growing and make extraction trivial where available [4]; scanned paper transcripts remain common for non-EU applicants.
• Failure modes: hallucinated content assumed from a course title, wrong ECTS equivalence for non-ECTS credit systems, over-crediting similar-sounding modules, systematic disadvantage for institutions with poor English documentation, and silent drift of precedent.

Shared resources & combinations

• Same stack (ocr-docparse + vlm + embeddings + reranker): DOC-003 form processing, DOC-014 PDF parsing, KNOW-002 RAG over documents, EDU-015 admissions screening and foreign grade conversion (shares transcript extraction and grade-scale tables).
• Curriculum knowledge graph: KNOW-013; competence mapping shared with KNOW-012 (ESCO).
• Downstream: EDU-005 adaptive paths and EDU-007 learning analytics (study plans after credit), EDU-001/EDU-009 for gap-closing bridging content.

Effort, cost & KPIs

• Path A: weeks for workflow; AI matching often not included.
• Path B: roughly 3-6 months including curriculum ingestion, regulation rules, examiner UI and evaluation.
• Path C: similar plus GPU operations.
• Cost drivers: curriculum data cleanup, student-record integration, legal review; per-case token cost is small.
• KPIs: turnaround time per case, examiner agreement rate, share of cases with precedent used, false-positive recognition rate, appeals and reversals, student satisfaction, consistency across examiners.

People & change

Start with management and the department lead (head of the recognition office or examination board) to agree that the AI only prepares proposals. Bring in data owners (registrar, faculty module coordinators) early, since their descriptions are the knowledge base, and IT for campus-system integration. The DPO is required: student files are personal data, a DPIA (Art. 35 GDPR) is advisable, an AVV (Art. 28 GDPR) is needed with every API provider, and Art. 22 GDPR demands a human decision on recognition. Works council: depends. If the office staff are employees and the system logs their handling and processing times, §87(1) no. 6 BetrVG applies (BAG: objective suitability suffices) and §90(1) no. 3 requires information; at public universities the Personalrat under state law applies.

Support AI literacy (Art. 4 AI Act asks for support measures, not a guaranteed level) with short training on checking cited evidence. Pilot with one faculty and a few experienced officers, compare against their own decisions, and keep their sign-off mandatory. Typical concerns are job loss and wrong rejections; address them by stressing time saved on research, a visible appeal route for students and full audit logs.

See GUIDE-004, GUIDE-005 and GUIDE-009.

Risks, regulation & compliance

• EU AI Act: Annex III point 3 lists AI systems intended to determine access or admission to education and to evaluate learning outcomes or the appropriate level of education. A system that decides or materially shapes whether credits count, whether a student can progress or enrol is likely high-risk (provider and deployer duties, human oversight, logging, transparency to affected persons, Art. 26/27 for deployers). A purely internal drafting aid whose output is independently assessed may fall under the Art. 6(3) derogation, but it must be documented and is risky where profiling of persons is involved. Treat it as high-risk by default [5]. No Art. 5 prohibition applies to curriculum matching; do not add emotion or biometric features.
• GDPR: transcripts and applicant files are personal data (grades, nationality, sometimes special categories); DPIA recommended; Art. 22: a recognition decision must not be solely automated, give the student a right to human review and explanation; EU hosting and processor agreements; retention limits.
• Administrative law: decisions need reasons and appeal routes (national higher-education laws, Lisbon Convention burden on the institution); keep the examiner's own reasoning, not just the AI text.
• Fairness: monitor outcomes by country of origin and institution type.
• Mitigations: human sign-off, citation of evidence, deterministic rules, audit log of proposal vs. final decision, periodic bias audits, student notice that AI assists.

Getting started

1. Take 100-200 closed cases (with decisions), anonymise them and compute the baseline: turnaround and consistency.
2. Clean the home module catalogue: one record per module with learning outcomes, ECTS, EQF level and prerequisites.
3. Build extraction plus retrieval first (top-5 candidate modules); measure recall against past decisions.
4. Add the outcome-level LLM comparison with mandatory citations; give examiners a review screen with reason codes.
5. Run a shadow pilot for one faculty or the Erasmus office, complete the DPIA and AI Act classification before go-live.

Sources

1. European Commission — Guidelines on the Erasmus+ Learning Agreement for Studies (KA131, automatic recognition) — https://erasmus-plus.ec.europa.eu/resources-and-tools/mobility-and-learning-agreements/learning-agreements/studies-agreement-guidelines-ka131
2. Council of Europe — Lisbon Recognition Convention (CETS 165) — https://www.coe.int/en/web/higher-education-and-research/lisbon-recognition-convention
3. ESN — Recommendations for the revision of the ECTS Users' Guide — https://esn.org/sites/default/files/news/esns_recommendations_for_the_revision_of_the_ects_users_guide.pdf
4. European Commission — Europass / Digital Credentials Infrastructure and EMREX — https://europass.europa.eu/en/stakeholders/european-digital-credentials
5. Regulation (EU) 2024/1689 (AI Act), Annex III point 3 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. Erasmus Without Paper — Onboarding guide (Erasmus+) — https://erasmus-plus.ec.europa.eu/sites/default/files/2024-03/erasmus-without-paper-onboarding-mar24_en.pdf
7. Docling project — https://github.com/docling-project/docling