Skip to main content

AI and Personal Finance in Germany: Where It Helps and Where It Falls Short

In the Geldchen survey (2026, 118 participants), 80% named AI among their sources of information about money, almost twice the share that named official sources (42%) [1]. For many, AI is already one of the main sources of information on money. It explains general principles soundly and makes mistakes with exact figures and procedures. In many cases, such errors change little.

How accurate are AI answers about money?​

The results diverge. Models pass exams for finance professionals. On everyday money questions, a third of answers or more are wrong or incomplete, and 88% of multi-step ones.

StudyWhat was testedResultIn favour of AI?
CFA Level III (2025)A mock exam for financial analystsReasoning models pass: o4-mini scores 79% against a pass mark of about 63% [2]yes
TaxCalcBench (2025–2026)A US tax return without a single error32% correct in 2025, 63–66% in 2026 [3]improving
University of Bayreuth (2026), 3,700 peopleA choice between two funds on the same index64% picked the better-value fund without advice, about 87% after correct AI advice and about 34% after wrong advice [4]depends on the advice
Saturn (UK fintech, 2026)121 questions on debt, pensions and taxes57% of answers had an error or lacked a needed warning, 88% on multi-step questions [5]no
Which? (2025), as reported by Stiftung Warentest (2026)40 UK questions to 6 chatbots, not only about moneyAbout one answer in three is wrong [6]mostly no

Geldchen found no published tests on everyday money questions in German from 2025–2026. Accuracy on German topics is an estimate by analogy.

Where does AI get it wrong, and when does that matter?​

AI most often errs on thresholds, fees, payouts, deadlines and steps in forms. Often this changes little, and tracking every yearly change in thresholds gains almost nothing.

In 2025, the WISO tax assistant gave the 2024 Grundfreibetrag (basic tax-free allowance) as €11,604 instead of the retroactively raised €11,784, even though the program itself calculated correctly [7].

The exact figure matters when:

  • the step is hard to undo: a transfer, a signed contract, a tax return sent through ELSTER (the tax portal);
  • there is a deadline: an Einspruch (objection) against a Steuerbescheid (tax assessment) is filed within one month of the day the assessment counts as delivered; for a letter sent by post, that is the fourth day after it was sent (§ 355, § 122 AO) [8];
  • a one-off amount is large for your budget.

In those cases, the figure comes from the authority's letter, an official website or the gross-to-net calculator; sources on taxes are in the article on the tax return.

What is AI useful for in money matters?​

AI usually sets out general principles soundly: a cash reserve before investing, broad diversification, low costs, no attempts to time the market. Models know the textbook material (they pass the CFA exam [2]), and correct AI advice improved fund choice [4]. In an experiment by Conlon and Schwardmann (2026, 1,500 participants, tasks not only about money), a conversation with AI made decisions more moderate on average, even though the model played along with the user [9]. Advice on a personal portfolio is weaker: in a study by TUM (Technical University of Munich, 2025), it overloaded portfolios with US stocks and recent growth leaders and raised costs [10].

AI also suggests who to turn to:

  • complex taxes: a Steuerberater (tax adviser);
  • an employee's tax return: a Lohnsteuerhilfeverein (wage-tax help association);
  • pension: DRV-Beratung (free pension advice from the state pension insurer);
  • investing without a product sale: the Verbraucherzentrale (consumer advice centre).

What might AI leave out?​

A chat answer solves exactly the task you set. What it often leaves out:

  • Other paths. Asked "which ETF?", a chat answers about ETFs but rarely asks whether paying off an expensive loan or building the reserve comes first.
  • The full cost. An option that looks expensive is dismissed quickly, though over its whole term it may cost no more. For a cheap one, the total cost of ownership (TCO) is often left out: fees, taxes, upkeep, time.
  • The chain of consequences. Outside standard scenarios, models handle "if–then" reasoning over several steps poorly: what happens after a job change, a move or a rate rise. Multi-step questions had the most errors (Saturn, 88%) [5].

Questions to the chat itself close these gaps: which other options to compare, what it costs over the whole term, and what happens one or two steps later.

What not to upload to an AI chat, and how do scammers use AI?​

The most sensitive documents are the payslip and the Steuerbescheid. A Lohnabrechnung (payslip) reveals your Steuer-ID (tax identification number), social insurance number and church-tax marker. The marker shows church membership, a special category of data under the DSGVO (GDPR) [11].

Since June 2025, BaFin (the financial supervisor) has warned about more than 700 near-identical "Investieren mit KI" (investing with AI) websites: "AI trading" from €250, no Impressum (legal notice of the site owner) and no BaFin licence [12]. Deepfakes serve as bait too, for example a fake video of Friedrich Merz (Stiftung Warentest, 2025) [13]; other red flags are in the catalogue.

What not to do when discussing money with AI?​

  1. We give PINs, TANs (one-time codes) and online-banking passwords neither to a chat nor to an AI agent.
  2. We don't upload documents until the Steuer-ID, IBAN and social insurance number in them are masked.
  3. We don't confirm a TAN before checking the recipient and amount in the bank's app, even if an AI agent prepared the payment (payment scams).
  4. We don't use a figure that a large or urgent decision depends on without a year and a source.
  5. We don't take a deadline from an authority's letter out of a chat translation without checking the original.
  6. We don't mistake "AI trading" and celebrity videos promising income for an investment.
  7. We don't hint at the desired answer or figure in the question: the model tends to start from it (anchoring effect).
  8. We don't take a step that is hard to undo on the strength of one chat answer.
  9. We don't treat a bank's or broker's assistant as neutral: that is a conflict of interest, and in the Bayreuth experiment, the label "bank's AI" barely weakened its influence [4].
  10. We don't count on compensation for a general chatbot's mistake.

FAQ​

This is not legal or financial advice.

Can ChatGPT be trusted on German tax questions?​

ChatGPT mostly explains general rules correctly; it does worse at working out a specific case. In a test at OTH Regensburg (a university of applied sciences, 2026), the chatbot reliably solved standard exam problems in accounting and tax law. It failed a complex tax case unless it got repeated follow-up prompts [14].

Is it safe to upload tax documents to an AI chat?​

The risk is lower if the identifiers are masked and training on your chats is switched off in the settings. A Steuerbescheid reveals income, IBAN, church tax and sometimes disability deductions, which is data on religion and health.

Who is responsible if AI gives a wrong financial answer?​

It depends on the case. Public AI tools that give investment advice are not authorised or supervised by financial regulators, don't owe you an authorised firm's duty to act in your interest, and may not offer access to an ombudsman, ESMA (the EU securities markets supervisor) warns [15]: the usual safeguards of an investment service are missing. A company is responsible for its own bot: in May 2026, on a consumer association's claim, the OLG Hamm (Higher Regional Court of Hamm) barred a clinic from false claims its chatbot made about doctors' qualifications under the unfair competition law (UWG) [16]. That is an injunction, not damages, and the ruling is not yet final. Geldchen found no cases about chatbots' financial answers.

Sources​

  1. Geldchen (2026) — Деньги после переезда в Германию ("Money after moving to Germany"), anonymous online survey, N = 118, 1 August – 1 September 2026; question on sources of information (the breakdown is not published on the results page). Survey description: https://geldchen.com/research/money-after-moving/
  2. arXiv — Advanced Financial Reasoning at Scale: A Comprehensive Evaluation of Large Language Models on CFA Level III (arXiv 2507.02954); Reasoning Models Ace the CFA Exams (arXiv 2512.08270), 2025. https://arxiv.org/abs/2507.02954, https://arxiv.org/abs/2512.08270
  3. Column Tax — TaxCalcBench: 2025 paper (arXiv 2507.16126) and v2 results table (tax years 2024 and 2025, with web search), June 2026. https://arxiv.org/abs/2507.16126, https://github.com/column-tax/tax-calc-bench
  4. Universität Bayreuth — Wie KI Finanzentscheidungen beeinflusst (press release, 3,700 participants; 64%, about 87% and about 34% across the whole sample), August 2026; Phys.org — AI financial advice can improve decisions or steer investors wrong, experiment finds, 2026. https://www.uni-bayreuth.de/pressemitteilung/ki-finanzentscheidungen, https://phys.org/news/2026-08-ai-financial-advice-decisions-investors.html
  5. Saturn — Artificial Authority: Should you trust AI to deliver financial advice?, September 2026; Financial Reporter — AI models give wrong financial advice 57% of the time, Saturn research finds, 2026. https://www.saturnos.com/report/artificial-authority, https://www.financialreporter.co.uk/ai-models-give-wrong-financial-advice-57-of-the-time-saturn-research-finds.html
  6. Which? — Can you trust AI? ChatGPT and other AI chatbots put to the test, November 2025; Stiftung Warentest — KI-Chatbots im Test: Perplexity schlägt ChatGPT und Meta AI, 2026. https://www.which.co.uk/news/article/can-you-trust-ai-chatgpt-and-other-ai-chatbots-put-to-the-test-aetjt5e0RnPB, https://www.test.de/KI-Chatbots-im-Test-Perplexity-schlaegt-ChatGPT-und-Meta-AI-6275046-0/
  7. heise Mac & i 2/2025 — Schlaue Steuererklärung (review of the WISO SteuerGPT AI assistant), 2025. https://www.heise.de/select/mac-and-i/2025/2/2502007565086966445
  8. Abgabenordnung § 355 Einspruchsfrist; § 122 Abs. 2 Bekanntgabe (fourth day after posting, since 1 January 2025) — gesetze-im-internet.de, 2026. https://www.gesetze-im-internet.de/ao_1977/__355.html, https://www.gesetze-im-internet.de/ao_1977/__122.html
  9. Conlon J., Schwardmann P. — AI Sycophancy and Decisions, CESifo Working Paper 12681, May 2026. https://www.ifo.de/en/cesifo/publications/2026/working-paper/ai-sycophancy-and-decisions, https://arxiv.org/abs/2607.28133
  10. Winder P., Hildebrand C., Hartmann J. (TUM, Universität St. Gallen) — Biased echoes: Large language models reinforce investment biases and increase portfolio risks of private investors, PLOS ONE, 2025. https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0325459
  11. Entgeltbescheinigungsverordnung § 1 (contents of the payslip: Versicherungsnummer, Steuerklasse, Kirchensteuermerkmale, Steuer-ID) — gesetze-im-internet.de, 2026; Regulation (EU) 2016/679 (GDPR), Art. 9 — EUR-Lex, 2016. https://www.gesetze-im-internet.de/entgbv/__1.html, https://eur-lex.europa.eu/eli/reg/2016/679/oj
  12. BaFin — BaFin warnt vor Plattformreihe "Investieren mit KI", 3 June 2025, updated in 2026. https://www.bafin.de/SharedDocs/Veroeffentlichungen/DE/Verbrauchermitteilung/unerlaubte/2025/meldung_2025_06_03_Investieren_mit_KI.html
  13. Stiftung Warentest — Cybertrading-Betrug: Abzocke mit gefälschtem Video von Friedrich Merz (Finanztest 11/2025), 2025. https://www.test.de/Cybertrading-Betrug-Abzocke-mit-gefaelschtem-Video-von-Friedrich-Merz-6251478-6251488/
  14. OTH Regensburg — Besteht ein Chatbot die Steuerrechtsklausur? (Prof. Claus Koss, exam in accounting and tax law), 2026. https://www.oth-regensburg.de/news/detailansicht/kuenstliche-intelligenz-auf-dem-pruefstand
  15. ESMA and national supervisors — Using Artificial Intelligence for Investing: What you should consider, 25 March 2025. https://www.esma.europa.eu/sites/default/files/2025-03/ESMA_Warning_on_the_use_of_AI_-_EN.pdf
  16. Oberlandesgericht Hamm — Urteil in Verbraucherzentrale Nordrhein-Westfalen e.V. gegen Aesthetify GmbH, 12.05.2026, 4 UKl 3/25 (press release; appeal on points of law to the BGH allowed), 2026. https://www.olg-hamm.nrw.de/behoerde/presse/Pressemitteilungen/16_26_PE_KI-Chatbot/index.php