A Stadtwerk or grid operator builds cables, pipes, stations and house connections all year. Every project collects costs on an Anlage im Bau (AiB) and, after commissioning, has to be settled into the right final assets with the right asset class, useful life and split between capitalisable cost and expense. This is slow detective work between project controlling, the GIS and the asset ledger. An agent can match projects, postings and GIS objects and prepare the settlement proposal. The accounting judgement stays with the asset accountant. Today this is realistic as a matching and proposal assistant, not as an autonomous booking agent.
Problem & value
- Today: the asset accounting of a Stadtwerk (Anlagenbuchhaltung) closes construction projects by hand. Costs sit on investment measures (internal orders or WBS elements) as Anlagen im Bau. When the site is commissioned, someone has to find out which assets were actually built (from the GIS and the as-built documentation), split the costs across them, separate maintenance expense from capitalisable cost, treat grants and Baukostenzuschuesse correctly and pick the useful life. The gap analysis behind this entry estimates about 25 h per month for grid asset accounting in a typical Stadtwerk (internal estimate, not a measured benchmark).
- Why it matters beyond bookkeeping: for grid operators the asset ledger feeds the regulated network costs. Calculatory depreciation is determined per asset, linear, based on the useful lives in Anlage 1 StromNEV, and the useful life stays fixed for the rest of the depreciation period [3][4]. A wrong class or commissioning year is therefore carried for decades and shows up in the cost audit.
- AI value: less searching across systems, consistent application of the in-house capitalisation guideline, earlier detection of projects that are commissioned but not yet settled (stale AiB), and a documented reasoning for every split.
- Evidence: we found no independent benchmark or published production case of AI in grid asset capitalisation. Savings figures are estimates; measure your own baseline first.
Variants
- Electricity and gas grid operator: cables, lines, transformers, stations, house connections, pipes.
- Water, district heating and fibre: same logic with other asset classes and funding rules (see OPT-036 for rollout planning).
- Municipal utilities with a city budget: additional rules from municipal accounting; asset ledger may sit in the city system.
- Construction companies building for customers: the revenue side (long-term contracts) is a different problem; here it is the owner's asset side.
- Service provider model: a shared IT or accounting service centre runs the process for several Stadtwerke.
How it works
- closed or commissioned investment measure (AiB with cost postings)
- collect evidence (postings, material withdrawals, invoices, own labour, order text)
- find built objects (GIS objects, as-built documents, commissioning date)
- classify each posting (capitalisable cost? maintenance expense? grant or contribution?)
- map to final assets (asset class, useful life, split key, commissioning year)
- check rules (capitalisation guideline, HGB/tax, regulatory classes, plausibility checks)
- draft settlement proposal (Abrechnungsvorschrift) + open items with reasons
- asset accountant review and release (four-eyes on classification and split)
- settlement in the asset ledger + audit log
- Deterministic first: amounts, sums and the check that the total of the split equals the total on the AiB are calculated in code. Asset classes and useful lives come from a maintained table, not from the LLM.
- LLM part: reading order texts, invoice lines and technician notes ("Kabel NS neu verlegt, 120 m, Hausanschluss ersetzt"), deciding which GIS object they refer to, and writing the reasoning for the accountant.
- Matching: embeddings and fuzzy rules link posting texts to GIS objects, street names and project numbers; past settlements serve as examples.
- Human-in-the-loop: release by the asset accountant. No write access to the ledger except drafts.
Data preparation
Which data: investment measures with postings, the chart of asset classes with useful lives, the written capitalisation guideline, a history of past settlements with the accountant's decisions, GIS exports of the network objects with installation and commissioning dates, grant notices, and Baukostenzuschuss records.
Where it lives: the ERP or utility system (SAP FI-AA/PS, Wilken ENER:GY, Schleupen, DATEV for small utilities), a GIS or network information system (for example Smallworld, ArcGIS-based or other NIS), project spreadsheets, PDF as-built plans and notices, and the head of the accountant.
Typical gaps: GIS and asset ledger do not share a key, so the link is manual; commissioning dates are missing or differ between systems; the capitalisation guideline is not written down or has exceptions known only to one person; maintenance and new construction are booked on the same order; old settlements are inconsistent.
Preparation steps: 1. Export two to three years of settled projects with their final assets and the accountant's splits (see PREP-005, PREP-020). 2. Write down the capitalisation guideline and the asset-class table with useful lives and valid-from dates (PREP-016, PREP-013). 3. Create a shared key between GIS object, order or WBS number and asset number; fix the worst mismatches first (DATA-003). 4. Collect commissioning dates and as-built documents per project in one place (PREP-006 if paper). 5. Label a test set of 50-100 settled projects with known correct results (PREP-018). 6. Restrict access to ledger exports and strip employee names from labour postings where they are not needed (PREP-022, PREP-023).
Tools: ERP exports, GIS export tools, spreadsheet for the test set.
Rough share of effort: high, about 40-50 percent of the project, mostly on the shared key and the written rules rather than the AI. See GUIDE-003.
Implementation paths
Which path for whom
- SMB (Stadtwerk) — Path B: no productised AI capitalisation agent is known; a small read-only build on cloud APIs in an EU region plus the ERP interface is the realistic route, ideally via the IT service provider.
- Enterprise (large utility) — Path B: volume justifies integration with SAP PS/FI-AA and an own evaluation set; strong governance.
- Developer (utility software vendor) — Path C: vendors of utility and asset software can embed self-hosted open models in their product.
Path A — Buy
- ERP and asset-accounting suites: SAP (FI-AA with investment measures and settlement rules), Wilken, Schleupen and regional utility IT providers offer rule-based settlement; AI-based proposals are mostly roadmap or pilot in 2026. Ask for production references on your release.
- Asset-accounting add-ons and finance automation tools: some tools propose asset classes from invoice lines; none that we found understands GIS objects.
- Choose A when: your ERP vendor ships a settlement assistant for your version.
Path B — Build on cloud APIs
- Models: an EU-hosted LLM or a frontier model in an EU region with structured output for classification of posting texts and drafting reasons.
- Orchestration: an agent framework or n8n-style workflow with read-only connectors to ERP and GIS; results are written back as a draft worklist or ticket.
- Matching: embeddings plus rules for GIS linking; classic tabular ML on past settlements for the split.
- Choose B when: you have an IT team or partner who can build the integration and an evaluation set.
Path C — Self-host
- Stack: open-weight LLM of roughly 30B class with retrieval over the guideline and past settlements on one workstation or pro GPU (HW2), GIS matching in code.
- Choose C when: ledger data may not leave the premises or you build a product for many utilities.
Path D — Custom training
Fine-tune a classifier on your own labelled settlements (asset class, capitalisable or expense). Worth it only with several thousand posting lines per year; a few hundred labelled projects already help prompting and evaluation.
State of the art (2026)
- Mature: rule-based settlement in ERP, document extraction from invoices (DOC-001), LLM classification of posting texts.
- Early: agents that follow a project across ERP, GIS and as-built documents and draft the split with reasons. We found no independent evidence of production use in German utilities; treat vendor claims with caution.
- Not advisable: unattended capitalisation. Capitalise-or-expense and useful-life choices are accounting judgement with tax and regulatory consequences (see HGB §255 on what belongs into production costs [1]).
- Limits: LLMs can invent rule references or misread an order text; every proposal must cite the posting, the GIS object and the guideline paragraph.
- Dependency: quality depends far more on the GIS-to-ledger key than on the model.
Shared resources & combinations
- Same stack (agent-framework + workflow-automation): AGENT-004 order entry, AGENT-006 reconciliation and AGENT-021 data entry reuse connectors and review UI.
- Upstream: DOC-001 reads supplier invoices and DOC-014 parses as-built PDFs; the agent consumes their structured output.
- Legacy UI: where the asset system has no API, AGENT-003 is a fallback, with the risks that brings.
- Domain neighbours: DATA-016 (asset models and context), PRED-033 (renewal planning needs the same asset register), OPT-036 (network rollout planning creates the projects), DOC-030 (quantity takeoff).
- Controls and governance: ANOM-006 (ledger anomalies), AGENT-018 (controls monitoring), AGENT-025 (agent monitoring).
- Best leverage together: a clean shared key between GIS and the asset ledger (DATA-003) pays off for settlement, regulatory cost audit and renewal planning at once.
Effort, cost & KPIs
- Path A: weeks of configuration if the vendor feature exists; add-on licence.
- Path B: 4-9 months for a read-only proposal assistant including key clean-up, evaluation set and a shadow period.
- Path C/D: 6-12 months plus operations effort.
- Cost drivers: integration with ERP and GIS and the data clean-up, not model fees.
- KPIs: share of proposals accepted unchanged, hours per settled project, number and age of commissioned-but-unsettled AiB, corrections found in audit, share of GIS objects linked to an asset.
People & change
Involve early: the head of accounting and the asset accountants as process owners, project controlling, the GIS team and network engineering (they hold commissioning data), IT for ERP and GIS interfaces, and the auditor or tax advisor before rules are encoded (GUIDE-005, GUIDE-010).
Works council: depends. Review logs show who classified and released what, which can monitor performance under §87(1) no. 6 BetrVG; §90 BetrVG requires information on AI-supported work. Agree that logs are not used for individual evaluation (GUIDE-004).
AI literacy: Art. 4 AI Act (as softened by the 2026 Digital Omnibus on AI, per secondary sources; check the consolidated text on EUR-Lex) requires measures that support AI literacy of staff; train accountants on checking cited rules and reporting wrong proposals (GUIDE-006).
Communication and pilot: present it as relief at year-end closing, not as replacement; pilot with one asset type (for example house connections or low-voltage cable) in shadow mode on last year's closed projects, name an AI champion (GUIDE-008, GUIDE-009).
Typical concerns: fear of audit findings from wrong classification (four-eyes release), fear that expert knowledge is devalued, and disputes between technicians and accountants about dates.
Humans in control: no capitalisation, split or useful-life decision enters the ledger without release.
Risks, regulation & compliance
- Commercial law: §255 HGB defines acquisition and production costs; what belongs in production costs (including mandatory and optional overheads) is a judgement call, and research and selling costs may not be included [1]. German GoBD requirements for traceable posting and procedural documentation apply to the ledger.
- Regulation (grid operators): network costs are determined from the asset base. Calculatory depreciation is linear, per asset, with useful lives from Anlage 1 StromNEV (for assets activated from 2006 on historical cost; legacy assets follow special rules with current replacement values) (for example medium-voltage cable 40-45 years, local-network transformers 30-35 years), fixed for the remaining period; Baukostenzuschuesse (paid by electricity-consuming connection customers) are released linearly over 20 years and reduce network costs [2][3][4]. Gas and water networks have their own ordinances.
- Tax: tax capitalisation and depreciation follow their own rules and may differ from the HGB and the regulatory view; the agent must keep these views separate.
- EU AI Act: minimal risk for back-office accounting support; the duty to support AI literacy (Art. 4, softened by the 2026 Digital Omnibus; check the wording) still applies [6].
- GDPR: mostly asset and cost data; labour postings and customer connection records (house connections) can identify persons; AVV (Art. 28) with providers, EU processing.
- Security: utilities are often critical infrastructure or in NIS2 scope; keep the agent read-only plus drafts and keep GIS data of the network in your own environment where required.
- Mitigations: versioned guideline with citations, shadow mode, restricted write rights, audit log, regular evaluation on fresh closed projects.
Getting started
- Measure the baseline: number of AiB closed per year, hours per project, age of unsettled AiB, past audit findings.
- Write down the capitalisation guideline and the asset-class table with useful lives.
- Check how GIS objects and ledger assets are linked today and fix the key for one asset type.
- Run a read-only matcher and proposal generator in shadow mode on last year's closed projects and compare with the real settlements.
- Roll out the best-performing asset types as drafts; keep release with the accountant. Do not plan autonomous booking.
Sources
- Federal Ministry of Justice — §255 HGB (Bewertungsmaßstäbe: Anschaffungs- und Herstellungskosten) — https://www.gesetze-im-internet.de/hgb/__255.html
- Federal Ministry of Justice — §9 StromNEV (Netzkostenmindernde Erlöse und Erträge, Baukostenzuschüsse) — https://www.gesetze-im-internet.de/stromnev/__9.html
- Federal Ministry of Justice — §6 StromNEV (Kalkulatorische Abschreibungen) — https://www.gesetze-im-internet.de/stromnev/__6.html
- StromNEV Anlage 1 (betriebsgewöhnliche Nutzungsdauern) — https://gesetze.legal/bund/stromnev/anlage_1
- SAP Help — Abrechnung von untergeordneten Objekten einer Investitionsmaßnahme (Anlagen im Bau) — https://help.sap.com/doc/cda2c6535e601e4be10000000a174cb4/700_SFIN20%20006/de-DE/5ca1c6535e601e4be10000000a174cb4.html
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj