German companies file a long tail of recurring reports: packaging quantities in LUCID, electrical equipment with stiftung ear, the e-balance sheet with the tax office, tourist-tax returns, cost surveys for the Bundesnetzagentur. Each has its own categories, forms and deadlines, and the numbers sit in the ERP, the shop, the PMS or the ledger. AI maps the internal data to the authority format, pre-fills the values, explains changes against the prior year and tracks deadlines. The filing itself stays with a named person.
Problem & value
- Today: a clerk or the tax advisor exports data, maps it by hand in a spreadsheet, types it into a portal and repeats this next year with half-forgotten rules. The work is small per report, but there are many reports, and mistakes carry real penalties.
- AI value: reuse of last year's mapping, automatic assignment of items and accounts to authority categories, plausibility notes ("packaging weight +40 % because of new product line"), and a deadline list. The gain comes from the mapping step, not from the form entry.
- Who needs it (from the gap analysis): wholesale (WEEE, batteries, packaging, about 5 h per period), online retail (LUCID packaging report, about 4 h), hotels (tourist tax, about 3 h), municipal utilities (Bundesnetzagentur cost and revenue-cap data, about 60 h), tax advisory firms (e-balance sheet about 4 h, filings and certificates about 5 h) [1][2][3][4].
- Stakes: a missing LUCID registration can lead to fines of up to 100,000 EUR, missing system participation up to 200,000 EUR, a faulty data report up to 10,000 EUR, plus sales bans and warnings by competitors [3].
Variants
- Product-compliance registers (retail, wholesale): LUCID/VerpackDG (successor of the VerpackG since 12 August 2026) packaging quantities per material, WEEE (ElektroG) and battery (BattG) registration and quantity reports with stiftung ear [2][3].
- Local-tax returns (hotels): tourist tax (Beherbergungs-/Übernachtungssteuer) per municipality from PMS overnight data, with exemptions for business stays where the municipality's bylaw says so.
- Regulator surveys (municipal utilities): cost data for the revenue-cap review (Erlösobergrenzen) on the Erhebungsbögen set by the authority; large, expert-signed, so AI only drafts [1].
- Tax-advisor workflows: account-to-taxonomy mapping for the e-balance sheet (§ 5b EStG) and bundling of filings and certificates for many clients [4].
- Out of scope: listed-company and bank reporting (ESEF, COREP) is DATA-020; ESG reporting is DATA-012.
How it works
- source data (ERP / shop / PMS / ledger export)
- schema mapping (item or account
- authority category, material, device class, taxonomy position)
- aggregation per reporting period and entity (quantities, weights, amounts)
- plausibility check (vs. prior year, vs. other registers, rule checks)
- AI explanation of changes + list of open questions
- human review and release (named person, four-eyes for large surveys)
- submission via portal / ELSTER / ERiC (by the human or certified software)
- archive of inputs, mapping version and receipt; deadline calendar update
- Mapping: an LLM or classifier proposes the category from item text, attributes and last year's assignment. Confirmed assignments are stored as a rule table, so the next period is largely deterministic.
- Calculation stays deterministic: sums, weights and VAT-relevant amounts are computed in code or SQL, never by the LLM. The LLM drafts the mapping and the explanation text.
- Human-in-the-loop: unknown items, new categories and changes above a threshold go to review. Registrations and quantity reports must be made by the obliged company itself [3].
- Submission: mostly manual portal entry or upload; the e-balance sheet needs XBRL via ELSTER/ERiC compatible software [4]. Agents driving portals (computer use) are possible but fragile.
Data preparation
Which data: item master with weight, material and packaging attributes; sales by period and country; device and battery categories; chart of accounts and trial balance; PMS overnight stays; prior-year reports and the authority's field descriptions.
Where it lives: ERP or shop system (JTL-Wawi, Shopware, SAP Business One, Business Central), DATEV or the ledger, PMS exports, spreadsheets of the compliance clerk, PDFs of authority notices.
Typical gaps: items without packaging weight or material, no flag for electrical devices or batteries, grown account structures with collective accounts, quantities that differ between ERP and shop, and mapping rules that exist only in one person's spreadsheet.
Preparation steps: 1. Add the missing attributes (weight, material, device category) to the item master, starting with top sellers. 2. Document last year's mapping and the authority's category list as the starting rule table (PREP-016). 3. Export clean period data via the ERP interface (PREP-005) and reconcile with the ledger. 4. Build a test set from two past filings with known correct values (PREP-018). 5. Restrict access to the exports and filings (PREP-023).
Tools: ERP exports, spreadsheet or SQL for aggregation, DATEV Kontennachweis for the e-balance sheet.
Rough share of effort: high, roughly half of the project, mainly attribute maintenance and mapping. See GUIDE-003.
Implementation paths
Which path for whom
- SMB — Path A: most obligations are small; compliance services, shop plug-ins and the accounting software cover them, and AI features come with the product.
- Enterprise — Path B: many entities, countries and registers justify an own pipeline on cloud APIs next to the ERP, with own rule tables and approval flow.
Path A — Buy (recommended for most SMBs)
- Compliance services and plug-ins: providers of packaging and WEEE full service (for example Lizenzero/Interzero, Reclay, Landbell, and others) manage registration and quantity reports; shop and ERP vendors offer packaging-report exports. Check EU hosting, who submits, and how the data is exported.
- E-balance sheet: DATEV, Lexware, onlinebilanz-type tools and tax-advisor software generate the XBRL and the account mapping; AI helps with mapping suggestions and consistency checks [4].
- Choose A when: reports are standard, volumes small, and the provider can read your ERP export.
Path B — Build on cloud APIs
- Pipeline: SQL aggregation from the ERP, an LLM API (EU region or EU provider) for category mapping and change explanations with structured outputs, n8n or Power Automate for deadlines and approvals.
- Choose B when: several registers and entities share the same item and account data, or no product covers the combination.
Path C — Self-host (open source)
- Stack: open-weight models (for example Qwen- or Mistral-class, quantised) via vLLM or Ollama, n8n, PostgreSQL for rule tables. HW2 is enough: the volume is small.
- Choose C when: item and sales data must not leave the company, which is rarely the main concern here.
Path D — Custom training
Rarely worth it. A small classifier for item-to-category mapping only pays off with tens of thousands of items; corrections from the review step are the training data.
State of the art (2026)
- Mapping and explaining are well within current LLM ability when the category lists are provided as context; the weak point is incomplete master data, not the model.
- Authority interfaces change slowly and are not AI-ready: LUCID and stiftung ear are portals, the BNetzA uses formal Erhebungsbögen as spreadsheets for the revenue-cap review [1], and the e-balance sheet is XBRL with a fixed taxonomy (6.9 for financial years starting after 31 December 2025, 6.10 announced for 2027) [4]. Anything beyond file-based submission means browser automation.
- Honest maturity: the deterministic part (export, aggregate, submit) is mature and sold as services. The AI layer (mapping, plausibility, explanation) works for small register reports and is productised only in parts. Large surveys such as the BNetzA cost review need expert sign-off, so AI helps only with drafting and checks.
- Limits: LLMs invent categories or weights if data is missing; they do not know current local bylaws or revised thresholds unless given the current text. Quantities must always come from computed data.
Shared resources & combinations
- Same data foundation: DATA-003 (master-data cleaning) and DATA-010 (product data enrichment with classification) deliver the item attributes this use case lives on; DATA-022 covers partner and EDI mapping.
- Same pattern: DATA-020 (financial and supervisory reporting) and DATA-012 (ESG/CSRD) reuse mapping, plausibility and audit-trail components; one rule-table and approval service can serve all three.
- Upstream and downstream: DOC-001 and DOC-017 supply invoice and bank data; AGENT-006 (reconciliation and close) provides reconciled ledger figures; DATA-002 produces the narrative for change explanations; DATA-024 (variance analysis) reuses the year-over-year logic.
- Large combined effect: the same item master with weight, material and device class feeds packaging, WEEE and battery reports at once, so one cleanup removes several recurring jobs.
- Deadlines: a deadline tracker for statutory inspections and certificates is a natural neighbour, see AGENT-004 for the general back-office agent pattern.
Effort, cost & KPIs
- Path A: days to weeks plus provider fees; most effort goes into attribute clean-up.
- Path B: 1–3 months for a pipeline with ERP export, rule tables and approval flow.
- Path C: similar to B plus one small server and operations.
- Cost drivers: master-data maintenance, number of registers and entities, expert sign-off for large surveys.
- KPIs: hours per filing, share of items mapped automatically, number of corrections after review, deadlines missed (target zero), authority queries or rejections.
People & change
Involve early: the head of finance or compliance as process owner, the owner of the item and ERP data, IT for exports, and the DPO if guest or customer data is used (tourist tax). Management decides who is legally responsible for each filing.
Works council: no for typical use. The tool processes items, quantities and accounts, not employee behaviour. If it logs individual clerks' work in detail, §87(1) no. 6 BetrVG may apply; §90 BetrVG still requires timely information (GUIDE-004).
AI literacy: Art. 4 AI Act: brief training on checking proposed mappings and on what the LLM does not calculate (GUIDE-006).
Communication and pilot: start with one register that is painful today (often packaging or WEEE), run the new process in parallel with the old one for one period, and name one owner (GUIDE-008, GUIDE-009).
Humans in control: the obliged person releases every filing; no automated submission of registrations or quantity reports without review.
Risks, regulation & compliance
- EU AI Act: minimal risk; no Annex III category. AI literacy duty (Art. 4) applies.
- GDPR: mostly no personal data; tourist-tax data and sole-trader accounts can contain it. Data processing agreement with the vendor (Art. 28) and EU hosting where personal data is involved.
- Sector law: VerpackDG, in force since 12 August 2026 and replacing the VerpackG (registration up to 100,000 EUR, system participation up to 200,000 EUR, data report up to 10,000 EUR) [3], ElektroG and BattG (registration with stiftung ear before placing on the market) [2], § 5b EStG for the e-balance sheet [4], EnWG/ARegV for revenue caps [1]. Local bylaws apply for tourist tax.
- Liability: the company stays responsible for correctness; wrong categories can mean fines or lost registrations. Keep the mapping version and source data per filing (GoBD-style traceability).
- Mitigations: deterministic calculation, prior-year comparison, review thresholds, audit log.
Getting started
- List all recurring filings with owner, authority, deadline, data source and hours per year.
- Pick the one with the best ratio of effort to risk and check its master-data gaps.
- Build the rule table from last year's filing and let the AI propose only the unknown items.
- Run one period in parallel, compare the numbers and log every correction.
- Add the next register on the same item master.
Sources
- Bundesnetzagentur — Kostenprüfung (Erlösobergrenzen, Beschlusskammer 8), formal Erhebungsbögen — https://bundesnetzagentur.de/DE/Beschlusskammern/BK08/BK8_05_EOG/52_Kostenpruefung/BK8_Kostenpruefung.html
- stiftung ear — who we are (ElektroG and BattG registration tasks) — https://www.stiftung-ear.de/en/about-us/who-we-are
- Lizenzero — Zentrale Stelle Verpackungsregister and LUCID (registration, quantity reports, fines) — https://www.lizenzero.de/verpackungsregister-lucid/
- Steuer-Schröder — E-Bilanz 2026 (§ 5b EStG, taxonomy 6.9/6.10, ELSTER/ERiC, account mapping) — https://www.steuerschroeder.de/e-bilanz.html
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj