AI in law firms: what is allowed under German law

In summer 2025, the family court in Cologne read a written submission based on commentary references and citations that do not exist. In its order, the court stated that the requirements cited there had ‘evidently been generated by means of artificial intelligence and were entirely made up’. Cases like this make headlines, and they distract from the real question many small law firms are currently asking themselves: what are we actually allowed to do with AI, and what not? The good news: AI is not banned in law firms. The rules are simply stricter than in most other industries, and they are not found in the AI Act but mainly in the professional rules for lawyers.

Key points at a glance

• AI is allowed in law firms. However, the German Federal Bar (Bundesrechtsanwaltskammer, BRAK) stresses that AI only provides support and that every result is checked by the lawyer on their own responsibility.
• Client secrets may only be passed to an AI provider under the conditions of Section 43e of the Federal Lawyers’ Act (BRAO): only to the extent necessary, after careful selection and with a contract in text form that binds the provider to confidentiality and informs it of the consequences under criminal law.
• A data processing agreement (DPA) is not enough for this. The standard contracts of OpenAI and Anthropic contain no obligation tailored to professional secrecy; Microsoft only offers one as a supplementary agreement.
• The safest route is genuine anonymisation. According to the BRAK, removing names is, as a rule, not enough if client information can be inferred from the context.
• The AI Act asks little of law firms: measures to promote AI literacy (with no fine attached) and transparency, for example for a chatbot on your website. AI used in law firms is generally not high-risk.
• Responsibility stays with you. Courts have already admonished lawyers for invented AI citations.

Are law firms allowed to use AI at all?

Yes. Neither the BRAO nor the professional code of conduct for lawyers (Berufsordnung) prohibits the use of AI. In early 2025, the BRAK published guidance on the use of AI (as of December 2024), which has not been updated since. Its basic line: AI may support a lawyer’s work but not replace it, and in every case the lawyer must review the results and carry out a final check on their own responsibility. In principle, the BRAK does not see any obligation to use AI.

Anything that does not involve a client matter is unproblematic: research on abstract legal questions, drafts of texts for the firm, outlines, help with wording, summaries of published judgments. It gets tricky as soon as information from a client matter comes into play. And it takes less for that to happen than many people think.

Confidentiality: broader than you might think

Lawyers’ duty of confidentiality covers everything that has come to your knowledge in the course of practising your profession (Section 43a(2) BRAO). This includes not only the contents of case files, but also the client’s name and even the fact that an engagement exists at all. Anyone who, without authorisation, discloses another person’s secret that was entrusted to them or otherwise became known to them as a lawyer may commit a criminal offence under Section 203(1) no. 3 of the German Criminal Code (StGB).

Since 2017, there has been a regulated way out: professionals bound by secrecy may pass secrets on to ‘other persons involved’ in their work (sonstige mitwirkende Personen), in other words service providers, to the extent that this is necessary for those providers’ activities (Section 203(3), second sentence, StGB). If you do not make sure that the service provider has been bound to secrecy, you can commit a criminal offence yourself if the provider discloses a secret without authorisation (Section 203(4), second sentence, no. 1 StGB). The counterpart to this rule in professional law is Section 43e BRAO, and this is exactly where it is decided whether and how a law firm may use an AI tool with client data.

Section 43e BRAO: the step-by-step check for AI providers

Section 43e BRAO governs the use of external services. The explanatory memorandum to the law expressly mentions IT service providers and the use of external applications and systems, and the BRAK counts providers of AI solutions among them. If you want to use an AI tool with client information, the best approach is to go through the provision subsection by subsection:

1. Is access to client secrets necessary? (subsection 1) The explanatory memorandum calls for strict standards. Based on the state of technology as of December 2024, the BRAK did not consider it necessary to transfer client secrets to language models such as ChatGPT, at least not with the freely available versions. So the task can often be done without any secrets.
2. Have you selected the provider carefully? (subsection 2) You must satisfy yourself that the provider is professionally suitable and reliable. If compliance with the requirements is no longer ensured, you must end the cooperation without delay.
3. Is there a contract in text form with the mandatory content? (subsection 3) The contract must bind the provider to confidentiality and inform it of the consequences under criminal law, limit its access to information to what is necessary, and specify whether and how subcontractors are brought in.
4. Is the service provided abroad? (subsection 4) This is only permitted if the protection of secrets there is comparable to that in Germany. For other EU countries, the explanatory memorandum generally assumes this to be the case. For providers from the USA, according to the BRAK, it has not been conclusively settled whether the level of data protection can be used as the benchmark for this; it therefore recommends giving preference to providers with servers in Germany or Europe.
5. Does the use directly serve one individual client matter? (subsection 5) Then you also need the client’s consent. The explanatory memorandum gives experts, private investigators and translators as examples. Whether an AI tool that you use in a specific client matter, for instance to translate or analyse a document, falls under this has not been settled.

And data protection law always applies alongside this (Section 43e(8) BRAO). So if personal data goes to the provider, you need both: the confidentiality agreement required under professional law and a data processing agreement under the GDPR.

Is a data processing agreement enough?

No. On confidentiality, a data processing agreement only requires that the people at the provider who work with the data have committed themselves to confidentiality or are under a statutory obligation of confidentiality (Article 28(3)(b) GDPR). That is not the same as being informed of the consequences under criminal law, as Section 43e(3) BRAO requires. A look at the contracts of the major providers, as of October 2026:

In practice, this means: with the standard plans from OpenAI and Anthropic, anonymisation is the main option left for client secrets. How the tools handle data in detail is covered in our tool checks on ChatGPT, Claude and Copilot.

Anonymisation: the safest route, done properly

If the prompt does not allow any conclusions about a client matter, no client secret is disclosed and Section 43e BRAO does not come into play. That is why the BRAK recommends, where possible, using only abstract prompts that allow no conclusions about a client matter and fully anonymising documents beforehand. The catch lies in the word ‘fully’:

• Removing names is not enough. According to the BRAK, deleting names and addresses is, as a rule, not sufficient if client information can be inferred from the context.
• Context gives a lot away. An unusual set of facts, a well-known company, a place and a date can together be enough to identify a client matter.
• Abstract rather than redact. Instead of uploading the written submission, phrase the legal question in abstract terms: ‘What is the legal position if a landlord …’ rather than the case file with the names redacted.
• Pseudonymising is not anonymising. If you replace names with placeholders and keep the key within the firm, the data remains personal data for you.

Should you inform clients or get their consent?

The BRAK does not see a general obligation under professional law to inform clients about the use of AI (as of its guidance from December 2024). It does, however, consider transparency and, in case of doubt, a contractual arrangement to be advisable. The Council of Bars and Law Societies of Europe (CCBE) is more specific in its guide from October 2025: if it can be assumed that an informed client would have objections or reservations about the use of AI, you should be transparent with them.

Three things apply regardless:

• Consent for an individual client matter: If your use of AI falls under Section 43e(5) BRAO, you need consent anyway.
• Privacy notices: If personal data goes to an AI provider, the provider must be listed as a recipient or category of recipients in the firm’s privacy notice, together with a note on the transfer to a third country if data goes to the USA (Article 13(1)(e) and (f) GDPR).
• Trust: Especially for sensitive client matters, a short clause in the engagement agreement is the easiest way to avoid misunderstandings.

Hallucinations: the responsibility stays with you

Language models occasionally invent judgments, case references and commentary passages, and German courts have already taken issue with this. In July 2025, the Cologne Local Court (Amtsgericht Köln) found that citations in a written submission were entirely made up and pointed out that deliberately spreading untruths violates the duty of objectivity under Section 43a(3) BRAO (order of 2 July 2025, 312 F 130/25). In November 2025, the Kammergericht (the Higher Regional Court in Berlin) admonished a lawyer who had cited, among other things, a non-existent decision of the Federal Court of Justice (BGH), and derived the duty to check from the engagement and from her position as an organ of the administration of justice (order of 20 November 2025, 17 WF 144/25).

In terms of liability, AI does not change the basic principle: you are liable to your client for culpable breaches of duty, with or without the help of tools. Professional indemnity insurance covers financial losses but may exclude losses caused by a knowing breach of duty (Section 51(3) no. 1 BRAO). Whether insurers will treat the unchecked use of AI results as a knowing breach is an open question. The simple rule to guard against this: every citation an AI provides is checked against the original before it goes into a written submission.

What the AI Act requires of law firms

Compared with professional law, the AI Act asks little of law firms. A law firm that uses AI tools in its professional work is a deployer within the meaning of the Regulation.

• AI literacy (Article 4): Since the Digital Omnibus, you must take measures to promote the AI literacy of your team; you do not have to guarantee a specific level. No fine is provided for this. Many guides still quote the old, stricter version. What lean training looks like is explained in the article on AI literacy training for employees.
• Transparency (Article 50): A chatbot on the law firm’s website must disclose that it is an AI no later than the start of the conversation. Formally, the obligation lies with the provider of the chatbot; if the firm has its own chatbot developed, the firm itself is the provider. Either way, make sure the notice appears. If you publish AI-generated texts to inform the public on matters of public interest, you do not have to label them if a human has reviewed them and someone holds editorial responsibility. Whether a specialist article on a law firm blog falls under this at all is unclear; with review and responsibility in place, you are on the safe side either way. Details in the article on AI labelling requirements.
• High risk: The high-risk category for the administration of justice (Annex III point 8(a)) concerns AI used by or on behalf of judicial authorities or in alternative dispute resolution, not the work of lawyers; the BRAK takes the same view. One exception may be AI in recruitment, for example to screen job applications. These obligations only apply from 2 December 2027.

Using AI in your law firm in six steps

1. Take stock: Which AI tools are already being used in the firm, including on private phones? Shadow AI is particularly risky in law firms.
2. Separate use cases: What works without any link to a client matter, what only in anonymised form, and for what do you actually need access to client secrets?
3. Choose tools: For anything linked to a client matter, only business plans with a data processing agreement, no training on your data and, where possible, data processing in the EU.
4. Clarify the Section 43e agreement: Does the provider offer a confidentiality agreement that includes information on the consequences under criminal law? If not: work with anonymised data only.
5. Write down the rules: An AI policy for the firm that sets out which data may go into which tool and that every result is checked. Your employees’ confidentiality obligation under Section 43a(2) BRAO is a good starting point.
6. Train the team and update the engagement agreement: a short training session with examples from everyday law firm practice, plus a clause on AI use in the engagement agreement.

The most common mistakes

• Using a personal account for client work. Without a data processing agreement and with possible use of your data for training, this is off-limits for client data.
• Thinking the DPA is enough. It fulfils the data protection obligation, not the confidentiality agreement required under professional law.
• Only redacting names. If the facts give away the client matter, nothing has been anonymised.
• Adopting citations without checking them. Invented citations get noticed, and the courts have already reacted.
• Not involving the team. If employees use their own tools, the best rule for the partners is of no use.
• Forgetting the law firm chatbot. Since 2 August 2026, it must identify itself as AI.

How Klaus helps law firms use AI

Most AI guides are written for companies without professional secrecy obligations. That is why Klaus explicitly asks in the free onboarding chat whether your law firm is subject to professional secrecy, and takes your answer into account when classifying your tools and in your AI policy. Klaus records your AI tools with their purpose and data categories and classifies each tool: allowed, allowed with conditions or prohibited. This produces your AI policy as a PDF, where everyone can look up which information may go into which tool.

The training module guides you through a short training session with comprehension questions (for one person on the Free plan, for the whole team from the Basic plan upwards) and records participation in your training records. And if you want to check a tool more closely, for example a chatbot for client intake, you get an assessment project with prioritised tasks and a compliance report as a PDF from the Basic plan upwards. The legal texts behind it are pre-drafted by legal professionals, not freely generated by an AI.

Conclusion

AI in law firms is allowed, but not without limits. Without any link to a client matter, you can use AI tools just like other businesses. As soon as client secrets come into play, Section 43e BRAO is decisive: necessity, careful selection, a contract with a confidentiality obligation and, abroad, comparable protection. The standard contracts of the major providers do not meet these requirements on their own, which is why genuine anonymisation is the most practical route for many law firms. If you also put clear rules in writing, train your team and check every AI answer, you can use AI safely in your law firm, and the AI Act is then taken care of with little extra effort.