Autonomous Tax Processing & Currency Normalization

David Prinz runs a German tax consultancy specializing in international clients - entrepreneurs, freelancers, and businesses with income across multiple countries and currencies. His clients send invoices, bank statements, and financial documents in a mix of languages: English, Dutch, French, and more.

The consultancy was growing, but the manual workload was growing faster. Each document required translation, currency conversion, and classification into German tax categories before it could even be processed. David needed a way to handle this without hiring more staff.

A German tax consultancy (managed by David Prinz) struggled with the manual burden of processing international financial documents. Inbound emails contained invoices and bank statements in multiple languages and various currencies (USD, GBP, etc.).

The manual workflow required staff to translate documents, manually look up historical exchange rates for the exact transaction date, and then classify each document into specific German tax categories (Einkunftsarten). This process was slow and highly susceptible to human error.

• Manual FX Lookups: Time-consuming historical currency conversion
• Language Barriers: Multilingual documents requiring manual translation
• Complex Classification: High-effort sorting into 7+ German tax categories
• Scaling Bottleneck: Growth limited by the number of manual processors

Deep Loom built an end-to-end autonomous pipeline using n8n and Gemini 2.5 Pro that 'thinks' like a German tax assistant.

The system triggers upon receiving a Gmail attachment, automatically uploads it to Google Drive, and performs multi-stage agentic processing:

• Agentic OCR & Translation: Intelligent text extraction and high-fidelity German translation using Gemini 2.5 Pro
• Autonomous FX Normalization: An AI agent that detects currencies and fetches historical exchange rates via API to convert all values to EUR automatically
• Tax Classifier Agent: A specialized agent that categorizes documents into legal German tax buckets (e.g., Kapitalvermögen, Gewerbebetrieb)
• Structured Extraction: Final cleanup and extraction of 20+ financial data points into a centralized Google Sheets audit trail

100% of FX conversions are now automated - every currency, every date, every document. Manual FX lookup errors dropped to zero. The time from receiving a document to having it classified, translated, and audit-ready went from 15–20 minutes per document to seconds.

The multi-agent architecture was the key technical decision. A single LLM prompt handling OCR, translation, FX lookup, and tax classification would have been fast but brittle. Splitting it into specialized agents - each with a narrow, well-defined task - made the system both more accurate and easier to audit. When a classification looks off, the audit trail shows exactly which agent made which decision and why.

The consultancy reduced manual task volume by 90%. David's team now focuses on the advisory layer - interpretation, client communication, strategic decisions - while the system handles the extraction and normalization work that was consuming most of their processing time.