Regulation & Ethics
GDPR and AI Tools for Lawyers: Lawful Basis, DPIAs and Vendor Terms
How GDPR applies when law firms use Claude, ChatGPT or Copilot: controller and processor roles, lawful basis, DPIA triggers, transfers and vendor terms.
On this page
- Controller, processor or neither: the tier decides, not the brand
- Lawful basis: why "the client agreed" is the wrong answer
- The upstream problem: EDPB Opinion 28/2024 and the Hamburg paper
- When a DPIA is required and what it should say
- Transfers, residency and what the Garante case teaches
- The vendor-term checklist
- Professional secrecy is the stricter layer
- What to do this week
A law firm that types client or counterparty data into Claude, ChatGPT, Gemini or Copilot is a GDPR controller and the vendor is its processor, but only on tiers with an Article 28 agreement, no training on inputs and defined retention. Consumer tiers fail that test. Below: roles by tier, lawful basis, when a DPIA is required, transfers and a vendor-term checklist.
Controller, processor or neither: the tier decides, not the brand
The question is never whether ChatGPT or Claude "is GDPR compliant". It is in what capacity the vendor processes the personal data you send it. For client and counterparty data the firm is the controller: it decides the purpose (advising the client) and the means (a particular tool). The vendor is a processor under Article 28 only where a contract binds it to act on the firm's documented instructions, keep the data confidential, delete or return it at the end, and use it for nothing else. Where the vendor reserves the right to train on inputs or to keep them for years, it processes for purposes of its own. The firm has then disclosed personal data to an independent controller and will struggle to name a lawful basis for having done so.
As of September 2026 the tiers sort as follows.
| Tier | Training on your inputs | Retention | Client data? |
|---|---|---|---|
| Claude Free, Pro, Max | User must choose (policy since 28 Aug 2025) | Opted-in chats 5 years; opted-out 30 days | No |
| Claude Team, Enterprise, API, Bedrock, Vertex | Never | API content not retained by default; Fable 5/5.1 and Mythos require 30-day retention and are not ZDR-eligible without express authorisation; Team and Enterprise chat not ZDR-eligible; flagged content up to 2 years; Compliance API feed and local session transcripts 6 years by default | Yes, once retention is configured |
| ChatGPT Free, Plus, Pro | Yes, unless the user opts out | Vendor-set | No |
| ChatGPT Business, Enterprise, Edu; OpenAI API | No by default | API inputs 30 days for abuse monitoring; zero data retention by contract | Yes |
| Gemini consumer app | Yes unless "Keep Activity" is off; a subset is read by human reviewers | 18 months by default; reviewed chats up to 3 years | No |
| Gemini for Workspace, Gemini Enterprise | Add-on required; enterprise terms | EU data residency for NotebookLM Enterprise since 3 Jun 2026 | Yes, with residency configured |
| Microsoft 365 Copilot | Enterprise terms | Enterprise terms; the Legal Agent for Word requires Anthropic enabled as a sub-processor (Apr 2026) | Yes, after the sub-processor list is reviewed |
Sources: Anthropic's data-retention documentation, OpenAI's enterprise privacy page and Google's Gemini Apps privacy notice, whose wording settles the argument for anyone still using a personal account:
"Please don't enter confidential information that you wouldn't want a reviewer to see." — Google, Gemini Apps Privacy Hub
Personal accounts, including the personal-only Google AI Pro and Ultra plans, are out for client work regardless of how good the model is. The firm also needs the privacy settings on each approved tool locked at admin level, because Thomson Reuters' Future of Professionals 2026 survey (June 2026) found 34% of respondents using unsanctioned "shadow AI". The short answer on ChatGPT and GDPR is the version to forward to colleagues.
Lawful basis: why "the client agreed" is the wrong answer
Consent is the basis lawyers reach for first and usually the worst available. It can be withdrawn at any time under Article 7(3), awkward halfway through a matter, and the file is full of people who never gave it: counterparties, witnesses, employees, beneficiaries, the other spouse. The workable structure is Article 6(1)(b), performance of the contract, for the client's own data, and Article 6(1)(f), legitimate interests, for everyone else's.
Legitimate interests requires the three-step test the EDPB restated in Opinion 28/2024: name the interest (delivering the engagement efficiently and accurately), show that the processing is necessary for it, and balance it against the data subjects' rights. Necessity is where AI deployments are weakest. If the task can be done with a pseudonymised extract, sending the whole file is not necessary. That is also the point of permission-bound DMS connectors such as the iManage MCP server (14 May 2026), which respects ethical walls and logs every call.
Special-category data under Article 9 (health, sexual life, trade-union membership, religion) and criminal-offence data under Article 10 are routine in personal-injury, family, employment and white-collar work. Article 9(2)(f), processing necessary for the establishment, exercise or defence of legal claims, covers contentious matters. It does not stretch to a marketing analysis of a client's customer database, and it does not remove the need for a DPIA.
Articles 13 and 14 mean the firm's privacy notice should name AI vendors among its processors, and the CCBE's guide on generative AI (2 October 2025) adds the professional layer: inform clients and give them the chance to object. That overlaps with the disclosure question but is not the same thing.
The upstream problem: EDPB Opinion 28/2024 and the Hamburg paper
The EDPB's Opinion 28/2024, adopted on 17 December 2024, set three positions that still govern the analysis (a summary is here). A trained model is anonymous only if the likelihood of extracting personal data from it is "insignificant". Training on personal data can rest on legitimate interests, subject to the three-step test. And a deployer of a model that was trained unlawfully must assess that upstream unlawfulness before relying on the model. That third point is the one law firms skip, and it makes how the model was trained a due-diligence question. Whether the vendor signed the EU's GPAI Code of Practice is a reasonable proxy: OpenAI, Google, Anthropic, Microsoft, Mistral, Amazon, IBM, Cohere and Aleph Alpha signed on 10 July 2025, xAI signed only the safety chapter, and Meta refused.
The Hamburg data protection authority's discussion paper of 15 July 2024 (PDF) takes a more relaxed view of the model itself: an LLM does not "store" personal data in the GDPR sense, so data-subject rights do not run against the weights. They run against inputs and outputs at the deployer's level. A subject-access request reaches every prompt and output the firm retains, including a six-year Compliance API feed and the local session transcripts that Claude Code and Cowork keep by default. Those logs need a retention schedule like any other matter record.
When a DPIA is required and what it should say
Article 35 requires a data protection impact assessment where processing is "likely to result in a high risk". The EDPB's criteria include innovative technology, sensitive data, vulnerable data subjects and large-scale processing; two or more usually means a DPIA. A generative-AI deployment is innovative technology by definition, and any litigation, employment or family file adds sensitive data. Firm-wide rollouts therefore need one. So does any agent that reads the DMS, drafts emails or files documents, because the data an agent moves can exceed its instruction: in July 2026 xAI's Grok Build was shown uploading 5.10 GiB of repositories where roughly 192 KB was used, including a secrets file users had told it to ignore.
A DPIA for an AI tool does not need forty pages; it needs these eight sections, answered honestly.
- Processing described. Tool, tier, data categories, practice groups, whether outputs are stored, whether agents act autonomously.
- Roles. Firm as controller; vendor as processor; sub-processors listed (for the Word Legal Agent that includes Anthropic); transfer route.
- Necessity and proportionality. Lawful basis per data category; why a pseudonymised extract would not do; retention period; how data-subject rights will be honoured against prompt logs.
- Risks to data subjects. Retention beyond the matter; training on inputs; unauthorised access at the vendor; hallucinated personal data in outputs; agent over-reach.
- Measures. DPA and no-training clause; admin-locked settings; zero data retention or the shortest available window; EU region; DMS connector with permission inheritance; human review before any output leaves the firm; approval gates for agents.
- Residual risk. What remains, and whether it triggers prior consultation with the supervisory authority under Article 36 (in Austria, the Datenschutzbehörde).
- Sign-off. Data protection officer's opinion, the partner responsible, the date, and a review trigger (new model, tier or connector).
- Register entry. Update the Article 30 record of processing activities.
The EU AI Act sits alongside this: the Article 4 literacy duty has applied since 2 February 2025, and the AI Act guide for lawyers covers what a deployer owes. The DPIA and the AI Act documentation can share most of their content.
Transfers, residency and what the Garante case teaches
Anthropic, OpenAI, Google and Microsoft are US-headquartered, so Chapter V applies to almost every deployment. The vendor DPAs carry standard contractual clauses and, where the vendor is certified, rely on the EU–US Data Privacy Framework; the transfer impact assessment is the firm's to write. Residency reduces the exposure without removing it. Enterprise settings and hyperscaler regions can keep data in the EU, and Mistral markets itself as EU-domiciled. Mistral's own argument against residency toggles applies to its US competitors as much as to anyone:
a "data-residency toggle changes where data sits — not which government can compel its production" — as summarised in a review of Mistral's sovereignty positioning
The Austrian bar's first practice rule, that client data stays on EU servers, makes residency a professional requirement in Austria whatever the GDPR analysis says. Few firms need to go as far as Latham & Watkins, which in September 2026 bought its own Nvidia H200 servers to fine-tune open-weight models in-house; an EU region plus a connector that leaves documents in the DMS gets most of the way.
Enforcement against vendors is slower and more fragile than the headlines suggest. The Italian Garante fined OpenAI €15m in December 2024. On 18 March 2026 the Court of Rome annulled the decision: once the Irish DPC became lead supervisory authority on 15 February 2024, the one-stop-shop stripped the Garante of competence (case report). The European Law Blog calls the result an "establish, then escape" enforcement gap. None of that helps a law firm. The firm is a controller supervised by its own national authority, and a complaint from a counterparty who found their medical history in a prompt log will land in Vienna, Munich or London, not Dublin.
The vendor-term checklist
Before a tool is approved for client data, someone should tick every line below with a page reference to the contract. The vendor due-diligence guide has the longer form.
- Article 28 data processing agreement signed, not merely linked in a terms page
- No training on inputs or outputs, stated without a "may" or an opt-out that resets
- Retention window stated in days; zero data retention where available, and confirmation of which models are excluded from it (for Anthropic, Fable 5/5.1 and Mythos)
- Abuse-monitoring and "flagged content" retention disclosed (Anthropic: up to 2 years; OpenAI API: 30 days)
- Sub-processor list with a change-notification right; look for model providers hidden inside productivity suites
- Processing region selectable and EU available; transfer mechanism named
- Customer-held encryption keys where offered (Lexis+ with Protégé introduced them in May 2026)
- Admin controls: training and memory switches locked, connector allow-lists, browsing-agent site restrictions
- Audit logs exportable, and their own retention understood (Anthropic's Compliance API defaults to 6 years)
- Deletion on termination within a stated period, with certification
- Vendor's position on lawful training and on the GPAI Code of Practice
Professional secrecy is the stricter layer
A tool can pass the GDPR test and still be off-limits. The German federal bar's guidance of December 2024 (BRAK, PDF) states that a tool which stores or forwards client data for model training breaches § 43a Abs. 2 BRAO; no balancing test rescues it. The Austrian bar's four rules from its first KI-Tag on 19 March 2026 (EU servers, disclose AI use to clients, the lawyer keeps final responsibility, no client data for training) rest on § 9 RAO's duty of confidentiality and reference § 40 Abs 3 RL-BA 2015. In England, the Upper Tribunal held in UK v SSHD [2026] UKUT 81 (IAC) that pasting client letters into public AI tools puts them "in the public domain", and the SRA's warning notice of 17 August 2026 adopted the point. The Austrian and German bar guidance has its own page, and the privilege analysis explains what the US courts have done with the same facts.
Dr. Niklas Schmidt, a partner at Wolf Theiss in Vienna admitted as both Rechtsanwalt and Steuerberater, teaches the AI Lab for Lawyers course, whose sixth module on risks and regulation covers the ground this page maps.
What to do this week
- List every AI tool in use, including personal accounts, and mark which tier each one is on; anything in the "No" rows of the table above stops handling client data today.
- Pull the DPA for each approved tool and run the checklist; note where retention is set and who can change it.
- Draft the DPIA using the eight sections above, one per firm-wide tool, with a separate entry for any agent with write access to the DMS or email.
- Update the privacy notice and the Article 30 register to name the vendors.
- Put your bar's professional-secrecy rules into the firm AI policy as bright lines, not guidance; the confidentiality checklist is the starting point.
Frequently asked questions
Is ChatGPT GDPR compliant for a law firm?
The business and enterprise tiers can be, because they come with a data processing agreement, no training on inputs by default and contractual zero data retention. The consumer tiers (Free, Plus, Pro) are not usable for client data, because OpenAI trains on conversations unless the user opts out and the firm has no processor contract. The same split applies to Claude and Gemini.
Do I need a DPIA before using generative AI in my law firm?
Almost always. The EDPB's criteria treat innovative technology and sensitive data as high-risk indicators, and two indicators usually trigger Article 35. A firm-wide rollout of any generative-AI tool meets the first, and litigation, employment, family or personal-injury files meet the second. An agent with access to the document management system or email needs its own entry.
What lawful basis does a law firm have for putting client data into an AI tool?
Performance of the contract under Article 6(1)(b) for the client's own data and legitimate interests under Article 6(1)(f) for counterparties, witnesses and other third parties, provided the three-step test is documented. Special-category data in contentious matters can rely on Article 9(2)(f), the legal-claims exemption. Consent is withdrawable and unavailable for third parties, so it rarely works.
Does data residency in the EU solve the transfer problem for AI tools?
It reduces exposure but does not remove it. An EU region keeps data physically in Europe, which matters under Chapter V and under the Austrian bar's EU-server rule, but a US-headquartered vendor can still be compelled to produce data under US law. The firm still needs standard contractual clauses or an adequacy mechanism in the DPA and a transfer impact assessment.
Can the vendor's own GDPR fine protect my firm?
No. Enforcement against vendors is slow and jurisdictionally fragile: the Court of Rome annulled the Garante's €15m fine on OpenAI in March 2026 because the Irish DPC had become lead authority. Your firm is a separate controller supervised by its own national authority, and a complaint about client or counterparty data in a prompt log lands there, not in Dublin.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Questions
Is ChatGPT GDPR-Compliant for Law Firms? A Practical Answer
ChatGPT can be used GDPR-compliantly by a law firm on Business, Enterprise or API terms with training off, a processor contract and a DPIA; consumer plans cannot. Seven steps.
7 min read
Regulation & Ethics
AI Vendor Due Diligence for Law Firms: The Questions to Ask Before You Sign
AI vendor due diligence for law firms starts with the subprocessor chain, then training and retention terms, ZDR eligibility, data residency, audit logs and exit rights.
11 min read
Glossary
Zero Data Retention (ZDR) — Definition for Lawyers
Zero data retention means the AI vendor stores nothing after replying. Which plans offer ZDR, the Covered Models exception, and why it matters for privilege and GDPR.
5 min read
Playbooks
AI Privacy Settings for Lawyers: Step-by-Step in ChatGPT, Claude, Gemini, Copilot and Perplexity
Exact steps to turn off training and check retention in ChatGPT, Claude, Gemini, Copilot and Perplexity as of September 2026, plus admin settings and a log.
10 min read
Guides
Which AI Tools Train on Your Data? A Confidentiality Checklist for Lawyers
Which AI tools train on your data, plan by plan as of September 2026: ChatGPT, Claude, Gemini, Copilot, legal vendors, plus bar rules and a 10-point checklist.
11 min read
Regulation & Ethics
AI Rules for Austrian and German Lawyers: ÖRAK, BRAK and CCBE Guidance
ÖRAK's four AI rules and § 40 Abs 3 RL-BA bind Austrian lawyers; BRAK's § 43a BRAO line binds German ones; CCBE guides cover both. With a bilingual checklist.
11 min read