You cannot govern what you have not documented
Ask the managing partner of a typical 80-fee-earner law firm to name every AI tool in active use across the practice. In most cases, the answer will be incomplete within thirty seconds. The litigation team uses GPT-4o for case summaries. The property team has Copilot embedded in Word and Outlook. The billing administrator adopted a legal AI tool at a conference and uses it for precedent drafting. The immigration partner integrated LEAP AI into their matter workflow. None were procured through a central governance process. None are documented anywhere.
This is the shadow AI problem. It is not a technology management failure. It is a documented-information failure under ISO/IEC 42001:2023 Clause 7.5. Before a firm can assess AI risk (Gap 4), manage the AI lifecycle (Gap 5), govern AI data (Gap 6), or disclose AI use to clients (Gap 7), it shall first establish — in a controlled document — what AI systems it operates, what they do, and what version of each is in production.
Ayinde v Haringey / Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin) — presided over by Dame Victoria Sharp P — established that 18 AI-generated citations submitted in court proceedings were fabricated. The common thread across AI citation hallucination cases is not reckless lawyers. It is undocumented AI systems operated without version control, without output review gates, and without any record of what the system was asked or what it produced.
A firm that cannot identify which AI system produced a given output, or which version of that system was in production at the time, shall be unable to demonstrate supervisory control under SRA Principle 5 or to establish the chain of accountability required by SRA Code of Conduct for Firms 2019, Rule 8.1(a). An AI system register is the first-line control against this failure mode.
This article explains what ISO/IEC 42001:2023 Clause 7.5 shall require for AI system documentation, why prompt version control is the overlooked element that most register templates omit, and how to build REG-AIMS-SYS-001 — the controlled register that closes Gap 3.
What ISO/IEC 42001:2023 Clause 7.5 requires of an AI system register
The organisation's AI management system shall include documented information required by this document and documented information determined by the organisation as being necessary for the effectiveness of the AI management system.
When creating and updating documented information, the organisation shall ensure appropriate identification and description (e.g. title, date, author, or reference number); format (e.g. language, software version, graphics) and media (e.g. paper, electronic); review and approval for suitability and adequacy.
Documented information required by the AIMS shall be controlled to ensure it is available and suitable for use, where and when it is needed; and it is adequately protected.
Per ISO/IEC Directives Part 2: shall denotes a requirement without discretion; should denotes a recommendation; may denotes a permission. Clause 7.5 is the documented-information clause — the mechanism by which everything the standard requires to be known is turned into something that can be controlled, retrieved, and audited.
Clause 7.5 does not stand alone. Read in conjunction with Clause 6.1 (AI risk assessment), Clause 8.2 (AI system lifecycle), Clause 8.3 (intended use documentation), and Clause 8.4 (supply chain), it establishes that an AI system register is a prerequisite, not an option. A risk assessment cannot be conducted on an AI system that has not been documented. A lifecycle cannot be managed for a system with no register entry. A supply chain cannot be evaluated for an AI tool that does not appear in any controlled record.
The following fields shall be present in a Clause 7.5-compliant AI system register:
Microsoft 365 Copilot is not one AI system — it is a family of AI features embedded across Word, Outlook, Teams, and SharePoint, each with different data access, different model behaviour, and different risk profiles. A single register entry for "M365 Copilot" shall not satisfy Clause 7.5. Each deployment context in which Copilot is used in legal work shall have its own entry, or a master entry with documented sub-entries for each practice area deployment.
The same principle applies to GPT-4o accessed via the OpenAI API versus via ChatGPT Enterprise versus via a third-party legal tool that uses the OpenAI API. Same model, different system boundaries, different risk profiles, different register entries.
In ITIL 4 terms, REG-AIMS-SYS-001 is the AI Management System's Configuration Management Database (CMDB) equivalent. The Service Configuration Management practice requires that all service assets — including AI systems — be registered as configuration items (CIs) with defined attributes: owner, version, state, dependencies, and relationships. The register's lifecycle states (Pre-deployment, Active, Under Review, Decommissioned) map directly to ITIL 4's CI state model, providing the authoritative system-of-record for all AI service asset transitions.
Without this register, AI systems are unowned service assets: deployed without configuration authority, changed without change control, and impossible to audit against the service baseline. The ITIL 4 Change Enablement practice further requires that any change to a CI — including a change to a system prompt — is subject to a change record and approval before deployment to production. The prompt version log in REG-AIMS-SYS-001 Part B is the change record for that control.
Why the SRA's AI guidance makes the register a de facto obligation
The SRA's guidance Compliance Tips for Solicitors Regarding the Use of AI and Technology, updated 9 February 2026, does not use the phrase "AI system register." It does not need to. The guidance's mandatory requirements create the register as a logical necessity.
The SRA requires firms to demonstrate which AI tools are in use and how outputs are supervised. This obligation shall not be met without a register. When the SRA asks a firm in a thematic review or an investigation — "what AI tools were being used on this matter at this time?" — the only defensible answer is a documented one. A COLP who responds with "I believe it was GPT-4o, possibly Copilot" is not demonstrating supervision under SRA Principle 5. They are demonstrating its absence.
"Firms using AI tools remain responsible for the output. Solicitors must consider how they will supervise AI-generated content before it reaches clients."SRA · Compliance Tips for Solicitors Regarding the Use of AI and Technology · 9 February 2026
Supervision of AI output presupposes knowledge of which AI system produced it. That knowledge shall be retrievable at the matter level — not only at the firm level. An AI system register is the instrument that makes this retrieval possible. Without it, any claim of supervisory control is structurally unfounded under SRA Code Rule 8.1(a).
The Pemberton Capital questionnaire reinforces this from the client side. Question 2 of the five-question AI governance questionnaire — following "Who owns AI risk?" — requires: "What AI tools does your firm use in client matters, and how are outputs checked?" Without a register, Halstead & Cole LLP (fictional, SRA number 12345678) cannot answer this question with the specificity a £620,000 per annum panel client has a right to expect.
Four non-conformance patterns — and the prompt control gap most firms miss
The prompt version control problem — in detail
Prompt version control distinguishes firms that have addressed AI governance substantively from those that have adopted a register template without understanding what it is governing. The following illustrates why it is a Clause 7.5 requirement, not an optional enhancement.
Consider Halstead & Cole's litigation team using GPT-4o for case chronology drafting. The fee earner who built the workflow created a system prompt instructing the model to identify key dates, extract party names, and flag missing evidence references. Over six months, different fee earners modified this prompt — adding instructions, removing constraints, changing the output format. None of these changes were documented. The prompt currently in use is the product of seventeen undocumented edits by four different people.
When a client dispute arises about a chronology that missed a critical date, the firm shall be unable to establish:
The v1.2 entry — an unapproved change that removed a critical safeguard — is exactly the event that a prompt version log makes visible and a governance process can prevent. Without the log, the unapproved v1.2 operated for over six weeks with no awareness that a control had been silently removed.
Building REG-AIMS-SYS-001: the two-part AI system register
REG-AIMS-SYS-001 addresses both problems — the missing inventory and the uncontrolled prompts — in a single controlled document with two linked parts. Part A is the AI system inventory; Part B is the prompt version log. Together they constitute the documented information required by Clause 7.5 at the system and operational levels.
-
Discovery sweep before documentation. Before creating the register, conduct a structured discovery of all AI tools in active use. Send a brief questionnaire to all practice group heads: "List every AI tool, plugin, or AI-enabled feature your team uses in legal work." Include browser extensions, embedded Word features, and third-party research tools with AI features. The discovery output — not the final register — is what surfaces shadow AI. At Halstead & Cole, this step reveals two to three tools unknown to the COLP or IT Lead at baseline.
-
Create Part A — system inventory. For each AI system discovered, create one register entry with the 14 mandatory fields defined in REG-AIMS-SYS-001. Mandatory fields include: system ID (e.g. AIMS-SYS-001), system name, vendor, product version or API model version, intended use, practice area scope, data categories processed, risk classification (cross-referenced to REG-AIMS-RISK-001), oversight mechanism, and the RACI-designated owner role. Each entry shall be version-controlled at the register level — changes to a system entry shall create a new version of that row, not overwrite the previous one.
-
Create Part B — prompt version log. For every system where the firm controls the prompt — any system accessed via API or where a system prompt has been configured — create a prompt version log entry. Each version shall capture: the prompt text or a verified hash reference, the date of change, the author, and the approver. The log shall record both the current version and the full change history. The approval column is the control gate: no prompt change shall be considered in-production until it carries an approval record from the designated role in RACI-AIMS-001.
-
Classify each system by risk tier. Assign each system entry a risk tier — High, Medium, or Low — based on the factors in the G4 risk assessment framework (REG-AIMS-RISK-001). High-tier systems — those used in client-facing advice, court submissions, or handling privileged data — shall carry a mandatory human review requirement for every output before use. Medium-tier systems attract supervisor spot-check review. Low-tier systems are subject to periodic audit. The risk tier shall drive the oversight mechanism field in Part A.
-
Issue under document control and set review cadence. REG-AIMS-SYS-001 is a living document. It shall be updated whenever a new AI system is adopted, an existing system changes version, or a prompt is modified. The RACI-designated IT Lead is the register owner (per RACI-AIMS-001). A quarterly review shall be set to verify the register reflects the current state of the firm's AI deployment. Any fee earner who adopts a new AI tool without triggering a register entry and an impact assessment is in breach of the AI Policy (POL-AIMS-001).
Register preview — Halstead & Cole LLP · AIMS-SYS baseline
| ID | System / Tool | Vendor | Model Version | Intended Use | Data Categories | Risk Tier | Oversight | Status |
|---|---|---|---|---|---|---|---|---|
| AIMS-SYS-001 | GPT-4o (API) | OpenAI | gpt-4o-2024-11-20 | Case chronology, correspondence drafting | Personal, Privileged | HIGH | Human review before use | Active |
| AIMS-SYS-002 | M365 Copilot — Word | Microsoft | Copilot for M365 v2 | Document drafting, precedent adaptation | Confidential, Personal | MED | Supervisor spot-check | Active |
| AIMS-SYS-003 | M365 Copilot — Outlook | Microsoft | Copilot for M365 v2 | Email drafting, client correspondence | Personal, Privileged | HIGH | Human review before send | Under Review |
| AIMS-SYS-004 | LEAP AI | LEAP Legal | LEAP AI 2025.3 | Matter management, time recording assist | Personal, Confidential | MED | Periodic audit | Active |
Preview — 9 of 14 fields shown · Full 14-field register + prompt version logs in REG-AIMS-SYS-001 · M365 Copilot split into separate entries by deployment context per Cl. 7.5 system-level documentation requirement
GPT-4o accessed via the OpenAI API (AIMS-SYS-001) has a different risk profile from GPT-4o accessed via ChatGPT.com, even though the underlying model is the same. API access means the firm controls the system prompt, the data sent, and the output handling — all of which shall be documented in Part B. Consumer product access means OpenAI's terms govern data use, and the firm may have reduced control over what is logged and retained. Both shall require separate register entries; the fields for "data processed" and "overseas transfer" differ materially between them.
REG-AIMS-SYS-001: AI System Register & Prompt Version Log
The controlled document that closes Gap 3 is available for download below. It includes Part A (14-field AI system inventory, pre-populated with the Halstead & Cole baseline entries as a worked example), Part B (prompt version log template with approval workflow for each configurable system), a discovery questionnaire for the initial sweep, risk tier classification guidance, and the version control and approval workflow.
Part A: 14-field AI system inventory with risk tier classification · Part B: Prompt version log with approval workflow · Discovery questionnaire · Risk tier guidance · Version control and owner assignment. Editable Word format — conduct discovery sweep, complete Part A, configure Part B for each API-accessed system, and issue under document control.
Priority actions within 14 days of downloading: First, distribute the discovery questionnaire to all practice group heads — this converts shadow AI into documented AI. Second, create Part A entries for every system identified. Third, identify which systems have configurable prompts and initialise Part B entries, recording the current prompt text as v1.0 regardless of actual history (the pre-register period is noted as "prior to AIMS implementation"). Once Part A is complete, the register satisfies the minimum Clause 7.5 requirement; Part B is the control layer that makes the register defensible under scrutiny.
REG-AIMS-SYS-001 shall only be as reliable as the process that keeps it current. The AI Policy (POL-AIMS-001) shall include an explicit prohibition on adopting any new AI tool — including browser extensions, embedded features in existing software, and third-party legal research tools with AI features — without first triggering a register entry and an impact assessment under Gap 4. Without this policy hook, the register will be accurate at issue date and progressively less accurate thereafter.
The RACI-designated IT Lead shall conduct a quarterly audit against the firm's software licence inventory to detect unregistered AI tools. Any unregistered AI system identified in a regulatory inspection or judicial proceeding shall be treated as a Critical nonconformance under the AIMS.
Where Gap 3 sits in the remediation sequence
Gap 3 is classified as High severity — below the Critical gaps (G1 AI Policy, G2 Internal Organisation, G4 Risk Assessment, G6 Data Governance, G7 Client Disclosure) but a hard prerequisite for several of them. A risk assessment shall not be conducted for a system that is not in the register. A data governance control shall not be applied to a data category processed by a system with no register entry. An AI disclosure shall not be made to a client for a system that has not been documented.
The next article in this series covers Gap 4: AI risk and impact assessment under ISO/IEC 42001:2023 Clause 6.1. Once the register is in place, every system in it becomes the subject of a structured risk assessment — the process that determines which controls are proportionate and which AI deployments require a formal impact assessment before continuing in production. Gap 4 — AI Risk & Impact Assessment →
Quality gate record
The following gates were checked before publication of v1.1. All 10 gates pass. This record satisfies ISO 9001:2015 Clause 7.5.3 (control of documented information) for the series and provides an auditable record of pre-publication compliance verification.
- G1Every regulatory claim carries a named, citable source with instrument reference. SRA AI guidance (9 Feb 2026), Ayinde v Haringey [2025] EWHC 1383 (Admin), SRA Code of Conduct for Firms 2019 Rule 8.1(a), and SRA Principle 5 are all cited by name, date, and reference number. No fabricated citations detected.PASS
- G2ISO/IEC 42001:2023 Clause 7.5 is correctly cited and accurately characterised. The three-clause dependency chain (Cl. 7.5 → Cl. 8.3 → Cl. 6.1) is accurately constructed. SHALL/SHOULD/MAY field card matrix is correctly applied per ISO/IEC Directives Part 2. No fabricated or mischaracterised clause references detected.PASS
- G3Scenario internally consistent with REF-AIMS-SCEN-001 v1.0 — firm name (Halstead & Cole LLP), SRA number (12345678), COLP name (Priya Anand), AI tools in use (GPT-4o, Microsoft 365 Copilot — Word, Microsoft 365 Copilot — Outlook, LEAP AI), and Pemberton Capital relationship (£620,000 per annum) all match the series character register. Fee earner reference (S. Chen) is consistent with prior series usage.PASS
- G4Prescriptive language per ISO/IEC Directives Part 2 is used correctly throughout. SHALL used for requirements. SHOULD used for recommendations. SRA regulatory obligations — softened to "expects" in v1.0 — corrected to "requires" and "shall" in v1.1 wherever the source instrument is mandatory. Regulatory posture accurately represented.PASS
- G5REG-AIMS-SYS-001 referenced with ungated download link. No email gate. Cross-references to POL-AIMS-001 (G1), RACI-AIMS-001 (G2), REG-AIMS-RISK-001 (G4), and PROC-AIMS-LIFE-001 (G5) all present with correct document IDs.PASS
- G6Article links to Gap 2 predecessor (RACI-AIMS-001) and Gap 4 successor (REG-AIMS-RISK-001). Gap 5 dependency (PROC-AIMS-LIFE-001) explicitly documented in remediation sequence. Cross-reference to G6 data governance and G7 client disclosure present in dependency analysis. Series navigation present in breadcrumb.PASS
- G7Limitations disclaimer present in the standalone disclaimer section. Article published for educational purposes only and does not constitute legal advice. Fictional scenario notice confirmed for all named parties: Halstead & Cole LLP, Priya Anand, Sophie Chen, Pemberton Capital LLP.PASS
- G8Not applicable — this article does not contain a policy or procedure template requiring a document-control block. REG-AIMS-SYS-001 document-control block is contained within the downloadable template, not reproduced in the article body.PASS
- G9H1 title includes target keywords "Shadow AI," "ISO 42001," and "AI System Register." Stat bar appears within first screenful of content. Structured data (Schema.org Article with isPartOf series) present in document head. Canonical URL set.PASS
- G10IEC 82079-1:2012 notice hierarchy fully corrected from v1.0. v1.0 erroneously used DANGER (reserved for immediate physical harm) for a regulatory enforcement consequence. v1.1 reclassifies as WARNING (risk of serious harm if action not taken) with two-paragraph structure: (1) factual judicial citation, (2) SRA regulatory consequence. CAUTION used for register decay and operational risk. NOTE used for informational API-vs-consumer-product guidance. ITIL 4 NOTE added and labelled distinctly as a non-IEC extension with indigo styling. No DANGER-level content present in this document. Notice hierarchy CSS comment block added to stylesheet for series maintainability.PASS