This isn't legal advice, just the operator's perspective I use when I build AI employees for businesses. AI for lawyers pays off today mainly in the groundwork: drafts of briefs and letters, research, contract review, client communication, and sorting incoming mail with suggested deadlines. Responsibility stays with you: you check every citation at the source, and client secrets only go into an AI tool under the conditions German professional law sets for service providers. The German Federal Bar (Bundesrechtsanwaltskammer, BRAK) says in its AI guidance that AI may support a lawyer's work, not replace it. Everything below refers to German and EU law; elsewhere, your bar's rules apply.
I'm a developer, not a lawyer. I run my own business with fourteen AI employees and build the same kind of employees for companies, including a tax advisory firm in Germany. That firm receives 30 to 40 digital tax assessments a day, and its AI employees cannot submit anything to the tax office or trigger payments, because their user in DATEV, the firm's accounting software, does not have those rights at all. German criminal law lists tax advisors next to lawyers for professional secrecy. This article is about lawyers; the tax case is described in AI Agents for Tax Firms: A Real Case. No interview with a law firm went into this article: the law firm examples are hypothetical, the legal framework comes from statute, the BRAK and case law. AI employees are explained in What Is an AI Employee?.
AI for lawyers: five areas of work in a law firm
The result is always a draft or a work list. For contracts, see AI Contract Review: What Works; for long files, Summarize Texts and PDFs with AI.
An AI employee has its own login with its own permissions: what it must never do, the system blocks. In a law firm, as an example, that could be a login that files drafts but cannot send anything to a court or client. How I introduce such roles is in Hiring an AI Employee: The Process.
Hallucinations: why every citation gets checked
The BRAK guidance on AI (as of December 2024, still listed as the current version on the BRAK website on November 1, 2026) describes hallucinations as content that seems logical but is factually wrong, and requires the lawyer's own final check of AI results, based on the duty of conscientious practice in § 43 of the Federal Lawyers' Act (Bundesrechtsanwaltsordnung, BRAO).
In an order of 2 July 2025 (case no. 312 F 130/25), the Local Court of Cologne found that requirements set out in a brief did not come from the decision cited and had "apparently been generated by artificial intelligence and freely invented" (my translation). The citations given were invented as well. The court pointed out to the lawyer that knowingly spreading untruths breaches § 43a(3) BRAO.
Three rules help:
- Provide the material instead of asking. The AI quotes word for word from the file, ruling or contract you give it, not from memory.
- Open every citation in the original. Statute, ruling, commentary: what you cannot find gets removed.
- Review separately. My house rule: whoever produced something does not check it.
More in AI Hallucinations: Why AI Makes Things Up and, for research tasks, in Deep Research: Which AI Researches Best. How a second AI checks the first is in LLM as a Judge: How AI Checks AI Work.
Confidentiality in Germany: § 43a BRAO and § 203 StGB
Under § 43a(2) BRAO, a lawyer in Germany is bound to confidentiality, and the duty covers everything that became known to them in the course of their profession. They must bind their staff to confidentiality in text form. § 203(1) no. 3 of the German Criminal Code (Strafgesetzbuch, StGB) makes unauthorized disclosure a criminal offense. Under § 203(3) sentence 2, secrets may go to other people who contribute to the lawyer's work, but only as far as necessary. Under § 203(4) sentence 2 no. 1, a lawyer who failed to bind such a person to secrecy can be liable if that person discloses the secret.
The professional code (Berufsordnung, BORA) requires in § 2(2) organizational and technical measures that are appropriate to the risk and reasonable for the profession. The BRAK guidance adds two points: even client names and the mere fact that a mandate exists are confidential. And removing names and addresses is usually not enough if the context reveals the mandate. Where possible, it recommends "abstract" prompts.
Technically, I place the protection in two spots: user permissions and a principle I call a data lock. Data is pseudonymized before it reaches the model, and the mapping back stays in-house, see Claude Code Mods for Tax Firms: Uses, Risks. That does not automatically cover the BRAK's point about context, and whether it suffices for a mandate is a legal question.
AI providers as service providers: § 43e BRAO
An AI provider that processes client data is a service provider under § 43e BRAO:
- Necessity (para. 1): access to confidential facts only as far as the service requires.
- Selection (para. 2): choose the provider carefully and end the cooperation without delay if the requirements are not met.
- Contract (para. 3): in text form, binding the provider to confidentiality with a notice of the criminal consequences, limiting its knowledge to what the contract requires, and stating whether it may bring in further people.
- Abroad (para. 4): only if secrecy protection there is comparable to protection in Germany, unless the protection of secrets does not require it.
- Single mandate (para. 5): if the service directly serves one individual mandate, the client must consent.
- Data protection (para. 8): remains unaffected, so the GDPR data processing agreement comes on top.
The BRAK adds that the mere possibility of access counts as access. For language models, it considered transmitting client secrets unnecessary as of December 2024, at least for freely available models, since they work without that data, and advises preferring providers with servers in Germany or Europe.
My view as an operator: a GDPR data processing agreement and a contract under § 43e(3) BRAO are different things. Whether a provider's standard terms contain the three points from para. 3 is in their wording; check that with your bar or a specialist before client data flows.
Tool options: business plan, EU cloud or local
Business plan with a data processing agreement. A free personal account is the wrong place for client data. At Anthropic, the data processing addendum covers commercial products such as Claude for Work and the API, whose inputs are not used for training by default. What applies per Claude plan is in Claude Privacy and GDPR: What Applies. For Microsoft 365, see Microsoft Copilot and GDPR at Work. How I decide what an AI may see is in AI and Privacy: What the AI Gets to See.
EU cloud. For processing in the EU, Claude runs through AWS Bedrock in EU regions, see Claude on AWS Bedrock: EU, GDPR and Cost. Providers based in Germany and the EU are covered in European AI Providers Beyond ChatGPT.
Local. Open models on your own hardware keep data in-house but take effort, see Local AI and Privacy: Cloud or Server?. What a private language model for an organization looks like is in Private LLM: Vendor, EU Cloud or Own Server?.
Legal services, client contact and the EU AI Act
For automated client communication, auto-responders and intake chatbots, the BRAK applies a higher standard of care, and there must be no impression that a lawyer's review is unnecessary. With heavy automation, it also sees tax and liability insurance risks. A non-lawyer who builds a tool that legally assesses individual cases runs into the German Legal Services Act (Rechtsdienstleistungsgesetz, RDG): under § 2(1) RDG, a legal service is any activity in specific matters of others that requires a legal examination of the individual case.
As of December 2024, the BRAK sees no general professional duty to tell clients about AI, but recommends transparency and, where in doubt, a contractual arrangement.
Under the EU AI Act, a firm that uses AI is a deployer, so Article 4 on AI literacy applies, newly worded since July 2026 (the BRAK guidance quotes the old version), see AI Literacy Under the AI Act: Training Staff. According to the BRAK, AI systems used in law firms are not among the high-risk systems for the administration of justice in Annex III no. 8(a), because those refer to judicial authorities. Deadlines and duties at a glance are in EU AI Act: Overview and Deadlines 2026.
How a law firm gets started
- Write a firm rule on which tool is allowed for which data, see Shadow AI: When Your Team Uses AI in Secret and AI Usage Policy for Small Teams.
- Start without client data. Abstract legal questions, template letters, your own texts, outlines.
- Check the tool against § 43e BRAO before client data flows, and bind your team to confidentiality in text form under § 43a(2) BRAO.
- Pick one process that comes up daily, follows clear rules and is quick to check. At the tax firm, the assessments came first: around 45 rules, sign-off after four clean runs.
- Define a reviewer role. Whoever writes the draft does not review it.
What a firm rule from step 1 can look like is shown in this hypothetical example. It does not replace a case-by-case review and is not legal advice.
Frequently asked questions
Can lawyers in Germany use ChatGPT or Claude?
Professional law does not ban AI. The limits come from § 43a(2) BRAO, § 203 StGB and § 43e BRAO. Questions without any link to a mandate are a different case from uploaded files. This is a pointer, not legal advice.
Can I enter client data into an AI tool?
Only if § 43e BRAO is met and data protection is settled. The BRAK recommends abstract prompts and fully anonymized documents.
Do I have to tell clients I use AI?
The BRAK sees no general professional duty (as of December 2024) but recommends transparency. If a service provider directly serves one individual mandate, § 43e(5) BRAO requires the client's consent.
Will AI replace lawyers?
No. According to the BRAK, AI supports a lawyer's work, and the final check stays with the lawyer.
Is AI in a law firm high-risk AI under the AI Act?
In the BRAK's assessment, usually not; that category refers to judicial authorities. Article 4 on AI literacy applies anyway.
What to do next
For one week, note which groundwork in your firm always runs the same way. Then check with your bar or a specialist which tool meets § 43e BRAO. We discuss how others train their first AI employees in my community AI Employee Framework. If you want to build an AI employee for your firm, get in touch.
Your first AI employee up and running within an hour
Join the community$5/month founding price, first 50 members
