Job costing and invoice assembly agent

Before an invoice can be written, someone has to find out what was actually done: hours from time sheets, material from delivery notes, extra work from a phone call, measurements from the site. In many Mittelstand firms this collecting and cross-checking takes longer than writing the invoice, and what is missed is revenue lost. An agent gathers the evidence per job or client, matches it against the order and the price terms, flags gaps, and drafts a verifiable invoice. A clerk checks and releases it. The tooling is feasible today; the hard part is scattered, unstructured source data.

Problem & value

• Today: in a plumbing and heating business (SHK-Handwerk), the final invoice needs time sheets from several fitters, material from the wholesaler's delivery notes, change orders agreed on site and prior advance invoices, spread over paper, WhatsApp, the ERP and the foreman's memory. In a tax practice, fee invoices need the services recorded per client and the correct StBVV items. In a hotel, the check-out invoice gathers folio postings; here the PMS already does most of the work and the value is low.
• AI value: fewer forgotten items (unbilled extras, material, hours), faster invoice cycle (and thus earlier payment), prüfbare invoices that survive objection, less rework and fewer disputes. The agent also gives a running job cost view (planned vs. actual) as a by-product.
• Gap noted in research: AGENT-030 covers time capture for professional services only; AUTO-015 generates documents from templates but has no job or fee logic. The estimated effort in the research is modest per firm (about 4 h per period for tax-practice fee invoices, 7 h per final invoice in SHK), which is why the value is in volume and in recovered revenue, not in a single big step.
• Evidence level: no independent benchmark for time savings is known to us; the savings range in the front matter is an estimate. Measure it in a pilot.

Variants

• Trades and construction (SHK, Elektro, Bau): final or partial invoice from order, Aufmaß, time sheets and delivery notes; billing per VOB/B with prüfbare Abrechnung and, for public clients, GAEB data exchange (X31 quantities, X86 award, X89 invoice) [2][4].
• Tax practice (Steuerkanzlei): fee invoice from recorded services, with the correct StBVV items, Gegenstandswert and hourly fees. The calculation must show fee amounts, short descriptions, the provisions applied and the matter value [1].
• Hospitality: assembling folio postings, extras and city tax at check-out. Normally solved in the PMS; the agent adds value only for group bookings or exceptions.
• Logistics: collecting billing items per client from tour data and delivery notes (see DOC-006).
• Engineering offices: phase-based fees (HOAI); same pattern, with phase logic as rules.

How it works

1. sources (time sheets, delivery notes, material withdrawals, Aufmaß, change orders, service records, mail)
2. capture & extraction (OCR/VLM for paper and photos, connectors for ERP/time-tracking)
3. assignment to job/client/contract (job number, address, project reference; confidence score)
4. completeness & plausibility check (hours vs. plan, material vs. bill of quantities, unapproved change orders)
5. pricing against the contract (unit prices, hourly rates, fee scale, discounts, advance invoices to deduct)
6. invoice draft (positions in contract order, evidence attached, XRechnung/ZUGFeRD/GAEB X89)
7. human review and release (open points listed first)
8. posting in ERP/accounting + GoBD-compliant archiving + dispatch

• Assignment and extraction: the AI part. A VLM reads handwritten time sheets and delivery notes; an LLM maps free-text entries to contract items and to the job. Unknown items go to an "unassigned" list rather than being guessed.
• Pricing and fee logic: deterministic code, not the LLM. Unit prices, hourly rates, VAT, reverse-charge cases and the fee scale (StBVV tables) must be reproducible. The LLM proposes which item applies; rules calculate.
• Human-in-the-loop: the clerk or foreman releases every invoice. Change orders without customer approval are shown as open points, never billed silently.

Data preparation

Which data: open orders and contracts with unit prices, time and service records, delivery notes and material withdrawals, Nachträge and customer approvals, Aufmaß, previous partial invoices, and customer and job master data.

Where it lives: ERP or craft software (e.g. weclapp, Business Central, Lexware, specialised Handwerker software), time-tracking apps, the wholesaler's portal, email and WhatsApp, paper delivery notes in the van, and the DATEV practice system for tax practices.

Typical gaps: time sheets filled out days later or without job number, delivery notes not tied to a job, verbal change orders without a written confirmation, price lists that are out of date, and several systems without a shared job key. This is why data preparation effort is high: the agent is only as good as the job key that links all sources.

Preparation steps: 1. Introduce one job or project number that every time entry, delivery note and withdrawal must carry (PREP-005). 2. Digitise recurring paper flows (time sheets, delivery notes) and agree on photo upload from the site (PREP-006). 3. Write down billing rules and fee-scale logic (PREP-016) and collect past good invoices as examples (PREP-020). 4. Build a test set of 30-50 closed jobs with the invoice that was actually sent (PREP-018). 5. Restrict access to personnel time data (PREP-023).

Tools: ERP exports, spreadsheet for rules and test set, scanner/phone upload. See GUIDE-003 for the readiness check.

Implementation paths

Which path for whom

• SMB — Path A: most trades and practices already use a vertical software (craft ERP, DATEV, practice software) and an add-on or built-in assistant is cheaper than a build. EU-hosted vendors exist.
• Enterprise — Path B: multi-site service firms with their own ERP, contract logic and approval rules need orchestration around their systems.
• Developer — Path C: open-source parsers (GAEB, e-invoice) plus a local LLM give a tailored prototype for software vendors and consultants.

Path A — Buy (recommended for SMBs)

• Vertical software with job costing: craft and construction ERP and AVA software (e.g. with GAEB import/export and Aufmaß-to-invoice), weclapp, Business Central with job module, Lexware Office for small volume. In tax practices: the DATEV practice suite and fee-invoicing modules, with time and service recording.
• AI add-ons: assistants for delivery-note capture and invoice drafting are emerging in these suites. Check what is in your product's roadmap before buying a separate tool. Be sceptical about "autonomous billing" claims; ask for a reference customer in your trade.
• Choose A when: your vertical software already holds jobs, prices and time entries and offers an AI or API extension. Check EU hosting, GoBD archive, XRechnung/ZUGFeRD output and GAEB support.

Path B — Build on cloud APIs

• Sources: ERP/REST APIs, Microsoft Graph for mail, time-tracking APIs. Models: EU-region endpoints or EU providers for extraction and mapping, with structured JSON output.
• Orchestration: n8n or Power Automate (AGENT-002), or a small service; deterministic pricing in code; e-invoice generation with open libraries (e.g. Mustang for ZUGFeRD/XRechnung, KoSIT validator).
• Choose B when: you need custom contract logic or integration to a non-standard ERP.

Path C — Self-host (open source)

• Parsing: pyGAEB (MIT) parses, validates and converts GAEB DA XML including X31 and X80-X89, and has optional LLM classification [4]; Docling or PaddleOCR for delivery notes; a local VLM/LLM (e.g. Qwen family) via vLLM or Ollama.
• Hardware: HW2 is enough; volumes are low and batched.
• Choose C when: personnel data or customer contract data must stay in-house, or you build a product for several firms.

Path D — Custom training

Rarely needed. Fine-tuning a small model on the firm's own delivery notes and booking history may help with unusual handwriting or abbreviations. Needs several hundred labelled examples.

State of the art (2026)

• What works: document extraction of delivery notes and time sheets (VLM), mapping of free text to contract items, completeness checks against bill of quantities, and drafting of invoice text. These are the same building blocks as in DOC-001 and AGENT-004.
• What is immature: end-to-end agents that cross ERP, time tracking, mail and site photos without a clean job key. Most production use is a semi-automatic assistant in a vertical software, not a free-roaming agent, hence maturity early-adopter.
• E-invoicing: domestic B2B issuing of structured e-invoices becomes mandatory from 1 Jan 2027 for firms with 2026 turnover above 800,000 EUR and from 1 Jan 2028 for all; receiving is mandatory since 2025 [3]. The agent's output must therefore be XRechnung or ZUGFeRD, not just a PDF.
• Limits: LLMs mis-assign items with similar names, cannot know about verbal agreements, and must not compute prices. Fee scales and tax rules are code. Handwriting and photo quality remain error sources.

Shared resources & combinations

• Same stack (ocr-docparse + vlm + llm-general): DOC-001 invoice processing, DOC-006 logistics documents, DOC-008 mailroom, DOC-003 forms. One VLM server serves all.
• Upstream: AGENT-030 supplies recorded time and narratives for professional services; AGENT-004 turns orders into ERP jobs with the job key this agent needs.
• Downstream: AGENT-023 handles collections, dunning and deductions after the invoice goes out; AGENT-006 reconciles; PRED-016 uses open receivables for cash-flow forecasts; AGENT-018 monitors billing controls.
• Highest combined effect: the biggest lever is the chain order (AGENT-004) → capture of time and material → this agent → collections (AGENT-023). Each step on its own saves hours; together they shorten the time from work done to cash received and cut forgotten billable items. Start where the job key is cleanest.
• Workflow engine shared with AGENT-002.

Effort, cost & KPIs

• Path A: weeks to a few months, mostly process clean-up; add-on pricing per user or per invoice.
• Path B: 2-4 months including ERP integration and test set. Path C: 3-5 months plus operations.
• KPIs: days from completion to invoice sent, share of invoices released without corrections, unbilled items found per month (recovered revenue), invoice disputes or objections, minutes of preparation per invoice, DSO.

People & change

Involve early: management (billing policy), the office manager or head of accounting as process owner, foremen or fitters who deliver the source data, IT, and the data owners of ERP and time tracking. The DPO joins because time data is employee data (GUIDE-005).

Works council: depends. If the agent reads individual time sheets or GPS site times, the system can be suited to monitor employee performance, so §87(1) no. 6 BetrVG co-determination can apply; §90(1) no. 3 requires early information on AI-supported work. Use the data per job, not per person, and agree this in a short works agreement (GUIDE-004). Many small trades have no works council, but the same data-protection duties apply.

AI literacy: Art. 4 AI Act requires measures; a short session for clerks on typical mis-assignments is enough (GUIDE-006).

Adoption: the real change is on site: foremen must enter time and photograph delivery notes with the job number. Present the agent as the end of "searching for missing notes" and not as control (GUIDE-008). Pilot with one trade or team and 10-20 closed jobs run in parallel (GUIDE-009).

Humans in control: release of every invoice, approval of every change order, and any price deviation stay with people.

Risks, regulation & compliance

• EU AI Act: minimal risk; no Annex III category for billing. Caution: do not reuse the data to rank or evaluate individual employees, which could move it toward the employment category. Art. 4 AI literacy applies.
• GDPR: time sheets and GPS site times are employee personal data (legal basis Art. 88 GDPR / §26 BDSG and works agreement; the scope of §26 BDSG is partly disputed after CJEU C-34/21, so check with your DPO); customer data on invoices is personal data for private clients and sole traders. AVV (Art. 28 GDPR) with every processor, data minimisation, EU hosting preferred.
• Tax and bookkeeping: §14 UStG mandatory invoice fields, GoBD (traceable processing, immutable archive, Verfahrensdokumentation), retention periods; structured e-invoice duties from 2027/2028 [3].
• Contract and fee law: under VOB/B §14 the contractor must bill in a prüfbar way, with quantity calculations, drawings and evidence attached, and changes clearly marked [2]; the final invoice is due within set periods after completion. For tax advisers, fees can only be demanded on a calculation in text form showing the items, provisions applied and Gegenstandswert [1]. Check the current text of both rules; the agent must implement them as rules, not as model judgement.
• Professional secrecy: tax practices are bound by §57 StBerG / §203 StGB; prefer EU or self-hosted processing and processor commitments.
• Billing integrity: invoicing items not delivered or unapproved extras is a legal and reputational risk. Keep an audit trail of source evidence per line.

Getting started

1. Pick one trade or fee type and measure the baseline: preparation time per invoice, days to invoice, forgotten items found afterwards.
2. Introduce the job key and photo upload of delivery notes (cheap and useful without any AI).
3. Check what your ERP or practice software already offers before buying anything.
4. Pilot on 20 closed jobs: compare the agent's draft with the invoice actually sent and list the differences.
5. Go live with release by a clerk; track recovered revenue and corrections monthly.

Sources

1. Bundesministerium der Justiz — § 9 StBVV (Berechnung der Vergütung) — https://www.gesetze-im-internet.de/stbgebv/__9.html
2. Haufe — VOB/B 2016 § 14 Abrechnung — https://www.haufe.de/id/norm/vobb-2016-14-abrechnung-HI10559881_p14.html
3. IHK Stuttgart — E-Rechnung: Ende der Übergangsfrist für Unternehmen mit mehr als 800.000 Euro Umsatz — https://www.ihk.de/stuttgart/fuer-unternehmen/recht-und-steuern/steuerrecht/steuermeldungen/e-rechnungen-5864496
4. pyGAEB documentation (GAEB DA XML parser, X31/X80-X89, optional LLM features) — https://pygaeb.readthedocs.io/
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj