AI Agents for Lawyers

Practice Areas

AI for Privacy and Data Protection Lawyers: DSARs, DPIAs and Advising on AI

AI for privacy lawyers: automating DSARs and DPIAs, reading vendor retention terms, and advising on LLM rollouts after EDPB Opinion 28/2024 and the Rome ruling.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. The enforcement picture clients ask about
  2. Advising on an LLM rollout: three questions that decide the memo
  3. The DPIA prompt skeleton
  4. DSARs: an agent that respects the clock
  5. Vendor reviews: what the terms say in September 2026
  6. Your own practice is the first client
  7. What to do this week

On 18 March 2026 the Court of Rome annulled the Italian Garante's €15m fine against OpenAI, not because the processing was lawful but because the Irish DPC had become lead authority before the Garante decided. AI for privacy lawyers runs on two tracks: advising clients on LLM deployments under a GDPR whose enforcement map keeps shifting, and automating the DSARs, DPIAs and vendor reviews that make up the day job.

The enforcement picture clients ask about

The client question in 2026 is rarely "is this lawful" in the abstract; it is "who will look, and under what theory". Four documents set the frame as of September 2026.

DocumentDateWhat it holdsWhat it means for a deployment
Hamburg DPA discussion paper on LLMs15 July 2024An LLM does not "store" personal data in the GDPR sense; data-subject rights apply to inputs and outputs at the deployer levelThe deployer, not the model, is where access, rectification and erasure requests are answered
EDPB Opinion 28/202417 December 2024A model is anonymous only if extraction of personal data is "insignificant"; legitimate interest requires the three-step test; deployers of models trained unlawfully must assess that upstream unlawfulnessThe DPIA must cover the provider's training, not only the client's use
Court of Rome, annulment of the Garante's OpenAI fine18 March 2026 (reasons 28 May 2026)The Irish DPC became lead authority on 15 February 2024; the Garante's November 2024 decision came after and lacked competence under the one-stop-shopEnforcement against the main providers runs through Dublin; national DPAs retain deployer-level powers
In re OpenAI Copyright Litigation (S.D.N.Y.)Preservation order 13 May 2025; 20-million-log production affirmed 5 January 2026ChatGPT output logs must be preserved and a de-identified sample produced; users "voluntarily submitted their communications"A US court order can override a provider's deletion promises; retention commitments to data subjects are conditional

The Rome judgment is the one to explain carefully. The Garante had fined OpenAI €15m on 2 November 2024 (announced 20 December) and ordered a six-month broadcast campaign; the court held that once the Irish DPC took over as lead supervisory authority, the Garante could not decide (PPC Land). The European Law Blog calls the result "establish, then escape": a provider that sets up its EU establishment in Ireland moves out of reach of the national regulators that had been most active (European Law Blog). For a deployer client the practical reading is that their own national DPA still supervises their use of the tool, and the EDPB opinion tells that DPA what to ask. The GDPR and AI tools page covers lawful basis in depth; the Norton Rose Fulbright summary of the EDPB opinion is a good client handout (Data Protection Report).

Advising on an LLM rollout: three questions that decide the memo

Most deployment advice reduces to three questions, and the answers depend on the tier of the tool more than on the vendor.

Who is processing what, on which terms. On a business, enterprise or API tier the provider is a processor under an Article 28 agreement, does not train on inputs, and states its retention; on a consumer tier the provider is closer to an independent controller with its own purposes, which is why the German three-lawyer firm in the employment page needed an Article 28 agreement plus a § 43e BRAO contract even for a self-hosted stack. Microsoft's Legal Agent for Word adds Anthropic as a subprocessor of the firm's documents, a fact one commentator noted "firm IT has not formally evaluated". Subprocessor chains are the first thing to map.

Whether the model itself is a problem. The Hamburg and EDPB positions above both end at the same place for a deployer: assume outputs can contain personal data, assume the provider's training is a question you must ask, and get the provider's answer in writing. The GPAI obligations in force since 2 August 2025 require providers to publish a training-content summary, which is the document to request.

What happens to inputs and outputs over time. Retention is where the surprises are. Claude consumer users who opted in to training have their chats kept five years; Gemini consumer activity is kept eighteen months by default with a subset read by human reviewers, and 72 hours even with activity off; OpenAI's API keeps 30 days for abuse monitoring unless zero data retention is contracted; Anthropic's Compliance API keeps session transcripts for six years by default. And the New York Times litigation shows a preservation order can override every one of those numbers. The privilege page covers the discovery side.

The DPIA prompt skeleton

A DPIA for an LLM deployment is a structured document, and a model drafts a competent skeleton from the client's facts, provided it is told what it may not invent. This is the working prompt:

Role: Data protection counsel drafting a DPIA under Article 35 GDPR for
the deployment described in the attached fact sheet (A), the vendor's
terms and DPA (B), the vendor's training-content summary or model card
(C) and the client's records of processing (D). Work ONLY from A–D.
Where a fact is missing, write [ASK CLIENT: ...] and continue.

Produce the DPIA in this structure:
 1. Description of the processing: purposes, data categories, data
    subjects, volumes, retention (from A and D); systematic description
    of the tool, tier, hosting region, subprocessors (from B).
 2. Necessity and proportionality: lawful basis per purpose; if
    legitimate interest, apply the three-step test from EDPB Opinion
    28/2024 with the facts in A; Article 22 analysis if any decision
    is solely automated.
 3. Model-level assessment: what C says about training data and
    personal data; whether the model can be treated as anonymous on
    the EDPB "insignificant extraction" standard; what B says about
    training on the client's inputs and about retention of inputs,
    outputs and logs.
 4. Risks to data subjects: a table with columns Risk | Likelihood |
    Severity | Source (A/B/C/D reference). Include confidentiality of
    inputs, inaccurate outputs about individuals, unauthorised
    retention, subprocessor access, cross-border transfer, and
    preservation orders overriding deletion.
 5. Measures: for each risk, the measure, who owns it, and whether it
    is contractual (B), technical, or organisational.
 6. Residual risk and whether prior consultation under Article 36 is
    indicated.
 7. Open questions for the DPO and the works council, if any.

Do not cite guidance, cases or regulator decisions other than those
quoted in A–D. Do not conclude that the processing is lawful.

The last two lines are the gate. A model asked for a DPIA will otherwise produce confident citations to regulator guidance it has not read, and HAQQ's June 2026 benchmark found 24% of answers across leading models cited or misapplied law that did not support the claim. The lawyer adds the authorities and writes the conclusion. Anthropic's open-source privacy plugin ships a "PIA/DPIA triage" skill on the same logic, alongside DPA review and DSAR response (github.com/anthropics/claude-for-legal); the plugins page explains the cold-start interview that records the firm's positions once.

DSARs: an agent that respects the clock

A data subject access request is an extraction, classification, redaction and drafting problem with a statutory deadline of one month, extendable by two for complex requests. That is an agent's shape of work, and the deadline is the constraint the agent is built around.

  1. Intake and clock. The agent logs the request, the date received, the identity verification status and the deadline, and schedules reminders at 21, 14, 7 and 2 days. A person confirms identity.
  2. Scope. The agent drafts the search parameters (name variants, email addresses, employee ID, date range, systems) from the request and the client's records of processing; the DPO approves them.
  3. Collection. Exports from mail, HR, CRM and chat systems go through an OCR pass and into a review workspace; the PDF extraction playbook covers the pipeline for scanned material.
  4. Classification. Every document is tagged: contains requester's personal data or not; contains third-party data; potentially privileged; potentially within an exemption (management forecasting, negotiations, legal professional privilege), with a reason and a passage reference per tag.
  5. Redaction proposals. The agent proposes redactions for third-party data and privileged content; a reviewer accepts or rejects each. Relativity's aiR for Privilege and the generative review tools covered on the e-discovery page do this at volume, and Relativity reports that 57% of its new workspaces are now non-litigation.
  6. Response drafting. The agent drafts the cover letter, the Article 15 information (purposes, recipients, retention, sources, rights) from the records of processing, and the schedule of documents. A lawyer decides every exemption and signs.
  7. Log. The agent files the decision log: what was searched, what was withheld and why, who decided. The regulator will ask for it.

Steps 1, 5 and 7 are the value: nobody misses the deadline, every redaction is a recorded decision, and the file explains itself. Ashurst's lawyers reportedly used AI to sift about 100,000 documents in 48 hours to decide what to submit to a regulator (an FT report seen only in snippet form; treat as reported), which is the scale at which a manual DSAR process fails.

Vendor reviews: what the terms say in September 2026

A privacy lawyer's vendor review of an AI tool is now a close reading of five things: training, retention, human review, subprocessors and residency. The current positions, as stated in the vendors' own published policies:

Vendor and tierTrains on inputsRetentionNotes
ChatGPT Free/Plus/ProYes unless opted outDeleted chats scheduled for removal within 30 days absent a legal hold; a US preservation order applied from May 2025Consumer terms, no processing agreement
ChatGPT Business/Enterprise, OpenAI APINo by defaultAPI 30 days for abuse monitoring; zero data retention contractualBusiness tier on the same models
Claude Free/Pro/MaxUser must choose (since 28 August 2025)Opted in: 5 years; opted out: 30 daysIncognito chats never used for training
Claude Team/Enterprise, APINoAPI not retained by default, except Fable 5.1, which requires 30 days and is not ZDR-eligible without express authorisation; Team and Enterprise chat not ZDR-eligible; flagged content up to 2 years; Compliance API 6 yearsThe strongest model is the least private (Anthropic retention policy)
Gemini consumerYes unless "Keep Activity" is off18 months by default; 72 hours with activity off; human-reviewed chats up to 3 yearsGoogle: "Please don't enter confidential information that you wouldn't want a reviewer to see" (Google)
Gemini for Workspace, Vertex, Gemini EnterpriseNo by defaultAdmin-controlled; EU data residency for NotebookLM Enterprise since 3 June 2026Requires Workspace add-ons; consumer AI Pro/Ultra plans are personal-account only
Microsoft 365 CopilotNoUnder the Microsoft DPA with Purview controlsConsumer Copilot runs without a DPA and is "not suitable for confidential or legal work"
Perplexity EnterpriseNever by default"AI data retention" toggle on consumer tiers
Meta Muse Spark API"Contributor tier" retains data; "standard tier" does notTier-dependentCheaper tier is the one that retains

Two 2026 incidents belong in every vendor memo. Grok Build's coding agent uploaded entire repositories, including an unredacted secrets file users had told it to ignore, at roughly 27,800 times the data the model used, and the "Improve the model" toggle did nothing until a server-side fix on 13 July 2026 (The Hacker News). A vendor's privacy switch is a representation, not a control, and the DPIA should say so. And the confidential-information warning on Google's own consumer page is the plainest statement any provider has made about what a consumer tier is for. The confidentiality checklist keeps the full table current, and the privacy settings playbook has the click paths for each opt-out.

Your own practice is the first client

Privacy lawyers are held to their own advice. In UK v SSHD [2026] UKUT 81 (IAC) the Upper Tribunal treated client letters pasted into public AI tools as being "in the public domain"; the SRA's warning notice of 17 August 2026 cites that case and counts 42 AI-misuse reports in a year. The BRAK's guidance says tools that store or forward client data for model training breach § 43a Abs. 2 BRAO; ÖRAK's four rules require EU servers and no training on client data; the CCBE's October 2025 guide says to inform clients and let them object; California's proposed comment to Rule 1.6 treats exposing client data to AI as "revealing" it where there is "a material risk of unauthorized access, retention, or use". The Austrian and German guidance page and the ChatGPT GDPR question page apply these to a firm's own stack.

The firm-side to-do list is the client-side one turned inward: an Article 28 agreement with every AI vendor, a per-tool record of tier and retention, the zero data retention arrangement for anything touching special-category data, and an AI Act inventory. Anthropic's AI-governance plugin ships a per-system EU AI Act inventory and an impact-assessment skill for exactly that; the AI Act page sets out which of a firm's tools fall where.

What to do this week

Run the DPIA skeleton on the client deployment you know best and count the [ASK CLIENT] flags; they are the questions the vendor's sales deck did not answer. Move the next DSAR onto the seven-step flow with the clock at step 1, and keep the decision log even if the rest is done by hand. Rewrite the vendor table above for your own firm's tools, tier by tier, and send it to the partner who signs the engagement letters. Module 6 of the AI Lab for Lawyers course walks through the retention terms and opt-out settings in the table, and Module 4 builds the DSAR clock as a scheduled workflow.

Frequently asked questions

Does a large language model contain personal data under the GDPR?

The regulators disagree in emphasis. The Hamburg DPA's July 2024 discussion paper argues an LLM does not store personal data in the GDPR sense and that data-subject rights apply to inputs and outputs at the deployer level. EDPB Opinion 28/2024 of December 2024 says a model can be treated as anonymous only where the likelihood of extracting personal data is insignificant, assessed case by case. Advise on the assumption that the model may contain personal data and document the assessment.

Can AI answer a data subject access request?

It can do most of the work: locate the requester's data across exports, classify each document, propose redactions for third-party data and legal privilege, and draft the response letter. A person decides the exemptions and signs. Anthropic's open-source privacy plugin includes a DSAR skill built around the statutory clock, and e-discovery platforms sell the same workflow. The one-month deadline, extendable by two months for complex requests, is the design constraint.

Is ChatGPT GDPR-compliant for a law firm?

On a business or enterprise tier with a processing agreement, no training on inputs and defined retention, it can be deployed compliantly, subject to a DPIA and the bar rules. On the consumer tiers it trains on conversations unless the user opts out and retains data for abuse monitoring, and a UK tribunal has treated material pasted into a public tool as being in the public domain. The tier, not the product name, determines the answer.

What changed when the Court of Rome annulled the OpenAI fine?

The Garante's €15m fine of November 2024 was annulled on 18 March 2026, with reasons on 28 May 2026, because the Irish Data Protection Commission had become lead authority for OpenAI on 15 February 2024, so the Garante lacked competence under the one-stop-shop mechanism. The substance of the Garante's findings was not decided. Commentators call it an enforcement gap: establish in Ireland, then escape national regulators.

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.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

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.

11 min read

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

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

Tools

Claude for Legal Plugins Explained: The 12 Practice-Area Plugins and 20+ Connectors

Claude for Legal gives paid Claude users 12 open-source practice-area plugins, 150+ skills and 20+ connectors to iManage, NetDocuments and CoCounsel.

8 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

The EU AI Act for Lawyers and Law Firms: What Applies in 2026

The EU AI Act applies to law firms as deployers. Article 4 literacy and Article 50 transparency are live; high-risk duties are deferred to December 2027.

9 min read