Care-grade and assistive-aid application and objection assistant

Outpatient care services (Pflegedienste) and medical practices spend hours on paperwork that decides whether clients get the right care grade and the right aids: applications, evidence, reminders when an insurer stays silent, and objections against rejections. The material needed is already in the care record. An LLM can turn it into a structured draft application and a deadline tracker, while the care professional checks and signs. This is a small, well-bounded back-office use case, not a product category of its own yet.

Problem & value

• Today: a client or relative applies for a Pflegegrad at the Pflegekasse, or an upgrade after the condition worsens. The care service often helps informally: it collects diagnoses, describes daily limitations, prepares the client for the MD (Medizinischer Dienst) assessment and writes objections when the grade is lower than expected. Aid requests (bed, walker, shower chair, consumables, home modification grant) go to the health or care insurer, often with a doctor's order. Deadlines are tracked in Excel or in someone's head; letters from insurers arrive by post.
• AI value: draft text and evidence sheets from data already documented, a checklist of missing documents before submission, and a calendar of deadlines that catches a lapsed objection period or a deemed approval. In the shadow-process analysis of the catalogue this is about 6 h per period for a Pflegedienst; treat it as an estimate, not a measurement.
• Why the deadlines matter (verified in law texts): the Pflegekasse must decide within 25 working days of the application; for missed deadlines, compensation of 70 euros per begun week applies unless the insurer is not responsible for the delay; since 1 Jan 2026 pauses of the deadline can run up to 15 further working days. The objection period is one month after receipt of the notice [1]. For aids, the health insurer decides within 3 weeks (5 weeks with an MD opinion) and the benefit is deemed approved if it neither decides nor gives a sufficient reason in time [2]; for Pflegehilfsmittel similar 3/5-week periods apply [3]. Check current wording before relying on any figure.
• No invented figures: no vendor-independent study on time saved exists; the savings range in the front matter is a planning assumption from the catalogue's process analysis.

Variants

• Pflegedienst (ambulant): main case. Care-grade application support, aid requests, objections, and the evidence trail from the care record.
• Medical practice (Arztpraxis): aid orders (Hilfsmittelverordnung), rehabilitation applications (Muster 61), medical statements for the MD. These are currently mapped weakly to HLTH-011.
• Nursing home (stationär): re-grading when residents deteriorate, because the grade drives the insurer payment; same pipeline, more clients.
• Care counselling / social service / hospital discharge management: same drafts, plus handover.
• Private use: relatives drafting an application or objection with a general chatbot; the same logic applies, but there is no care record and no professional check.

How it works

1. trigger (new client | condition change | insurer letter | deadline close)
2. collect sources (care plan, SIS, assessments, doctor's order, previous Bescheid / MD report)
3. map facts to the five assessment modules (mobility, cognition & behaviour, self-care, illness-related demands, daily life & social contacts) and to aid needs
4. draft application + evidence sheet (only facts present in the record, each with source reference)
5. completeness check (missing diagnoses, signatures, doctor's order, insurer form)
6. human review and signature by care professional / client or legal representative
7. send (post, insurer portal, e-mail) + register deadlines in the tracker
8. monitor: insurer letter scanned
9. classified
10. deadlines updated
11. on rejection or low grade: compare MD report with record, draft objection, human decides

• Extraction and classification: a document model reads scanned insurer letters and extracts type (approval, rejection, request for documents), date of receipt and amounts; a classifier or LLM with a schema does the rest.
• Drafting: the LLM may only use facts from the record. It must mark gaps ("not documented") instead of filling them. The mapping to modules is a suggestion; it does not assess or score the client. The Pflegegrad itself results from the MD assessment with weighted modules (mobility 10 %, cognition/behaviour 15 %, self-care 40 %, illness-related demands 20 %, daily life 15 %) and fixed point thresholds [4].
• Deadline logic: the date rules (25 working days, one month for objection, 3/5 weeks for aids) are coded as plain rules and not left to the LLM. Receipt date is entered by a human.
• Human in the loop: every outgoing document is released by a named person. Objection letters on legal merits go to the client, a Pflegeberatung or a lawyer when the case is complex.

Data preparation

Which data: the care plan, SIS/risk assessments and daily entries; diagnoses and doctor's orders; previous grade decisions and MD reports; insurer master data and forms; deadline history.

Where it lives: the care software (Medifox Dan, Connext Vivendi, CGM Sozial and similar), the PVS in a practice, a mix of Word/Excel templates, e-mail inboxes and paper folders. Insurer letters arrive as paper or PDFs, usually unstructured.

Typical gaps: free-text entries that never mention the limitation that matters for the grade; missing consent or power-of-attorney documents; letters filed under the wrong client; deadlines tracked only in individual calendars; no template library of successful applications and objections.

Preparation steps: 1. Digitise and file insurer letters per client in one place (PREP-006, PREP-008). 2. Agree on a small glossary and the mapping of care-plan fields to assessment modules (PREP-015). 3. Collect 20-30 past applications and objections with outcomes as examples and gold standard (PREP-016, PREP-020), pseudonymised for tests (PREP-022). 4. Define a deadline sheet with receipt date, type, due date, owner. 5. Check the legal basis and the access rights per client record (PREP-025, PREP-023).

Rough share of effort: about 30 percent; mostly document intake and template cleanup, not model work.

Implementation paths

Which path for whom

• SMB — Path B: no mature product covers the combination of care record, drafting and deadline tracking; a small build on an EU-hosted LLM API plus a workflow tool is realistic, ideally by an IT partner of the care-software vendor. Where the care software or an office suite already has an assistant, Path A may start the pilot.
• Developer — Path C: a vendor of care or practice software can build this on open-weight models and keep health data on-premises.

Path A — Buy

• Care and practice software: vendors add AI assistants and letter modules (check Medifox Dan, Connext Vivendi, CGM roadmaps); today these focus on documentation (see HLTH-025), not on insurer applications.
• General assistants in office suites (Microsoft 365 Copilot, Gemini for Workspace) for drafting from attached documents; fine as a start but they have no deadline logic and no link to the care record.
• Choose A when: volume is small (under ~20 applications a month) and a human prepares the data.

Path B — Build on cloud APIs

• Models: EU-hosted LLMs (e.g. Mistral) or frontier models via EU regions with structured outputs.
• Document intake: a document-parsing service for scanned insurer letters (see DOC-014, DOC-008).
• Orchestration: n8n or Power Automate for intake, tracker (Excel/SharePoint/Nextcloud) and reminders.
• Integration: most care software has no open API; expect export of care-plan and master data as files or partner interfaces.
• Choose B when: you have a standard set of templates and 10-50 users.

Path C — Self-host (open source)

• Open-weight LLMs of 8-30B (Mistral Small, Qwen, Llama-class) via Ollama or vLLM on HW2, OCR/layout with Docling or PaddleOCR, workflow in self-hosted n8n.
• Right for care software vendors, large providers and church providers with strict data rules.

Path D — Custom training

Not needed. Prompting with examples and rules suffices; fine-tuning would only help for a vendor with thousands of anonymised applications and outcomes.

State of the art (2026)

• LLMs draft German formal letters and summarise records reliably; with a source reference per fact, hallucination is controllable. Extraction from standard insurer letters works well with document AI.
• Maturity is early-adopter at best: the pieces are mature, but we found no dedicated product for German Pflegedienste that covers application drafting, evidence mapping and deadline tracking. Teams assemble it from generic components.
• Limits: an LLM cannot judge what the MD will find. It may overstate needs, which is a risk of misleading the insurer; it cannot replace the assessment. Insurer rules and forms change (for example the 2026 change in the pause of the deadline [1]); legal advice on objections is for lawyers or counselling services, so the tool must draft, not decide.
• Rights and fees: the Pflegedienst offering legal help must respect the Rechtsdienstleistungsgesetz; as a rule only ancillary help to the care service is allowed. Check with a lawyer before offering objection drafting as a paid service.

Shared resources & combinations

• Best combined with HLTH-025 (voice nursing documentation): better and more timely entries are the evidence base for every application and objection; HLTH-025 raises quality of the input.
• Same stack (llm-general + ocr-docparse + workflow-automation): DOC-003 form processing, DOC-008 mailroom classification, DOC-014 PDF parsing, AGENT-002 back-office automation; one parsing service and one workflow engine can serve all of them.
• Neighbouring cases: HLTH-011 (reports and letters; covers practice rehab applications), HLTH-018 (record summarisation as input), HLTH-030 (reply drafting), HLTH-005 (billing).
• Biggest effect with least effort: if only two things are implemented, take the deadline tracker with insurer-letter intake and the evidence sheet. They protect revenue and clients' rights and do not need integration into the care software.

Effort, cost & KPIs

• Path A/B: a few weeks for a pilot with templates and deadline rules; running cost is mainly API usage and a workflow tool. Path C: 2-4 months incl. evaluation and a GPU server.
• KPIs: hours per application, share of applications complete at first submission, lapsed deadlines (target zero), number of objections filed in time, approval rate before/after (only as an observation, not a target to steer the assessment), correction rate on drafts.

People & change

Involve early: management and the nursing director as process owner; the people who handle insurer paperwork (often back-office or head nurse); IT; the DPO before any client data is processed (health data, Art. 9 GDPR; DPIA likely; processing agreement with every vendor; GUIDE-005). Church providers apply KDG or DSG-EKD in place of GDPR.

Works council: usually no. The tool processes client data and documents, not employee behaviour, so §87(1) no. 6 BetrVG is typically not triggered; still inform the works council (§90(1) no. 3 BetrVG) if one exists, and check if per-user logs could monitor staff (GUIDE-004).

AI literacy: Art. 4 AI Act requires measures for staff competence; teach the check of draft against the record and the rule that no unsupported statement is sent (GUIDE-006).

Communication and pilot: start with one team and 5-10 applications in parallel to the old way; name a champion; clients or relatives sign the final document; explain that the tool saves writing time but does not change the assessment (GUIDE-008, GUIDE-009).

Typical concerns: fear of wrong statements to insurers (every fact must have a source), fear that paperwork becomes "templated" (keep the personal description), uncertainty about legal advice.

Risks, regulation & compliance

• EU AI Act: minimal risk for drafting and deadline tracking; AI literacy applies. If the tool begins to score the client or to predict the Pflegegrad, classify again, because health-related assessments may become medical-device territory or an Annex III case for access to benefits; keep the tool out of the assessment.
• GDPR: health data of clients, special category (Art. 9); a legal basis under the care or treatment contract, consent of the client or legal representative for transfers to the insurer, EU hosting, deletion rules, no training on client data. Professional secrecy (§203 StGB) applies when external service providers see client data.
• Sector rules: SGB XI (§§15, 18, 40) for the assessment, deadlines and aids, SGB V (§13(3a), §33) for health-insurance aids, SGG for objections and lawsuits, Hilfsmittelverzeichnis of the GKV-Spitzenverband for product listings [5].
• Risks and mitigations: overstating needs (source reference, human check), wrong deadline (rule engine, receipt date entered by a human, weekly list), wrong client (client selected, not guessed), insurer letter in the wrong file (intake review).

Getting started

1. Measure the baseline: applications, aid requests and objections per month, hours, lapsed deadlines.
2. Build the deadline sheet first, even without AI, and enter all open insurer procedures.
3. Pick 10 closed cases, run a draft with a model on the pseudonymised record and compare with what was sent.
4. Agree DPIA, processing agreement and release rule; then pilot with one team for 4-8 weeks.
5. Extend to practice rehab and aid orders if the pilot holds.

Sources

1. betanet — Pflegeantrag und Pflegebegutachtung (deadlines, compensation, 2026 pause, objection) — https://www.betanet.de/pflegeantrag-und-pflegebegutachtung.html; primary law: SGB XI § 18c (25 working days, 70 euro, pauses) — https://www.gesetze-im-internet.de/sgb_11/__18c.html
2. SGB V § 13 — Kostenerstattung, Absatz 3a (decision periods, Genehmigungsfiktion) — https://www.gesetze-im-internet.de/sgb_5/__13.html
3. SGB XI § 40 — Pflegehilfsmittel und wohnumfeldverbessernde Maßnahmen — https://www.gesetze-im-internet.de/sgb_11/__40.html
4. SGB XI § 15 — Ermittlung des Grades der Pflegebedürftigkeit (modules, thresholds) — https://www.gesetze-im-internet.de/sgb_11/__15.html
5. GKV-Spitzenverband — Antragsverfahren Hilfsmittelverzeichnis / Pflegehilfsmittelverzeichnis — https://www.gkv-spitzenverband.de/krankenversicherung/hilfsmittel/hilfsmittelverzeichnis/antragsverfahren/antragsverfahren.jsp