AI in German Tax Firms: What Professional Law, § 203 StGB and the GDPR Require

A firm that uses AI is working in several areas of law at once. Professional law, criminal law, data protection and the EU AI Act each ask their own questions, and meeting one requirement does not replace another. This guide sets out the layers and points to the articles that go deeper on individual questions. Its main basis is the FAQ of the Federal Chamber of Tax Advisors (Bundessteuerberaterkammer, BStBK) of 27 October 2026.

May a German tax firm use AI?

Yes. The BStBK considers AI in tax advice permitted, but only "im Rahmen von Datenschutz, Berufsrecht und klaren Compliance-Regeln", meaning within data protection, professional law and clear compliance rules. According to the FAQ there is currently no duty to use it. The Tax Advisory Act (Steuerberatungsgesetz, StBerG) is technology-neutral: it applies to all professional work, including work where an AI tool helps. For lawyers, the Federal Bar (Bundesrechtsanwaltskammer, BRAK) has published its own guidance on AI.

Which rules apply side by side?

Six layers apply side by side, and each asks something different of the firm.

What does confidentiality mean for AI use?

It does not prohibit external AI services, but it attaches conditions. The duty in § 57(1) StBerG covers everything learned in the course of professional work. § 203 of the Criminal Code (Strafgesetzbuch, StGB) makes the unauthorised disclosure of another person's secrets a criminal offence, and it expressly includes business and trade secrets. Protection therefore reaches further than the GDPR: data of a limited company is covered too.

A vendor is not automatically an unauthorised third party. § 203(3) StGB permits secrets to be disclosed to persons who contribute to the professional's work, as far as that is necessary for their contribution. The firm must, however, ensure that these persons are bound to secrecy. If it fails to do so and such a person discloses a secret without authorisation, § 203(4) StGB makes the professional liable to prosecution as well.

Without this safeguard, the BStBK leaves two routes: the client's consent, or anonymisation or pseudonymisation. Details that point to clients or engagements therefore do not belong in publicly accessible AI services. The FAQ treats professional use of private AI as impermissible. According to the FAQ, the same requirements apply accordingly to general cloud applications.

What must the contract with the AI vendor say?

§ 62a StBerG requires a contract in text form that binds the vendor to secrecy and warns of the criminal consequences. It limits knowledge to what is necessary and settles whether further persons may be used. The vendor must be chosen carefully, and for services performed abroad the protection of secrets must be comparable. If the service directly serves a single engagement, the client's consent is needed as well.

For lawyers the parallel provision is § 43e of the Federal Lawyers' Act (BRAO), and for auditors § 50a of the Auditors' Code (WPO). The checklist is in the article Confidentiality agreement under § 62a StBerG.

What does the GDPR add?

As soon as an AI system processes personal data, the GDPR applies in full, and § 62a(8) StBerG confirms that it remains unaffected. The firm needs a legal basis under Art. 6. Depending on the use, the BStBK names performance of the client contract or legitimate interests, and for dictation AI usually consent.

If an external vendor runs the application on the firm's behalf, the guidance of the German data protection authorities (Datenschutzkonferenz, DSK) says this is often processing on behalf of a controller (para. 32). A contract under Art. 28 GDPR is then required. Security measures under Art. 32, data minimisation, deletion concepts and informing data subjects under Art. 13 come on top.

A risk assessment comes before the start. If it points to a likely high risk, a data protection impact assessment under Art. 35 GDPR is required, which according to the DSK will often be the case for AI applications (para. 39). That has a consequence for small firms: under § 38(1) of the Federal Data Protection Act (BDSG) a data protection officer must then be appointed regardless of headcount.

On location, the BStBK recommends European vendors or data centres in the EU. The DSK prefers technically closed systems (para. 20) and applications that do not use inputs and outputs for training (para. 24). Ask where each individual processing step takes place, not just where data is stored. The vendor has to supply information about the system, because without it the firm can neither inform people nor assess the risk (paras. 21, 40).

What applies to recorded conversations and AI minutes?

Here § 201 StGB comes in, which makes it an offence to record the non-public spoken word without authorisation. According to the BStBK a recording is only permitted with the prior consent of everyone involved, and consent after the fact is not enough. The chamber classes file memos, conversation notes and transcriptions as dictation AI, considers them permitted in principle and regards offline use as preferable. The consent requirement applies to every tool, including one without a meeting bot.

The details are in the article May an AI tool transcribe client meetings?. What matters when choosing a meeting assistant is described in the guide for tax firms.

Do clients have to be told about AI use?

That depends on the situation. According to the BStBK there is currently no express professional duty to disclose it. Data protection law does require information once personal client data flows to an AI service, and any recording needs consent in any case. The article Do I have to tell clients that I use AI? separates these cases.

What duties does the AI Act bring?

For firms, AI literacy and transparency are the main points. Anyone using AI in a firm on their own responsibility is, according to the BStBK, generally a deployer under the regulation.

Art. 4 of the AI Act requires deployers to take measures to support the development of AI literacy among their staff. This wording has applied since 27 July 2026, when Regulation (EU) 2026/1744 entered into force. No particular level of literacy has to be guaranteed for anyone. The BStBK FAQ still reflects the earlier wording here. Separately, the duty to keep up professional training under § 57(2a) StBerG remains.

Art. 50 of the AI Act covers specific situations, such as chatbots, emotion recognition, deep fakes and published text on matters of public interest. According to the BStBK, purely internal drafts and meeting notes generally need no label. The FAQ says typical firm uses such as research or drafting do not generally fall into the high-risk categories.

What applies towards the firm's own staff?

The same data protection duties, plus employment law rules of their own, because AI solutions regularly process staff data too. The BStBK names information duties towards staff and, where there is a works council, its co-determination rights, for example where performance and behaviour are monitored. The DSK recommends clear internal rules and points out that introducing an AI application can trigger co-determination (paras. 36, 37). § 201 StGB also applies to internal meetings: its wording does not distinguish between clients and colleagues.

Who is responsible for AI results?

The professional. The duty to work independently and conscientiously under § 57(1) StBerG also applies to AI-assisted work. The BStBK requires a qualified person to check and approve every result, and compares AI output to the work of inexperienced staff. Towards the client, the tax advisor is liable for errors. According to the FAQ such errors are generally covered by professional liability insurance, but the policy terms remain decisive.

Where should a firm start?

With an inventory: which AI tools does the team already use, including unofficially? That produces an AI register and a short policy that sets out which data may go into which systems. According to the FAQ there is no statutory form for this documentation, and a lean table is often enough. The scope follows the risk: the closer AI gets to core work and client data, the more should be recorded.

New tools go through an approval process and a test phase with harmless data. The questions for vendors are in the article What should I look for when choosing a vendor?, and the chamber's statements are summarised in The FAQ in 12 points.

For client meetings there are tools designed for this setting. TaxMeet (Next Intelligence GmbH, Berlin) is designed for § 203 StGB: no meeting bot joins the call, capture runs per participant on the device, and nothing is executed without human approval. Checking against this guide and obtaining consent from everyone involved remain the firm's job.