May lawyers use AI, and what is it worth using for in a law firm?

What do the professional rules require when lawyers use AI?

Three duties carry the answer. § 43a(2) BRAO imposes a duty of confidentiality about everything that has become known in the course of professional work, which according to the BRAK (German Federal Bar) guidance includes clients' names and the fact that a mandate exists. § 43e BRAO governs when service providers, including AI providers, may gain access to these facts. § 43 BRAO requires conscientious professional practice and thus your own review of every work result.

• Access only as far as necessary (§ 43e(1)), and only after careful selection of the provider (para. 2).
• Contract in text form (para. 3): a confidentiality obligation with instruction on the consequences under criminal law, knowledge only as far as necessary to perform the contract, and a rule for further persons.
• Service providers abroad (para. 4): only if the protection of secrets there is comparable to that in Germany.
• The client's consent (para. 5) if the service directly serves a single client matter.

The contractual obligation also protects you. Under § 203(4) sentence 2 no. 1 StGB (German Criminal Code), anyone who has failed to ensure that a person involved in their work was bound to secrecy commits an offence if that person discloses a secret. According to the BRAK guidance, whether the provider actually looks at the data is irrelevant for access; the possibility is enough. The details are in the article AI and § 203 StGB. For tax advisers, § 62a StBerG is a provision with the same structure, as the table in the article AI for tax advisers shows.

What do the BRAK and DAV advise?

Both consider AI permissible under professional law, but draw the line on client secrets differently. In December 2024, the German Federal Bar published guidance on the use of artificial intelligence, and in July 2025 the German Bar Association (DAV) published its own-initiative statement no. 32/2025. Read side by side:

In practice, this means: if you follow the BRAK's more cautious line, you do not put client secrets into a language model at all. The DAV's line requires a provider that is bound in text form under § 43e(3) BRAO and excludes training on client data by contract. Both agree that responsibility for the result stays with the lawyer.

Which Legal AI tools are there for German law firms?

Besides general chat assistants, legal publishers and legal tech providers offer their own working environments. They connect a language model to a legal database and support answers with citations. As of September 2026, the providers describe their offerings as follows:

The table is not a ranking and not a recommendation; it reflects what the providers write themselves. For each offering, check the contract documents under § 43e BRAO, not the marketing page. General assistants such as ChatGPT or Claude do not access these publishers' databases; their data routes are compared in the articles Biwak or ChatGPT and Biwak or Claude.

Which tasks in a law firm is AI suitable for?

AI helps most where a lot of reading, sorting and wording is done and a lawyer can check the result against the file. The DAV names research, document analysis, for example in due diligence reviews, drafting contracts and documents, summaries and the transcription of conversations as typical areas of use.

Why must citations in AI texts always be checked?

Because language models can invent judgments, commentary passages and margin numbers that look genuine. In July 2025, the Cologne Local Court criticised a submission whose legal arguments and citations, in the court's view, had evidently been generated with artificial intelligence and were entirely made up. Under one cited FamRZ reference, the court found decisions on completely different legal questions, and a cited commentary passage had been written by a different author.

How often this happens depends on the model and on whether it accesses a database. In January 2024, a Stanford University study found hallucination rates, that is, shares of invented answers, of 69 to 88 per cent for the general models widely used at the time when asked specific questions about decisions of US courts (Stanford HAI). The figures apply to older models and a different legal system, but the pattern remains: the more precise the question about a citation, the greater the risk of an invented answer.

What does the EU AI Act mean for law firms?

Law firms are usually deployers of general AI tools, not providers of a high-risk system. Annex III point 8(a) of the EU AI Act covers systems intended to be used by a judicial authority or on its behalf to assist it in researching and interpreting facts and the law and in applying the law to a concrete set of facts, as well as systems used in a similar way in alternative dispute resolution. The BRAK concludes from this that the systems used in law firms are not covered, because the provision refers to judicial authorities. If you use AI close to the decision in conciliation or arbitration proceedings, you should check the classification separately. In any case, after the amendment by Regulation (EU) 2026/1744, the obligations for high-risk systems under Annex III only apply from 2 December 2027.

For law firms, Art. 4 is what matters most. Since 27 July 2026, it has required providers and deployers to take measures to support the development of their staff's AI literacy; they do not have to guarantee a particular level. The BRAK guidance of 2024 still quotes the old, stricter version. Training and a written firm policy nevertheless remain sensible, as the article AI literacy training describes. If you run a chatbot for clients on your website, also check the transparency obligations under Art. 50; the deadlines are set out in EU AI Act: which deadlines apply from when.

How does a law firm get started with AI safely?

With a written firm policy, two or three tasks without client secrets, and a review of the contract under § 43e BRAO before the first real file goes into a tool.

1. Write a firm policy

   Which tools are approved, which data may go in, who checks results? Rule out private accounts for firm work.

2. Start with tasks that involve no client matter

   Compare contract templates, summarise training materials, draft website texts and checklists. No client secret is involved here.

3. Check the provider

   Contract in text form under § 43e(3) BRAO, processing on behalf under Art. 28 GDPR, server location and subcontractors, no training on your data, retention period, access in a support case.

4. Train the team

   Strengths, weaknesses and the checking routine for citations. You have to bind your employees to confidentiality in text form under § 43a(2) BRAO anyway; remind them that this also applies to AI inputs.

5. Check and record

   Note which tool was used for what and who checked the result.

6. Only then client data

   Only with a provider bound under § 43e BRAO and, if the tool directly serves a single client matter, with the client's consent.

Templates and clauses yes, client secrets no: Biwak in a law firm

Today only for work without client secrets or with fully anonymised files. In a law firm, that mainly means contract templates, clause libraries, checklists and training materials. Biwak's AI agent works in a folder you assign to it and reads Word files, PDFs with a text layer, Excel, CSV and emails saved as .eml files there. It saves results as CSV, HTML or Markdown, so a clause comparison comes as a table and not as a finished submission. Because Biwak backs up the folder before each task, an unwanted change to a template can be undone with one click.

For § 43e BRAO, the actual data route is what counts; a local app on its own does not meet any requirements of professional secrecy. With a model task, Biwak transmits the question, the context needed from the conversation and the contents of the files used to the model provider. Processing happens in the EU: at Microsoft Azure in the EU Data Zone, or as a fallback via Prem AI (contracting party PREM SA, Switzerland). At Azure, according to Microsoft, inputs are accessible neither to other customers nor to the model makers and are not used for training. Before selecting a provider under § 43e(2) BRAO, you should nevertheless know four points: Microsoft may store suspicious requests for abuse detection and have them reviewed by humans, with storage in Germany. For the fallback route, when Azure rejects a task, the following applies: Prem names its compute partners Nebius and TensorX, but a DPA with Prem has not yet been signed; Prem assures, by its own account, by contract and organisationally, not technically, that inputs are not stored; and Biwak does not currently use the encrypted mode that Prem recommends for legal documents. Biwak's DPA contains a confidentiality undertaking in text form; none of the model providers has expressly committed itself under Section 203 StGB. If you follow the BRAK's line, you do not put client secrets into any language model anyway; if you follow the DAV's line, you check the provider yourself beforehand. Conversations from the desktop app are stored on your computer and only reach Biwak if you send an error report or a report yourself; the details are in the privacy policy.

Frequently asked questions