Most people with a valid claim never enforce it: the flight was delayed, the rent is above the cap, the service-charge statement is wrong, the warranty was refused. Checking eligibility, finding the right paragraph, writing the letter and chasing the deadline is tedious. An assistant that combines deterministic rule checks (distance bands, rent-cap formulas, limitation periods) with LLM drafting turns a vague complaint into a sourced claim letter and a tracked case. In Germany the line between helping and "legal services" is regulated by the RDG, which shapes every product in this space.
Problem & value
- Today: consumers search forums, use template letters, or hand the case to a commission-based provider (typically paid as a share of the recovered amount). Many never claim; deadlines pass; letters are too vague to trigger the statutory consequence.
- AI value: (1) plain-language triage ("do I have a claim?"), (2) extraction of facts from documents, (3) a correct, cited letter in the right register and language, (4) deadline tracking and escalation. Rules stay in code or curated knowledge bases; the LLM explains and drafts.
- Legal anchors that make rule checks feasible: Regulation (EC) 261/2004 (fixed compensation bands by distance, extraordinary-circumstances defence), the German rent-cap rules in BGB §§556d ff., the statutory warranty rules (Directive (EU) 2019/771, in Germany BGB §§433 ff., 475 ff.), and the Directive (EU) 2020/1828 on representative actions [1][2][3][4].
- No reliable independent figure for "claims not enforced" is used here; measure your own baseline instead.
Variants
- Air passenger rights (EU261/UK261): flight, delay, cancellation, denied boarding; best-defined rules.
- Tenancy: rent-cap check (comparative local rent, exemptions such as new builds), service-charge statement audit, deposit return, rent-reduction notice for defects.
- Utilities, telecoms, subscriptions: price-increase objections, cancellation letters, billing errors (overlap with AGENT-019).
- Goods & services: warranty/conformity claims, withdrawal (14 days), package-travel defects.
- Cross-border: claim in another member state via the European Consumer Centres and EU ADR/ODR channels [5].
- For professionals: law-firm intake tools and legal-tech providers use the same pipeline under their own licence (see DOC-010, KNOW-006). Out of scope here.
How it works
- intake (chat + uploads: contract, bill, ticket, emails)
- document parsing & fact extraction (dates, amounts, parties, flight no., rent, area)
- issue classification (claim type, jurisdiction, counterparty, limitation period)
- rule engine / calculators (EU261 band, rent-cap ceiling, deadline, cost-benefit)
- retrieval over curated law, case-law summaries and authority guidance (RAG, with citations)
- LLM reasoning: merits, counter-arguments, missing evidence, questions to the user
- drafting (letter/email/form, user language + counterparty language)
- HUMAN REVIEW & SEND (user confirms; optional lawyer/legal-tech hand-off)
- tracker (reminders, reply parsing, escalation: ADR body, order for payment, small claims)
- Rule engine first: distance from airport coordinates, delay at arrival, rent ceiling as comparative rent plus 10%, limitation periods are computed in code and shown to the user, not "recalled" by the LLM.
- Retrieval: law texts (EUR-Lex, gesetze-im-internet.de), court-decision databases, ministry and consumer-centre guidance; every statement in the output carries a citation, otherwise it is marked uncertain.
- Human-in-the-loop: the user always approves and sends; for high-value or contested cases the assistant recommends a lawyer, legal-aid advice (Beratungshilfe) or a consumer centre.
- Agentic filing: web forms of airlines or ADR bodies can be filled by a computer-use agent, but submission needs explicit user consent.
Data preparation
The data are the dispute documents: contracts, terms and conditions, invoices, order confirmations, e-mail and chat correspondence, screenshots, tickets, photos of defects and deadlines. They sit in mailboxes, phone galleries, paper folders and sometimes customer portals. Typical gaps are missing dates, correspondence spread across channels, screenshots without context, no proof of delivery or of a complaint, and unclear versions of terms and conditions.
Preparation steps: - Build one case folder with a timeline (what happened when, with proof) before asking the assistant anything. - Export mail threads to PDF (PREP-007), scan paper (PREP-006) and name files by date and sender. - Redact data that is not needed, such as bank details of third parties. - For a developer or SMB tool: curate a small authoritative knowledge base of statutes, court decisions, templates and authority addresses (PREP-013), and maintain it as laws change (PREP-017). - Collect gold-standard letters that a lawyer approved (PREP-016).
Effort is low: the timeline and evidence collection matter much more than technology and typically take over half of the time. For a small knowledge base with RAG the curation of sources is the main work. See GUIDE-003.
Implementation paths
Which path for whom
- Private: A. Licensed claim services carry the cost risk for standard cases (flights, rent), and consumer centres and ADR bodies are free or cheap.
- Developer: B. A drafting and deadline tool on EU-region APIs with curated legal templates is feasible and keeps the build small.
- SMB: A. A small business with a standard consumer or contract dispute is better served by a lawyer or licensed provider supported by a general assistant than by a custom build; the use case is occasional, so volume does not justify more.
Path A — Buy (recommended for most consumers)
- Claim-enforcement services (commission-based, licensed): Flightright, AirHelp, EUclaim for flights (Flightright and several others are German or EU-based); wenigermiete.de (Conny) and similar for rent-cap and service-charge claims; consumer-law chat tools of consumer centres.
- Free official routes: national enforcement bodies (Germany: Luftfahrt-Bundesamt), the ADR body Schlichtungsstelle für den öffentlichen Personenverkehr (söp), and European Consumer Centres [5][6].
- General assistants: Claude, ChatGPT, Gemini or Mistral Le Chat for drafting letters, with the caveats below. Tenants' associations and consumer-advice services remain strong human options.
- Choose A when: the case type is standard and a licensed provider will carry the cost risk.
Path B — Build on cloud APIs
- Frontier or EU models (Mistral, Claude, Gemini, GPT via EU regions) with structured outputs for fact extraction, plus a deterministic rules layer (Python/TypeScript) for calculators and deadlines.
- OCR/VLM for contracts and statements (Azure Document Intelligence, Mistral OCR, Google Document AI).
- RAG over statutes and decisions with a vector DB (pgvector, Qdrant) and a reranker; open legal data from EUR-Lex and the Federal legal portals; mandatory citation checks before display.
- Workflow and reminders via n8n, calendar and email APIs. Choose B when building a product or a consumer-association service, which needs its own RDG analysis.
Path C — Self-host (open source)
- Private assistant at home (PERS-005): open-weight models such as Qwen3, Llama, Mistral Small/Medium or gpt-oss via Ollama/vLLM; Docling or PaddleOCR for documents; a local vector DB with the user's own contract folder.
- Suited to users who do not want contracts, rental agreements and medical or bank data to leave the device. Hardware: HW2 (24 GB GPU) or a recent laptop for ~8B–14B models with modest quality on legal reasoning; larger models need HW3.
Path D — Custom training
Rarely sensible. Fine-tuning on German legal text and on claim-letter pairs can improve tone and terminology, but retrieval plus rules beat parametric knowledge for freshness. Needs labelled cases with outcomes, which only operators of claim services hold.
State of the art (2026)
- Reasoning LLMs can read a lease clause or a service-charge statement and flag inconsistencies, draft letters in formal register and translate them for cross-border claims.
- Strongest pattern: LLM as explainer and drafter, deterministic code for numbers, retrieval for law. Pure LLM answers still produce wrong paragraph numbers, outdated case law and invented decisions.
- Failure modes: hallucinated citations, ignoring recent court rulings or changed thresholds, jurisdiction mix-ups (national rent-cap rules differ per federal state and city), overconfident "you will win" statements, missing deadline awareness.
- Evaluation: test with a gold set of real, anonymised cases per claim type; track citation correctness and outcome agreement with a lawyer.
- Courts and regulators are increasingly fast at digital filing, which favours assistants that produce complete, structured submissions.
Shared resources & combinations
- Same stack (llm-general + ocr-docparse + RAG): PERS-008 (understanding letters from authorities), DOC-009/DOC-010 (contract analysis and drafting, professional), KNOW-006 (legal research).
- Upstream/downstream: PERS-003 (trip documents feed flight claims), PERS-001 (bank statements as evidence for charges), AGENT-019 (subscriptions, bills, cancellations), PERS-016 (GDPR rights requests use the same letter-and-tracker engine), PRED-038 (outcome prediction, professional).
- One private local server (PERS-005) can host all of these.
Effort, cost & KPIs
- Path A: minutes for the user; commission-based providers typically take a percentage of any recovery, so check fees and contract terms.
- Path B: 2–4 months for an MVP with rules for one or two claim types; legal review is the main cost.
- Path C: a weekend to set up a private assistant; ongoing curation of law sources is the effort.
- KPIs: share of claims correctly classified, citation accuracy, time from intake to sent letter, recovery rate, deadlines missed, user-reported understanding, escalation correctness.
People & change
For a private user the people are the opponent's customer service, a consumer centre or tenants' association, and sometimes a lawyer. Involve a lawyer or advice centre before any court step, for high amounts or unclear law; the assistant does not give binding legal advice.
In an SMB, the owner decides on use and budget, the staff member handling the dispute is the end user, and legal/compliance (often an external lawyer) reviews letters before sending. Involve the data protection officer if customer data is pasted into tools: use a provider with an AVV (Art. 28 GDPR) and no training on inputs (GUIDE-005). A DPIA is normally not needed for occasional drafting. Add the tool to the company AI policy so staff know what may be pasted (GUIDE-007).
Works council: no, as long as the assistant is used occasionally for external disputes and no employee behaviour is evaluated. If it becomes a logged standard tool for staff, §87(1) no. 6 BetrVG may apply, and §90(1) no. 3 requires early information when work processes change (GUIDE-004).
AI literacy: staff should understand that the assistant can cite wrong statutes or invent decisions, so every reference is verified and deadlines are checked manually (GUIDE-006). Start with one experienced person as pilot, keep the human signature on every letter, and communicate that the tool prepares drafts rather than decides (GUIDE-008).
Risks, regulation & compliance
- EU AI Act: minimal risk (Reg. (EU) 2024/1689). Systems used by or on behalf of judicial authorities to research and apply the law are high-risk under Annex III; a consumer-side assistant is not. Art. 50 transparency applies to chatbots (users must know they talk to AI); Art. 4 AI literacy for providers and deployers. No Art. 5 prohibition is touched [7].
- Legal services law (Germany): the RDG reserves individual legal advice in third-party matters to lawyers and registered persons. A generic tool giving individual legal assessments to the public, or a provider collecting claims commercially, needs a licence, e.g. as a registered collection service, or a lawyer behind it; the 2021 legal-tech reform widened what collection services may do, but case-law boundaries (unauthorised legal advice, conflicts of interest, fee rules) continue to evolve [8]. Self-help for one's own matters is allowed; position products as information and drafting aids with a clear scope, and get RDG advice before commercial launch. Other member states have their own rules.
- GDPR: contracts and correspondence contain personal data of the user and third parties (landlord, co-tenants). Art. 6(1)(b) or consent; EU hosting, processor agreements, retention limits, no training on user data. Special-category data (health, union membership) can appear in disputes.
- Liability and consumer protection: disclaimers do not replace correct output; unclear warnings, dark patterns around success fees and unfair terms are enforceable.
- Mitigations: rules layer, mandatory citations, "last verified" date on legal sources, escalation triggers (high value, court deadline, criminal angle), no automatic sending.
Getting started
- Pick one claim type with crisp rules (EU261 flights) and build the calculator plus checklist first.
- Index primary sources only (EUR-Lex, gesetze-im-internet.de, ministry and consumer-centre pages).
- Draft templates by lawyers, with LLM filling facts and adapting tone; keep an audit log.
- Test on 30–50 anonymised past cases with expected results; measure citation accuracy.
- Add deadline tracking and an escalation page (ADR body, consumer centre, lawyer, legal aid).
Sources
- Regulation (EC) No 261/2004 (air passenger rights) — https://eur-lex.europa.eu/eli/reg/2004/261/oj
- Directive (EU) 2019/771 on the sale of goods — https://eur-lex.europa.eu/eli/dir/2019/771/oj
- BGB §556d, Mietpreisbremse (gesetze-im-internet.de) — https://www.gesetze-im-internet.de/bgb/__556d.html
- Directive (EU) 2020/1828 on representative actions — https://eur-lex.europa.eu/eli/dir/2020/1828/oj
- European Consumer Centres Network — https://commission.europa.eu/live-work-travel-eu/consumer-rights-and-complaints/resolve-your-consumer-complaint_en
- söp Schlichtungsstelle öffentlicher Personenverkehr — https://soep-online.de
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Rechtsdienstleistungsgesetz (RDG) — https://www.gesetze-im-internet.de/rdg/
- Flightright — https://www.flightright.com
- wenigermiete.de — https://www.wenigermiete.de