The Question No One Asked Sophie
When Sophie Chen pastes a case chronology brief into GPT-4o, she is doing three things simultaneously: using an AI system, processing personal data under UK GDPR, and potentially transmitting privileged communications to an external processor. Three obligations attach to a single action — and at baseline, Halstead & Cole has documented none of them.
"Solicitors must consider whether using AI tools is consistent with their duty to keep client information confidential, including where data is processed by third-party AI providers."
— SRA · Compliance Tips for Solicitors Regarding the Use of AI and Technology · 9 February 2026The confidentiality and data protection double exposure
Law firms face two simultaneous exposures when client data enters an AI system without a governance framework. First — SRA Principle 6 (client confidentiality): submitting privileged data to an AI vendor's server may constitute disclosure, depending on the vendor's data terms. Second — UK GDPR Article 6 (lawful basis): the firm must identify a legal basis for every processing activity. "We needed to draft the chronology" is not a lawful basis. Without a documented data governance register, the firm cannot demonstrate to the ICO or the SRA that either obligation was met.
Reading Clause 8.3 and UK GDPR Together
The firm does not need two separate registers. It needs one integrated register that documents each AI system's data flows in terms that satisfy both ISO 42001 Clause 8.3 (data quality and provenance) and UK GDPR (lawful basis, DPA, overseas transfer, special-category safeguards). The frameworks complement rather than duplicate each other — Clause 8.3 refers to UK GDPR rather than recreating its obligations.
Plain English summary
ISO 42001 Clause 8.3 is the "data quality and fitness for purpose" clause — it asks whether the data entering each AI system is appropriate for what the system does, whether its provenance is documented, and whether access and use are controlled. UK GDPR is the "lawful basis and protection" regime — it asks whether processing is lawful, whether the data subject is informed, whether the vendor has a DPA, and whether international transfers are covered. A single register documents both — which is what makes the response to Pemberton Capital Question 5 definitive rather than aspirational.
- Documented data provenance for each AI system — where does the data come from?
- Data quality assessment — is this data fit for the AI system's intended use?
- Data access controls — who can submit data to each AI system?
- Data lineage — can the firm trace what data was used to produce a given output?
- Privacy and security controls consistent with legal obligations
- Retention and deletion procedures for AI-processed data
- Lawful basis identified for each processing activity (Article 6)
- Special-category data: explicit consent or Schedule 1 DPA 2018 condition (Article 9)
- Data Processing Agreement with AI vendor if they are a processor (Article 28)
- International transfer mechanism if data leaves the UK (Articles 44–49)
- Privacy notice updated to reflect AI processing
- DPIA mandatory for high-risk processing
Lawful basis for AI processing of client data
The most available lawful bases for AI processing of client data in legal practice are Article 6(1)(b) — processing necessary for performance of the contract with the client — and Article 6(1)(f) — legitimate interests, where the firm's interest in using AI to provide efficient legal services is not overridden by the client's interests. Article 6(1)(a) (consent) is generally unsuitable as a primary basis for service delivery. For special-category data (health, race, religion, sexual orientation, criminal convictions), the firm requires a Schedule 1 DPA 2018 condition in addition to an Article 6 basis — most commonly paragraph 10 (legal claims) or paragraph 6 (health professional processing, where relevant). A firm that cannot identify a lawful basis for each data category it processes through an AI system is processing that data unlawfully.
Where Law Firm AI Governance Fails
Law firm AI deployments typically involve four categories of data — each with distinct governance requirements. Most firms have no documented framework for any of them in the AI context. The table below shows the categories, their governance requirements, and the failure mode that typically applies at baseline.
| Category | Examples in legal practice | ISO 42001 Cl. 8.3 obligation | UK GDPR additional requirement | Common failure |
|---|---|---|---|---|
| Special Category | Health data in personal injury or clinical negligence; immigration status; criminal convictions in defence; religious beliefs in employment | Highest data quality and access controls; privacy impact addressed in impact assessment | Explicit consent or Schedule 1 DPA 2018 condition required in addition to Art. 6 basis; DPIA mandatory for systematic processing | Firm uses GPT-4o to draft a medical chronology — no Schedule 1 condition documented and no DPIA conducted |
| Privileged / Confidential | Counsel's opinions; without-prejudice correspondence; client instructions; strategy documents | Data provenance documented; access restricted to authorised fee earners; vendor data terms reviewed for confidentiality | Not personal data if anonymised — but privilege is a separate legal concept not displaced by GDPR compliance | Privileged documents submitted to AI vendor without checking whether the vendor's terms permit use for model training or logging |
| Personal Data — Standard | Client names, addresses, financial details; counterparty contact details; witness names | Data minimisation for AI use — submit only what the AI system needs for the specific task | Article 6 lawful basis required; privacy notice must cover AI processing; DPA with vendor if they are a processor | Full client file submitted to AI for "summarisation" when only the relevant section was needed — no data minimisation applied |
| Matter Reference / Anonymised | Anonymised case facts; redacted documents; fictional scenarios for AI testing | Provenance documented even for anonymised data; re-identification risk assessed | GDPR does not apply to truly anonymous data — but pseudonymised data (with key held separately) remains personal data | Firm believes redacted documents are "anonymous" but counterparty names, case numbers and dates remain — re-identification possible |
The "anonymised data" trap
Many firms believe that redacting names makes documents anonymous. It does not. A clinical-negligence chronology with the patient's name removed but with dates of treatment, treating hospital, age, and condition type remains strongly re-identifiable. Pseudonymised data — where identifiers are removed but a key exists that could re-identify — remains personal data under UK GDPR. Truly anonymised data is exempt from GDPR; data that is merely de-identified is not. The AI register must record which category each data flow falls into, and what level of identifiability remains after any "redaction" has been applied.
Five Ways an AI Data Governance Process Fails
Most firms do not fail because they ignored data governance — they fail because some framework was already in place (the firm's general ROPA, an M365 DPA, an IT cyber register) without being extended to cover AI-specific processing. Click each non-conformance category to see what an audit would flag.
The firm has GDPR records (the Record of Processing Activities, or ROPA) but has not extended them to cover AI-specific processing. ISO 42001 Clause 8.3's requirements for data provenance, lineage, and quality are entirely absent for AI systems.
Audit finding: "AI processing activities are not documented in the firm's data governance framework. Clause 8.3 obligations unaddressed; UK GDPR Article 30 record incomplete for AI processing."
The ROPA may record "legitimate interests" for general CRM processing but has not been updated to cover AI-specific processing activities. The lawful basis for submitting client data to GPT-4o — for any purpose — has not been identified or recorded. Without a documented lawful basis, the processing is unlawful under UK GDPR Article 6, regardless of any other controls in place.
Audit finding: "No lawful basis recorded for AI processing of client data. Article 6(1) compliance not evidenced."
Some firms have a Microsoft 365 agreement but have not confirmed that it covers Copilot's AI processing specifically. OpenAI enterprise zero-data-retention terms may be in place but not documented in the AIMS governance framework. UK GDPR Article 28 requires a binding contract before personal data is submitted to a processor — this obligation is system-specific and AI-specific, not satisfied by the firm's general commercial terms.
Audit finding: "Vendor data processing terms for AI components not separately documented. Article 28 obligations for AI-specific processing unevidenced."
Fee earners submit entire client files to AI tools when only a specific section was needed. No guidance exists on what data may be submitted to which system, resulting in systemic over-sharing that violates both data minimisation (UK GDPR Article 5(1)(c)) and Clause 8.3's "data access and use" requirement. Every byte of data submitted beyond what is needed for the task increases the regulatory footprint.
Audit finding: "No data minimisation controls applied to AI submissions. Article 5(1)(c) (data minimisation) compliance not evidenced for AI processing."
Personal injury, clinical negligence, immigration, and employment matters routinely involve special-category data. If a fee earner uses GPT-4o to draft a medical chronology, the firm is processing health data through an AI system — an activity that requires both an Article 9 condition and a DPIA, neither of which exists at baseline in the typical firm. The ICO treats this as a Tier 1 complaint risk.
Audit finding: "Special-category data processed via AIMS-SYS-001 without Article 9 condition, Schedule 1 DPA 2018 condition, or DPIA. Article 9 and Article 35 obligations unaddressed."
The Data Flow Map & Why Zero-Retention Isn't Enough
The diagram below maps the data flows for Halstead & Cole's four registered AI systems. Each row shows the data category entering the system, the vendor's processing location, and the governance status at baseline. The final row flags the special-category exposure that applies to personal injury and immigration matters.
Why "zero-retention" isn't enough
OpenAI's enterprise API offers zero-data-retention as a contractual term — the firm's data is not used for model training and is not retained after the API session. This is a meaningful protection. It does not eliminate the UK GDPR obligations. The data is still processed (transmitted, held in memory, used to generate output) — that is processing under UK GDPR Article 4(2). The transfer still occurs. The lawful basis is still required. Zero-retention reduces the retention-related risk and the training-data risk; it does not displace the requirement for a lawful basis, a DPA, and an overseas transfer mechanism. REG-AIMS-DATA-001 documents all of these — not just the retention status.
Building REG-AIMS-DATA-001 in Six Steps
REG-AIMS-DATA-001 is a single controlled document that satisfies ISO 42001 Clause 8.3 and the ICO's AI data governance requirements simultaneously. It is structured as a register of data flows — one entry per AI system per data category combination. Below are the six steps Halstead & Cole takes to build it.
From REG-AIMS-SYS-001, identify every data category each system handles
The register uses four standard categories (Special Category, Privileged/Confidential, Personal Data Standard, and Matter Reference/Anonymised) but allows sub-categories where needed — for example, distinguishing health data from criminal conviction data within the Special Category column, as they have different Schedule 1 conditions.
Document the Article 6 basis — and the Article 9 + Schedule 1 condition for special-category data
For each row, record the Article 6 lawful basis and, where special-category data is involved, the Article 9 condition and the Schedule 1 DPA 2018 paragraph. Where the lawful basis has not yet been determined, the row is marked "PENDING — system not cleared for this data category." A system is not cleared for a data category until a lawful basis has been documented and verified by the COLP.
Confirm vendor DPA status — AI-specific, not generic
For each AI system acting as a data processor, confirm: whether a Data Processing Agreement is in place, what it covers, whether it explicitly addresses AI processing, and whether it includes a Sub-Processor Schedule. Record the DPA date and reference. For systems where the vendor is a data controller in their own right (rare in enterprise contexts but possible), record the basis for transfer and the privacy notice obligations that follow.
Record the Articles 44–49 mechanism for every cross-border data flow
Where an AI system processes data outside the UK (OpenAI operates US servers; Microsoft operates across multiple regions), document the transfer mechanism — UK Adequacy Regulation, International Data Transfer Agreement (IDTA), or binding corporate rules. The zero-data-retention claim made by some vendors does not remove the transfer obligation — the data still crosses borders even if it is not retained.
Define what fee earners may and may not submit — specifically
For each AI system, document what data categories may and may not be submitted. These rules are the operational expression of UK GDPR Article 5(1)(c) within the AI governance framework — and they are also the Clause 8.3 "data access and use" control. They should be specific: "AIMS-SYS-001 — Personal data: submit only the specific passages relevant to the task. Do not submit full client files. Do not submit special-category data without prior COLP approval and a DPIA in place."
Flag combinations that require a DPIA — and complete them before production
Any processing that is "likely to result in a high risk to individuals" requires a Data Protection Impact Assessment before processing begins. In the AI context, this includes: systematic processing of special-category data; automated decision-making with legal or similarly significant effects; large-scale processing of personal data by AI systems. The register flags which system/data combinations require a DPIA; the DPIA itself is completed as a separate document and cross-referenced in the register.
Priority actions on issue of REG-AIMS-DATA-001
The two highest-priority rows at baseline are AIMS-SYS-001 (GPT-4o) processing personal data — the vendor DPA and overseas transfer mechanism must be confirmed and documented within 14 days. The second priority is the blanket prohibition on special-category data in any AI system until a DPIA and Schedule 1 condition are in place — this prohibition must be communicated to all fee earners in writing at the time REG-AIMS-DATA-001 is issued, cross-referenced to POL-AIMS-001.
Module 07 Completion Checklist
Tick every box below before proceeding to Module 08. Module 07 satisfies the dual data-governance framework for all five of Halstead & Cole's registered AI systems — and answers Pemberton Capital's fifth and final questionnaire question with documented evidence.
Module 08 — Preview
Module 08 closes the operational gap. With the policy (G1), the AI system register (G3), the risk register (G4), and the data governance register (G5) all in place, the final module addresses the human oversight procedure that makes the entire framework work in practice. PROC-AIMS-HITL-001 defines what review fee earners must conduct before AI output is used in client work, how the review is recorded, how sampling audits are conducted for lower-tier systems, and what happens when a review identifies an error. By the end of Module 08, Halstead & Cole will have a complete, defensible AI governance framework — and a full set of answers to Pemberton Capital's panel questionnaire.