Data Minimization Playbook for Legal AI

Metric 0 Pre-Check

Complete all five gates before beginning any data minimization programme or assessment:

• Gate M0.1 — GOV-02 verified: AI Use Policy is current and in force; data minimization obligations aligned with approved AI usage categories.
• Gate M0.2 — AI BoM verified: AI Bill of Materials entries exist for all AI tools involved in data collection, processing, or storage activities covered by this playbook.
• Gate M0.3 — DAT-01 verified: Data Governance Framework is in force; this playbook implements DAT-01’s minimization governance rules.
• Gate M0.4 — DAT-02 verified: Data Inventory and Classification Matrix is current; data levels (Public/Internal/Client Confidential/Privileged) inform minimization requirements throughout.
• Gate M0.5 — STR-07 verified: AI Task Force Charter is active; escalation channels open for Class 6 Shadow AI incidents identified during minimization assessment.

All five gates must be confirmed before minimization programme work begins.

---

1. Purpose, Scope, and When to Use

Purpose. Provide structured guidance for applying data minimization principles across legal AI initiatives, aligning with GDPR, ABA ethics rules, EU AI Act, US state privacy laws, and Risk Taxonomy 2026.

Scope. All AI-related data collection, processing, storage, sharing, and disposal activities within the legal function, including vendor tools and internal builds.

When to Use.

• Blueprint stage: Pillar 2 — Data and Infrastructure.
• During AI vendor selection and contracting.
• When designing or updating data governance and retention.
• During compliance monitoring, audits, and incident response.

---

2. Regulatory Framework Compliance

2.1 GDPR (Global Applicability)

• Article 5(1)©: Personal data must be adequate, relevant, and limited to what is necessary.
• Implementation: Purpose-limitation analysis before collection; documented justification per data element; automated retention and deletion.
• Enforcement: Fines up to 4% of global turnover or €20m; DPA investigations.
• Risk Taxonomy 2026: Primarily Class 4 (Privacy/Data Protection); Class 7 (Regulatory Compliance Drift) when guidance changes.

2.2 ABA Model Rules (US Legal Profession)

• Rule 1.6: Protect confidential client information; obtain informed consent before using client data in AI systems; maintain reasonable safeguards.
• Rule 1.1: Competence includes understanding AI capabilities, limitations, and data protection.
• Rule 5.3: Lawyers remain responsible for AI-assisted work.
• Risk Mapping: Class 2 (Privilege/Confidentiality) and Class 6 (Shadow AI) when unapproved tools process client data.

2.3 EU AI Act

• Article 10: Data and data governance obligations for high-risk AI systems.
• Recital 69: Data minimization applies throughout the AI lifecycle.
• Risk Mapping: Class 3 (Bias/Fairness), Class 7 (Regulatory Compliance Drift), Class 9 (Operational Resilience).

2.4 US State Privacy Laws

• CCPA/CPRA, VCDPA, Colorado Privacy Act: Necessity and proportionality for collection, use, retention, and sharing; enhanced rules for sensitive data and automated decision-making.
• Risk Mapping: Class 4 (Privacy/Data Protection), Class 3 (Bias/Fairness) for impact assessments and audits.

---

3. Data Lifecycle Minimization Framework

3.1 Stage 1 — Data Collection

• Conduct a necessity assessment for each data element: purpose, alternatives, and minimum required.
• Limit collection to data directly relevant to the current legal matter; avoid “just in case” collection.
• Use standardised intake forms with only essential fields and progressive collection as matters evolve.
• Obtain explicit client consent for each data category where required.
• AI vendor pre-check: Confirm AI BoM registration, identify minimum data needed, consider anonymisation/synthetic data, and document necessity.

3.2 Stage 2 — Data Processing

• Process data only for the original purpose or clearly compatible secondary purposes.
• Configure AI systems to process only necessary data; implement filters and validation.
• Maintain human oversight for significant processing decisions and AI-assisted analysis.
• Implement bias detection and mitigation for training and operational data (Class 3 monitoring).

3.3 Stage 3 — Data Storage

• Apply storage minimization: regular necessity reviews and automated retention policies.
• Define maximum retention periods per DAT-02 level; implement automated deletion and quarterly reviews.
• Enforce encryption, access controls, audit logs, and matter-based segregation.
• Ensure backups follow the same minimization and deletion rules.

3.4 Stage 4 — Data Sharing

• Evaluate and document necessity and proportionality for each sharing arrangement.
• For AI vendors: share only minimum data, prohibit training on client data, require segregation, and enforce contractual controls on retention and deletion.
• Put DPAs in place with explicit minimization, security, and audit rights.
• For cross-border transfers, use adequacy decisions or SCCs and conduct transfer impact assessments.

3.5 Stage 5 — Data Disposal

• Review data against retention schedules; identify items eligible for disposal, subject to legal holds.
• Use secure deletion methods (cryptographic deletion, multi-pass overwriting, physical destruction where needed).
• Require vendors to delete client data at termination and provide certificates of destruction.
• Maintain detailed disposal logs for DPS Defensibility.

---

4. AI-Specific Data Minimization

4.1 Training Data Minimization

• Prohibit vendors from using client data for model training via contract and technical controls.
• Prefer public, synthetic, or anonymised data for training legal AI models.
• Regularly assess training data quality and bias; remove outdated or irrelevant data.
• Set retention limits for training data and document disposal decisions.

4.2 Operational Data Minimization

• Filter inputs to include only data necessary for the specific legal task; apply redaction and masking.
• Manage context windows to minimise exposure and persistence across sessions.
• Use prompt engineering to request only necessary information.
• Filter outputs to remove unnecessary or sensitive information; apply human review and retention limits.

4.3 Class 6 Shadow AI — Escalation Protocol

• Scope: Any unapproved AI tool processing legal data without AI BoM registration, a matching DPA, or attorney authorisation for Level 3–4 data.
• Primary risk: Uncontrolled data proliferation across all lifecycle stages.
• Severity tiers: Critical (Level 4 data), High (Level 3), Medium (Level 2), Low (scope drift in registered tools) with defined stop, notify, and logging actions.
• Detection: Network monitoring, DLP, staff surveys/self-reporting, and AI BoM checks.
• Prevention: Approved alternatives, web filtering, clear communication, and TAL-02-based literacy.
• Post-incident: Exposure assessment, privilege review, notification assessment, vendor assessment, AI BoM update, GOV-03 closure, and targeted training.

4.4 Agentic Tier Data Minimization

For Agentic Tier (Level 4 — AI as Executor) tools:

1. Minimal Data Scope per Task: Task-scoped, time-limited access only.
2. Automatic Data Expiry: Discard working data after task completion unless explicitly authorised.
3. Audit Logging with Privilege Protection: Log all access in privilege-protected records.
4. Kill-Switch with Data Halt: Immediate halt of all data access and queued processing.
5. No Cross-Matter Data Access: Strict cross-matter isolation.

If any provision cannot be confirmed, the tool may not process Level 3 or Level 4 data; General Counsel approval and STR-07 notification are required.

---

5. Role-Specific Implementation Guidance