Turing College Logo

German Tax Assistant: Work Expenses, Calculated & Explained

German Tax Assistant:  Work Expenses, Calculated & Explained

About this project

What it is. A German tax return is long, and Anlage N - where employees claim work-related expenses - is where deductions are easy to miss. The current release prepares Anlage N for the 2025 tax year and shows where every figure comes from. Nothing is filed with the Finanzamt, and the application does not provide tax advice.

How it works. An interviewer agent asks about your working year - commuting, home-office days, equipment and training - and chooses each next question from what you have already answered, prioritising the areas most likely to matter. Some questions are gates: they determine whether an entire expense category applies. The application will not finish the interview while a gate remains unanswered, regardless of the model's confidence.

Once every gate has been answered, the agent may propose stopping if the remaining categories cannot bring the total above the €1,230 employee allowance. This is only a proposal: you can accept it or continue the interview.

The models choose actions, but ordinary unit-tested Python performs every calculation. Each expense is presented with its formula, substituted values, provenance, Anlage N line and supporting passage from a knowledge base of 53 official German tax documents, including EStH 2025, LStH 2025, BMF circulars and the official Anleitung zur Anlage N. A second agent, running on a different model from a different provider, audits the completed case. A separate cited chat answers general Anlage N questions from the same official sources.

What you receive. The completed case produces a detailed printable report and a filled draft of the official Anlage N PDF, with the calculated figures placed on the relevant form lines. It also shows the total work-related expenses that may be claimed as Werbungskosten, together with a complete breakdown of how that total was calculated. This is the deductible amount to declare in the tax return, not an estimate of the eventual tax refund, which depends on the person’s complete tax situation. You review the draft and file it yourself.

Documents and privacy. You may upload a Lohnsteuerbescheinigung or an invoice. A vision model performs two independent extraction passes. If the results disagree, nothing is saved until you retry the extraction or enter the value yourself.Even when both passes agree, no extracted value enters the tax case until you review and confirm it.

The document is sent to the model as an image because there is no local OCR. Every model request requires a Zero Data Retention endpoint; if no compliant endpoint is available, the request fails. The original file is never stored. Only its file name and the values you confirm remain.

Extraction is deliberately limited. From a Lohnsteuerbescheinigung, the application requests only the tax year and employment period. From an invoice, it requests only the purchase details needed by the calculators. The full page is visible to the vision model, but the employee’s name, eTIN, tax class, salary and church-tax amount are not requested as extracted fields and are never persisted by the application. The interview itself contains no name, address, email address or tax ID.

What comes next. The next step is to expand from Anlage N to the additional forms needed for a complete return by an employed taxpayer. The longer-term direction is to support other filer profiles, including self-employed people and pensioners, with their corresponding forms and interview paths.

Each tax year will be supported as a separate, versioned package of official documents, rates, limits, form lines, a blank PDF and its field-position map. The knowledge base will be rebuilt from the official sources published for that year, while previous years remain frozen and available for late returns. The architecture already separates the agent workflow and calculator interfaces from year-specific rules, so the product can grow by adding verified yearly content and forms rather than redesigning the entire application.

Created byElena Krukenberg
Published atSeptember 15, 2026
CourseAI engineering
Looking for
New jobCo-founderFreelance projects
OpenOpen LinkedIn