Meet the People Whose Decisions Define the Gap
An ISO 42001 gap is not an abstract standard failure. It is a real person, in a real seat, making a real decision that creates a real exposure. This module follows four of them. Each one carries a different facet of the firm's risk — and each one will own a different document once the remediation programme begins.
Priya Anand · COLP
The default AI risk owner — and the named accountable officer under SRA Rule 9.1. Priya has inherited every AI exposure the firm currently carries, by default, because no one else has been named. She is the personal registration behind the regulatory risk; the gaps are, in regulatory terms, her gaps until they are reassigned.
Jonathan Halstead · Managing Partner
The signature on every AI deployment. Jonathan approved the rollout of GPT-4o and Microsoft Copilot across the firm without a policy, a register, or a system owner. Under ISO Cl. 5.2, leadership commitment is a mandatory, evidence-based control — and the absence of that evidence is the most acute governance gap in the firm.
Marcus Webb · Head of IT
Owns the tech, not the governance. Marcus manages the licences, the integrations, the access controls. He is, however, completely unaware of his AI governance obligations under ISO Cl. 7.5 (system register) and Cl. 8.4 (supplier governance). The technical controls he runs every day are unmanaged from a compliance perspective.
Sophie Chen · Paralegal
The end-user whose unsupervised AI exposure defines the firm's front-line risk. Sophie uses GPT-4o for research, drafts emails with Copilot, and runs LEAP AI for billing prompts. None of her usage is logged. None of her outputs are reviewed. She is the living embodiment of the human-in-the-loop gap — and of the hallucination risk documented in Ayinde.
"The first thing Priya did, after reading the SRA thematic review, was open her firm's intranet and search for 'AI policy.' Nothing came back. The second thing she did was ask Jonathan. He said he'd assumed IT was handling it. IT said they'd assumed the COLP was handling it. Nobody was handling it. The firm had been running GPT-4o for ten months."
— Halstead & Cole LLP — AIMS Implementation NarrativeEvery system below is live. None is documented in any register. None has a system owner. None has been the subject of a risk assessment, a data protection impact assessment, or a supplier governance review. This is the firm's "before" picture — the nine gaps are measured against it.
| System | Use Case | Primary User | Status Today |
|---|---|---|---|
| SYS-AI-001 OpenAI GPT-4o (API) |
Legal research, drafting, case-summary generation | Sophie Chen & dispute resolution team | Unsanctioned |
| SYS-AI-002 Microsoft Copilot (M365) |
Email drafting, document summarisation, calendar drafting | All fee earners | Unsanctioned |
| SYS-AI-003 LEAP AI (Practice Management) |
Billing narrative generation, time-recording prompts | All fee earners | Unsanctioned |
| SYS-AI-004 Browser Bookmarks / Personal Prompts |
Ad-hoc prompt libraries held by individual fee earners | Various — no central visibility | Unrecorded |
Why "we don't have an AI policy" is not a defence
The SRA's existing Code of Conduct rules — competence (3.2), supervision (3.3), confidentiality (6.3), client information (7.1), and COLP accountability (9.1) — already apply to AI use at Halstead & Cole. The absence of a written AI policy does not suspend those rules. It removes the firm's ability to demonstrate that those rules have been honoured. The gaps in this module are the difference between the rulebook the SRA expects the firm to operate under and the rulebook the firm actually has.
Why the C-Suite Cannot Defer — Two Triggers, Both Live
AI compliance is not an IT problem wearing a fancy name. It is a regulatory exposure that lands on the COLP and a commercial exposure that lands on the Managing Partner. Halstead & Cole has just been hit by both triggers within the same week. Every gap in this module is a consequence of those triggers not having been acted upon sooner.
Priya opens her email on a Monday morning and reads two documents that change her working week. Each one maps to a separate category of exposure — the first to SRA conduct risk, the second to court-process risk. Together they make clear that the firm's current AI usage is incompatible with the regulator's expectations.
Most COLPs cannot outline their AI responsibilities
The SRA's thematic review of compliance officers found that the majority of COLPs in the sample could not outline all of their responsibilities. The review specifically called out technology-adjacent obligations. Priya reads this and recognises herself in the finding. The SRA expects the COLP to understand the firm's full risk surface — including the AI tools the fee earners are using without supervision. The current gap between that expectation and the firm's reality is the source of every critical gap in this module.
18 fabricated citations, one unrecoverable regulatory moment
The Ayinde / Al-Haroun judgment is the case that ends the debate about whether LLMs are reliable legal research tools. ChatGPT produced 18 fabricated case citations in a single document. The fee earner used the outputs without verification. The High Court's finding is now binding authority on the standard of human oversight required when an LLM is in the workflow. Priya knows that Sophie Chen uses GPT-4o daily — and that the firm has no prompt logs, no output reviews, and no system capable of telling her whether their own work contains hallucinations.
The hallucination Priya cannot rule out
Halstead & Cole has no prompt logs and no register of AI-generated outputs. There is no procedure for verifying an LLM's citations against primary sources. There is no record of which matters used AI-generated research. The firm therefore cannot, today, look Priya in the eye and confirm that none of its current caseload contains a fabricated citation. The Ayinde defect could already be present in work the firm has filed. This is not a future risk. It is an open exposure today.
Five days after Priya reads the SRA thematic review, the firm's largest client sends a procurement questionnaire. It is not a regulatory document. It is a revenue document — and Halstead & Cole cannot answer it.
Pemberton's questionnaire asks three questions. Halstead & Cole cannot answer "yes" to any of them.
Does the firm have a documented AI policy?
Pemberton's procurement team wants to see an AI policy signed by leadership. The firm has none. Priya drafts a holding response; Jonathan is asked whether to commit to a deadline. The default answer, today, is "no." Pemberton's procurement language is increasingly common in 2026 panel agreements.
Does the firm maintain a register of AI systems in use?
Pemberton wants evidence that the firm can list every AI system touching client data, its vendor, its data classification, and its review status. Marcus, the Head of IT, could produce a list of licences — but no AI system register exists. The procurement team interprets the absence as governance immaturity, not administrative gap.
Does the firm operate a documented human-in-the-loop review for AI outputs?
Pemberton wants the procedure. The firm does not have one. Sophie Chen's daily GPT-4o usage has no documented review step. The honest answer to Pemberton's question is "we have not formalised this." Pemberton's response, if the answer does not change, is to begin a panel review.
The £620k question
Pemberton Capital accounts for 7.6% of Halstead & Cole's annual turnover. Losing the panel is not a paper-cut; it is a material revenue event for a mid-market firm. The procurement questionnaire is not adversarial — it is becoming standard market practice. The deadline the firm faces is not regulatory. It is commercial. The ISO 42001 controls that close the gaps in this module are the same controls that answer Pemberton's questions.
Advanced Track — Thornbury Legal Ltd.
For learners working with boutique or highly confidential client scenarios, the same nine gaps apply — but the architecture changes. Thornbury Legal Ltd. is a fictional 14-fee-earner firm self-hosting an open-weights model on-premises to protect client confidentiality. Their Gap 9 challenge is different from Halstead & Cole's: they fail to check the model's weight provenance and to confirm that vendor safety filtering is still active when inference is local. Self-hosting eliminates the data-transfer risk and introduces a weight-trust risk. The same 7-step framework applies; the controlled documents are different.
The Five Gaps That Block ISO 42001 Certification
A Critical gap means immediate regulatory exposure for the firm and a hard blocker at certification. Each gap below is presented in the 7-step teaching pattern this programme uses for every control: the executive hook, the ISO requirement, the regulator's parallel rule, the scenario failure, the controlled document that fixes it, a free template pointer, and the call to action. The five critical gaps cover the foundational, accountability, risk, data, and transparency layers of the AIMS.
How to read each gap
Each gap is structured to mirror how an auditor, a regulator, or a panel-client procurement team will interrogate it. The Hook places the executive in the moment of failure. What ISO 42001 Requires uses the prescriptive language of the standard ("shall" = mandatory; "should" = recommended; "may" = permitted). What the Regulator Requires maps the ISO clause to the SRA Code or UK GDPR. The Gap in the Scenario walks through the live failure. The Fix introduces the controlled document.
POL-AIMS-001 — AI Policy ISO Cl. 5.2 · Annex A.2
Jonathan Halstead approved the rollout of GPT-4o and Copilot across the firm without a written policy. There is no documented statement of acceptable AI use, no prohibited-use list, no signature from leadership, and no firm-level position the COLP can point to. Every other governance control is optional in practice until this document exists.
The decision that is now on the Managing Partner's record
Six months ago Jonathan approved the rollout of GPT-4o to all fee earners in a single email. There was no policy attached. There was no register entry. There was no system owner. That email is now the most consequential technology decision the firm has made this year — and the only evidence that it happened is an Outlook message.
Top management shall establish an AI policy
Clause 5.2 requires top management to establish an AI policy that is appropriate to the purpose of the organisation, provides a framework for AI objectives, includes a commitment to satisfy applicable requirements, and commits to continual improvement. Annex A control A.2 requires the policy to be documented, communicated, and available to interested parties.
SRA Code of Conduct Rule 2.1 & Rule 9.1
Rule 2.1 requires the firm to have effective governance structures, arrangements, and processes. Rule 9.1 requires the COLP to take all reasonable steps to ensure compliance. An undocumented rollout decision is, in regulatory terms, an unmanaged technology programme — and "reasonable steps" cannot be evidenced without a signed AI policy.
No documented position on responsible AI use exists
The firm has no policy on acceptable AI use cases, no position on prohibited activities (e.g., uploading client documents containing personal data to commercial APIs), no supervision expectations, and no sign-off requirement. Fee earners infer what is allowed from the absence of any rule. Sophie Chen's daily GPT-4o usage is therefore either implicitly endorsed or implicitly tolerated — neither position is defensible.
Implement a formal AI Policy
POL-AIMS-001 — the firm's first controlled document under the AIMS — sets out: the firm's commitment to responsible AI use, the permitted and prohibited use cases, the human-in-the-loop principle, the data classification rules, the disclosure obligation to clients, the incident reporting procedure, and the named accountable officer. The policy is signed by Jonathan Halstead as Senior Partner and countersigned by Priya Anand as COLP.
Free Template — POL-AIMS-001 Starter
A starter template for POL-AIMS-001 is provided as a downloadable starting point. It contains the standard policy sections, sign-off block, and revision history fields. Disclaimer: this template shall be reviewed by a qualified solicitor before adoption. It is an educational scaffold, not legal advice.
RACI-AIMS-001 — Roles & Responsibilities Register ISO Cl. 5.3
Priya holds every AI risk accountability by default. There is no AI Risk Owner, no AI System Owner for any tool, no Human-in-the-Loop reviewer designated for any practice area. The entire firm's AI governance rests on one individual's personal registration — the most fragile operating model an SRA-regulated firm can have.
One person, every risk, no backstop
Priya is the named COLP. She is also the default AI Risk Owner, the default AI System Owner for every tool in use, the default data-protection reviewer, the default incident responder, and the default human-in-the-loop reviewer. If she is on leave for two weeks, no AI governance decision can be made. If she resigns, the AIMS has no continuity. The firm has built a one-person control structure for a problem that crosses every practice area and every system.
Roles, responsibilities, and authorities shall be assigned and communicated
ISO/IEC 42001 Clause 5.3 requires roles, responsibilities, and authorities to be formally assigned and communicated. At Halstead & Cole this means naming an AI Risk Owner, AI System Owners (one per tool), an AIMS Programme Lead, and Human-in-the-Loop reviewers. At the moment, nobody has any of these titles. Module 02 fixes this with RACI-AIMS-001.
SRA Code of Conduct Rule 3.3 & Rule 9.1
Rule 3.3 requires managers to ensure adequate supervision. Rule 9.1 requires the COLP to evidence reasonable steps. A single-individual accountability model cannot evidence either — there is no documented second line of defence, no documented escalation path, and no evidence that any AI decision was reviewed by anyone other than Priya. SRA Standards and Regulations 2019 expects the firm to demonstrate that compliance responsibilities are distributed and that supervision is operative, not theoretical.
Dangerous single-owner concentration of risk
The firm has not defined who owns AI risk, who owns each AI system, who reviews AI outputs in each practice area, who reports AI incidents, who signs off on new AI tool deployments, or who reviews model updates. Every one of these decisions defaults to Priya by default — and defaults that are not documented do not exist for ISO purposes. The concentration of risk on a single individual is, in audit terms, a single point of failure that no firm can defend.
Define a Roles & Responsibilities RACI Matrix
RACI-AIMS-001 assigns named individuals to each AI governance role: AI Risk Owner (Priya), AI System Owner per tool (Marcus for infrastructure, named partners for practice-area tools), Human-in-the-Loop reviewers per practice area, Incident Reporting Lead, and Document Controller. The RACI is signed off by the Senior Partner and is the second controlled document in the AIMS. It converts the firm's implicit accountability model into an explicit, auditable, sign-off-ready structure.
Module 03 will build the actual RACI, name the role holders, define the escalation path
Module 02 introduces the gap and names the document. Module 03 will build the actual RACI matrix, name the actual role holders in Halstead & Cole's structure, and define the actual escalation path for AI-related decisions. The gap is identified, the fix is named, the severity is set, and the free starter template is available for download.
Free Template — RACI-AIMS-001 Starter
A starter RACI matrix is provided, with role definitions and the standard Responsible / Accountable / Consulted / Informed grid. Disclaimer: shall be reviewed by qualified HR and compliance counsel before adoption.
REG-AIMS-RISK-001 — AI Risk & Impact Assessment Register ISO Cl. 6.1 · 6.1.2
The firm has never assessed the risks of using AI. Hallucination, fabricated citations, biased outputs, privilege waiver, data leakage, model drift — none has been formally identified, scored, or treated. The firm operates AI tools without knowing the size of the risk surface it is standing on.
"What could go wrong?" — the question the firm cannot answer
When Priya is asked by Pemberton Capital to describe the firm's top three AI risks, she cannot answer. The firm has no AI risk register. The firm has never scored the likelihood or impact of an AI-related incident. The firm has no documented risk treatment plan. The honest answer is "we don't know what we don't know."
The organisation shall assess AI risks and determine treatment
Clause 6.1.2 requires the organisation to assess AI risks, considering likelihood, impact, and the effectiveness of existing controls. Clause 6.1.3 requires the organisation to determine and apply risk treatment options. Annex A control A.5 requires a structured, documented AI risk assessment process with defined criteria.
SRA Rule 1.2 & UK GDPR Article 35
SRA Rule 1.2 requires the firm to identify and mitigate risks to clients. UK GDPR Article 35 requires a Data Protection Impact Assessment where processing is likely to result in a high risk to the rights and freedoms of natural persons — automated decision-making and large-scale processing of client data by an AI tool clearly trigger this. The absence of both assessments is a current breach.
Hallucination, bias, and fabricated citations are unassessed risks
The firm has not assessed: (a) the hallucination risk — fabricated citations in court filings, the precise Ayinde defect; (b) the bias risk — discriminatory outputs in employment or property work; (c) the privilege-waiver risk — uploading client documents to a commercial LLM; (d) the data-protection risk — international transfers without valid mechanisms. None has been scored. None has a treatment plan.
Create an AI Risk & Impact Assessment Register
REG-AIMS-RISK-001 documents each identified AI risk, its likelihood, its impact, its inherent risk score, the existing controls, the residual risk score, and the agreed treatment plan (accept / mitigate / transfer / avoid). The register is reviewed quarterly and is the third controlled document in the AIMS.
Free Template — REG-AIMS-RISK-001 Starter
A starter risk register template includes pre-populated risk categories (hallucination, bias, privilege waiver, data transfer, model drift) and a 5×5 scoring matrix. Disclaimer: shall be reviewed by qualified risk and compliance counsel before adoption.
REG-AIMS-DATA-001 — Data Governance & Transfer Impact Assessment ISO Cl. 8.3
Client data is being transmitted to commercial AI APIs (GPT-4o, Copilot, LEAP) without proper data transfer mechanisms, without signed Data Processing Agreements, and without a documented classification of what client data can and cannot be submitted. This risks waiving legal professional privilege — an exposure no firm can survive.
The paralegal's open tab is the firm's privilege exposure
Sophie Chen has GPT-4o open in a browser tab next to a client matter file. She pastes a paragraph of client correspondence to "summarise this for the partner." The paragraph contains the client's name and the substance of their dispute. By the time Sophie closes the tab, that text has travelled to a US-based commercial API, been logged in OpenAI's infrastructure, and may now form part of model training data or future retention. Privilege may already be lost.
Data used in AI systems shall be governed by documented controls
Annex A control A.7 (Cl. 8.3) requires the organisation to determine and document the data used in AI systems, including the source, quality, and appropriateness of the data; to define controls on data access, processing, and retention; and to verify that data flows are subject to valid legal mechanisms for cross-border transfer.
UK GDPR Article 28, 44, 46 & SRA Rule 6.3
UK GDPR Article 28 requires a signed Data Processing Agreement with every processor. Articles 44–46 require a valid transfer mechanism for personal data leaving the UK. SRA Rule 6.3 requires the firm to keep client affairs confidential. Submitting client data to a US-based API without these mechanisms is, today, an active breach of three regulatory layers simultaneously.
No DPA, no transfer mechanism, no privilege protection
OpenAI has not been engaged under a UK GDPR-compliant DPA. The firm's data flows to the US, which is not an adequacy jurisdiction for personal data without the UK International Data Transfer Agreement (IDTA) or UK Addendum to the EU SCCs — neither is in place. There is no documented rule forbidding fee earners from submitting client documents to commercial LLMs. There is no transfer impact assessment. The privilege waiver risk is current and unmanaged.
Establish an AI Data Governance Register & TIA
REG-AIMS-DATA-001 documents each AI system's data flows: what categories of client data can be submitted, what is prohibited (e.g., identifiable client names in sensitive matters), the data classification rules, and the Transfer Impact Assessment for every cross-border flow. A signed DPA is mandatory before any personal data enters a cloud AI tool. The register is reviewed annually and whenever a vendor changes its data policy.
Free Template — REG-AIMS-DATA-001 Starter
A starter data governance register includes a vendor-by-vendor DPA tracker, transfer mechanism register (IDTA / UK Addendum / SCCs), and data classification matrix. Disclaimer: shall be reviewed by qualified data protection counsel before adoption.
DOC-AIMS-DISC-001 — AI Transparency Disclosure Statement ISO Annex A.2
The firm is sending AI-assisted correspondence to clients without disclosing that AI was used. This breaches the SRA's February 2026 transparency guidance and exposes the firm to a regulatory finding that is straightforward to evidence — every email Copilot helped draft is a record.
The client-side evidence the firm is producing daily
Every email a Halstead & Cole fee earner drafts with Copilot assistance is evidence of undisclosed AI use. Pemberton Capital — a client already asking about AI governance — receives correspondence drafted this way every week. When Pemberton's compliance team eventually asks to see the firm's disclosure statement, the absence of one is the most visible gap in the procurement file.
Transparency about AI use shall be documented and communicated
Annex A control A.8 requires the organisation to document its policies and procedures for providing transparency about the use of AI, including the audiences to whom AI use shall be communicated, the format of disclosure, and the points in the AI lifecycle at which disclosure shall occur.
SRA Code Rule 7.1 & SRA AI Guidance Feb 2026
SRA Rule 7.1 requires clients to be given information in a way they can understand. The SRA's February 2026 AI compliance tip states that "it should always be made clear to clients where they are interfacing with AI." This applies today, to every AI-assisted communication the firm sends — not to a future state once governance is "mature."
The firm is in breach of SRA guidance today
Copilot-assisted emails go to clients daily without disclosure. LEAP AI-assisted billing narratives go to clients without disclosure. ChatGPT-assisted research summaries inform advice without disclosure. None of these communications contains a transparency statement. The firm is currently in breach of the SRA's published AI guidance — and that breach is documented in the firm's own email server.
Implement an AI Transparency Disclosure Statement
DOC-AIMS-DISC-001 is a standard disclosure statement inserted into client engagement letters and AI-assisted correspondence, in language approved by Priya and countersigned by Jonathan. The disclosure covers: which categories of work the firm uses AI for, the human review commitment, the data classification rules applied to AI inputs, and the client's right to request human-only delivery. The statement is reviewed annually.
Free Template — DOC-AIMS-DISC-001 Starter
A starter disclosure statement, suitable for adaptation into engagement letters and email footers. Includes the four mandatory disclosure elements and a sign-off block. Disclaimer: shall be reviewed by qualified solicitor before adoption.
Peer Group — The Governance Academy
The five critical gaps above are the topic of the live Governance Academy working sessions for COLPs and compliance leads. Each gap is worked through in a small-group setting with peer firms at similar stages of the AIMS journey. The first working session is scheduled within 30 days of module completion. Details are issued in the Module 02 follow-up email.
The Four Gaps That Will Sink an Audit if Not Closed
A High severity gap is not "tomorrow's problem." It is a current, significant compliance weakness — one that will surface in any serious internal or external audit, and that will be the first thing a regulator, a panel client, or an ISO certification body asks about. The four high-severity gaps cover the system register, the lifecycle procedure, the human oversight requirement, and the supplier governance regime.
Severity difference — Critical vs High
A Critical gap means the firm is in current breach of a regulatory obligation and cannot achieve certification. A High gap means the firm is in current significant compliance weakness and will be unable to evidence the requirement at audit. The four gaps below are not "less urgent" than the five critical gaps; they are gaps whose remediation can be sequenced over the next 60–120 days, while the critical gaps must be closed within 30.
REG-AIMS-SYS-001 — AI System Register + Prompt Repository ISO Cl. 7.5
There is no central register of AI systems in use at the firm. Fee earners have embedded AI prompts in browser bookmarks, email signatures, and one-off Word macros — with no version control, no sign-off, and no visibility for the IT team or the COLP. The firm does not know what AI tools it depends on, and cannot evidence that the tools it does know about are managed.
The Head of IT discovers the firm is running 14 AI tools, not four
When Marcus Webb is asked to list every AI tool in use, he produces four. A subsequent anonymous staff survey reveals ten more — including two browser extensions, an Excel add-in, and a personal ChatGPT Plus account paid for on a fee earner's personal card. None of the ten is in any register. None has been risk-assessed. None has a DPA.
Documented information shall be maintained for every AI resource
Clause 7.5 requires the organisation to determine and provide the resources needed for the establishment, implementation, maintenance, and continual improvement of the AIMS. Documented information shall be maintained to the extent necessary to support the operation of the AIMS and to retain evidence of conformity.
SRA Rule 2.1 & UK GDPR Article 30
SRA Rule 2.1 requires effective governance arrangements. UK GDPR Article 30 requires a Record of Processing Activities covering all personal data processing. AI tools processing client personal data are processing activities that must be recorded — and the record does not exist.
Prompts, tools, and licences are unversioned and unmanaged
Staff embed AI prompts in browser bookmarks and emails. There is no central prompt library. There is no version control. When a vendor updates its model, prompts that worked yesterday may not work tomorrow. The firm has no procedure for tracking prompt changes, no record of which prompt produced which client output, and no ability to reproduce a high-quality result.
Deploy an AI System Register + Prompt Repository
REG-AIMS-SYS-001 catalogues every AI system: system ID, vendor, version, deployment date, system owner, data classification permitted, last review date, and current status. A companion Prompt Repository holds approved prompts with version history, the system they are associated with, and the review approval. New tools cannot be added to production without an entry in the register.
Free Template — REG-AIMS-SYS-001 Starter
A starter system register with vendor, version, and approval fields. Includes a one-page prompt repository template. Disclaimer: shall be reviewed by IT and compliance counsel before adoption.
PROC-AIMS-LIFE-001 — AI System Lifecycle Management Procedure ISO Cl. 8.2
When vendors (OpenAI, Microsoft, LEAP) release a new model version, the firm has no procedure for assessing the update, regression testing the change, or informing affected clients. The firm's risk profile can shift overnight without anyone noticing.
The model Sophie's prompts were tested on is no longer the one answering
OpenAI retires GPT-4o on a Tuesday at 6am US Pacific. By Tuesday lunchtime UK time, Sophie is using GPT-5 without knowing it. The prompts she has refined over six months behave differently. The contract review tool's outputs have changed. Nobody at the firm has been notified, and no procedure exists for assessing the change.
AI systems shall be managed through defined lifecycle stages
Clause 8.2 requires the organisation to implement a documented process for planning, designing, developing, testing, deploying, operating, monitoring, and retiring AI systems. The standard is explicit: lifecycle management includes the management of changes to AI systems, including vendor-side model updates.
SRA Rule 3.2 (Competence) — Continuous
A fee earner who believes they are using GPT-4o but is actually using GPT-5 is, in regulatory terms, operating outside their competence — they have not assessed the new model's failure modes. The firm has a duty to ensure fee earners know which model they are using and what its limitations are.
No regression test, no impact assessment, no client notification
Marcus has no alert configured for vendor model updates. The procurement function does not flag a model change. There is no procedure for regression-testing the change against a known set of expected outputs. There is no client-facing notification when a change in tool or model affects the work the firm is producing for them.
Create an AI System Lifecycle Management Procedure
PROC-AIMS-LIFE-001 defines the firm's lifecycle stages: Plan, Acquire, Onboard, Operate, Update, Decommission. The procedure requires: vendor change notifications, regression test suites, impact assessments, sign-off before re-deployment, and client notification where the change affects work product. The procedure is reviewed annually and after every material vendor change.
Free Template — PROC-AIMS-LIFE-001 Starter
A starter lifecycle procedure with stage gates, sign-off blocks, and a regression-test checklist. Disclaimer: shall be reviewed by IT and compliance counsel before adoption.
PROC-AIMS-HITL-001 — Human Oversight & AI Output Review Procedure ISO Cl. 9.1
AI outputs are used by staff without a defined human-in-the-loop review step or sign-off requirement. The firm has no documented expectation that fee earners verify AI-generated research, drafts, or correspondence before they reach a client. This is the Ayinde defect, waiting to be reproduced.
The output that reaches the court with no human review on record
Sophie uses GPT-4o to generate a summary of recent case law on a property dispute. The summary includes three case citations. Sophie does not check them against a primary source — there is no procedure requiring her to. The summary is incorporated into a skeleton argument and filed. One of the three citations is fabricated. The Ayinde defect has been reproduced.
AI performance and human oversight shall be monitored
Clause 9.1 requires the organisation to monitor, measure, analyse, and evaluate the performance and effectiveness of the AIMS. This includes monitoring the operation of AI systems, the rate of human overrides of AI outputs, and the effectiveness of human review procedures. The oversight requirement is mandatory, not aspirational.
SRA Rule 3.2 (Competence) & Rule 3.3 (Supervision)
The SRA's competence and supervision rules require fee earners to verify AI-assisted work before it constitutes a client deliverable. The Ayinde judgment makes clear that a court will not accept "the AI told me" as a defence. The duty of verification is the fee earner's — and the supervisor's — and is non-delegable.
No defined review step, no documented override rate, no incident log
The firm has no procedure specifying: which AI outputs require a second-pair-of-eyes review, which can be used unsupervised, what the review check-list is, how human override decisions are recorded, or how AI-related incidents are reported. Fee earners make these decisions individually, inconsistently, and with no record.
Document a Human Oversight & AI Output Review Procedure
PROC-AIMS-HITL-001 defines: which AI outputs require human review before client delivery; the review check-list (citations verified, logic checked, bias flagged, data classification confirmed); the recording requirement for human override decisions; the incident reporting procedure for AI-related errors; and the monthly review of override rates per system.
Free Template — PROC-AIMS-HITL-001 Starter
A starter human-in-the-loop procedure including a review check-list, override log template, and incident reporting form. Disclaimer: shall be reviewed by compliance counsel before adoption.
REG-AIMS-SUP-001 — AI Supplier Governance Register + Third-Party Checklist ISO Cl. 8.4
The firm has not reviewed the Terms of Service of its AI vendors for data retention, model training opt-outs, or sub-processor arrangements. No Data Processing Agreements are in place. The contract review tool has a vendor nobody at the firm can name. The supply chain is unmanaged.
The contract review tool's vendor is unknown to the firm
The commercial property team uses a contract review tool. When Priya asks the team for the vendor name, the partner cannot recall. When Priya asks Marcus, he can confirm the licence is paid but not the underlying vendor's data processing terms. The firm is processing client contracts through an AI tool whose data-handling practices it has never assessed.
Supplier relationships shall be governed and documented
Clause 8.4 requires the organisation to control externally provided processes, products, and services that relate to the AIMS. This includes AI vendors. The standard requires documented supplier assessments, defined controls over the supplier relationship, and ongoing monitoring of supplier performance.
UK GDPR Article 28 (DPAs) & Article 32 (Security)
Article 28 requires a binding contract with every data processor. Article 32 requires appropriate technical and organisational measures. The ICO has issued clear guidance that AI vendors processing personal data are processors and require a DPA. Submitting client data to a vendor without a DPA is a current breach of Article 28.
No DPAs, no ToS review, no third-party risk assessment
OpenAI's Terms of Service have not been reviewed against the firm's data classification rules. Microsoft Copilot's DPA status has not been confirmed. LEAP AI's sub-processor list is unknown. The contract review tool's vendor is unknown. The firm has no register of supplier relationships, no review schedule, and no procedure for adding a new AI vendor to the supply chain.
Implement an AI Supplier Governance Register + Third-Party Checklist
REG-AIMS-SUP-001 catalogues every AI vendor: vendor name, DPA status, transfer mechanism in place, data retention period, model training opt-out confirmation, sub-processor list, last review date, and risk rating. The Third-Party Checklist is a pre-onboarding questionnaire that must be completed before any new AI vendor is added. The register is reviewed annually and on every vendor policy change.
Advanced Track — Thornbury Legal Ltd.
For firms self-hosting open-weights models, Gap 09 takes a different shape. Thornbury Legal Ltd. fails to check the model's weight provenance and to confirm that vendor safety filtering is still active when inference is local. Self-hosting eliminates the data-transfer risk — and introduces a weight-trust risk. The supplier governance procedure for a self-hosted model shall document: the source of the model weights, the licence terms, the absence of vendor-side filtering, and the firm's own pre-deployment safety review.
Peer Group — The Governance Academy
The four high-severity gaps are covered in the Module 05–10 working sessions. The combined critical + high gap set (9 total) is the topic of the live cross-firm remediation workshops. Working sessions for the procurement-facing gaps (System Register, Supplier Governance) are scheduled in tandem with the risk and oversight sessions. Details are issued in the Module 02 follow-up email.
Nine Controlled Documents — The Master Inventory
Every gap in this module closes with a controlled document. The table below is the firm's master document register — the complete inventory of policies, registers, and procedures that Halstead & Cole will produce across Modules 03–10 to remediate the nine gaps. The order is deliberate: it reflects the implementation sequence, not the document numbering. POL-AIMS-001 comes first because nothing else can be approved without the policy in place.
How to read this register
Each row is one controlled document. The ISO Clause column names the clause the document satisfies. The Severity column matches the gap it closes. The Owner column names the individual accountable for the document's production, approval, and maintenance. The free starter template referenced in each gap section above is the starting point for the document in that row.
| Doc ID | Document | Closes Gap | ISO Clause | Severity | Owner |
|---|---|---|---|---|---|
| POL-AIMS-001 | AI Policy & Management Direction | Gap 01 | Cl. 5.2 / A.2 | Critical | Jonathan Halstead (sign), Priya Anand (owner) |
| RACI-AIMS-001 | AI Roles & Responsibilities Register | Gap 02 | Cl. 5.3 / A.3 | Critical | Priya Anand |
| REG-AIMS-SYS-001 | AI System Register + Prompt Repository | Gap 03 | Cl. 7.5 / A.4 | High | Marcus Webb |
| REG-AIMS-RISK-001 | AI Risk & Impact Assessment Register | Gap 04 | Cl. 6.1 / A.5 | Critical | Priya Anand |
| PROC-AIMS-LIFE-001 | AI System Lifecycle Management Procedure | Gap 05 | Cl. 8.2 / A.6 | High | Marcus Webb |
| REG-AIMS-DATA-001 | AI Data Governance Register & Transfer Impact Assessment | Gap 06 | Cl. 8.3 / A.7 | Critical | Priya Anand + Marcus Webb |
| DOC-AIMS-DISC-001 | AI Transparency Disclosure Statement | Gap 07 | Annex A.2 / A.8 | Critical | Priya Anand |
| PROC-AIMS-HITL-001 | Human Oversight & AI Output Review Procedure | Gap 08 | Cl. 9.1 / A.9 | High | Senior Partner designate |
| REG-AIMS-SUP-001 | AI Supplier Governance Register + Third-Party Checklist | Gap 09 | Cl. 8.4 / A.10 | High | Marcus Webb + Priya Anand |
Each of the nine gaps in this module was presented using a single, consistent 7-step structure. The same structure is used in every working session in the Governance Academy programme. Executives who learn to read a gap through this lens will read any ISO 42001 control the same way — and will not need a translator every time a new requirement lands.
The Hook — put the executive in the moment of failure
ISO/IEC 42001 Clause 4.1 requires the organisation to understand internal and external issues relevant to AI governance. For Halstead & Cole, the internal context includes: the undocumented AI tools in use, the single-COLP accountability structure, and the absence of an IT function with AI expertise. The external context includes: the SRA's regulatory expectations, the Ayinde judgment, client-side AI-governance questionnaires, and the EU AI Act's phased implementation (high-risk AI obligations apply from August 2026). Map these before scoping — they determine which risks are highest priority.
What ISO 42001 Requires — the prescriptive language of the standard
The scope shall cover every AI system the firm uses, regardless of whether it was formally approved. Excluding a tool from the scope does not remove the risk — it removes the visibility. At Halstead & Cole, the scope shall include: HC-SYS-001 (commercial LLM), HC-SYS-002 (contract review tool — vendor to be confirmed), and HC-SYS-003 (Word macro prompts). These three system IDs will be carried through every subsequent module.
What the Regulator Requires — the parallel rule that makes the gap live
This is one of the most important distinctions in ISO 42001 implementation, and most firms get it wrong. A commercial LLM accessed via an API (like Halstead & Cole's current setup) sends data to a third-party server for processing. Governance must address: international data transfer mechanisms, DPA obligations, and vendor terms of service. A self-hosted model — where inference happens on the firm's own servers — eliminates the data-transfer risk but introduces different risks: model weight provenance, licence compliance, and the absence of vendor-managed safety filtering. Halstead & Cole currently uses commercial LLMs only. If the firm ever considers a self-hosted model for confidentiality reasons, the scope statement must be amended before deployment begins.
The Gap in the Scenario — the live failure, in plain terms
Clause 4.2 requires the firm to identify interested parties and their AI-relevant requirements. For Halstead & Cole: the SRA (regulatory compliance), the panel client (procurement questionnaire), the LLM vendor (DPA, terms of service), the ICO (UK GDPR compliance), individual fee earners (competence obligations), and the firm's PI insurer (disclosure of AI use). Each interested party has different requirements, and the governance framework must be able to address each one. The scope statement shall list them.
Free Template — a downloadable starting point, with the required disclaimer
A starter template is provided for the controlled document. The template contains the standard sections, sign-off block, and revision history fields. The disclaimer is non-negotiable: the template shall be reviewed by a qualified solicitor before adoption. It is an educational scaffold, not legal advice. Templates are starting points, not finished products.
Call to Action — the peer-group invitation that completes the loop
The executive is invited to join a working peer group — typically the Governance Academy's monthly COLP roundtables — where the same gap is worked through with peer firms at similar stages of the AIMS journey. The peer group is where templates get pressure-tested, where the COLP's name is on the attendance list, and where the work that started in this module becomes an issued, signed, working document.
What Halstead & Cole will have at the end of Module 02
- A complete inventory of the firm's nine AI governance gaps, with severity ratings and the document that closes each one
- A named cast of stakeholders — Priya, Jonathan, Marcus, Sophie — each with a defined role in the remediation
- A mapped understanding of the two triggers driving the C-Suite to act: regulatory exposure and client procurement
- A master document register that names every controlled document the firm will produce across Modules 03–10
- The 7-step teaching method, internalised — the C-Suite now reads every ISO 42001 control through the same lens
Module 01 Completion Checklist
Every item below shall be confirmed by Halstead & Cole's C-Suite before the firm proceeds to Module 03. This module is the bridge between understanding and action — if any item is unclear, return to the relevant section and re-read it before continuing. A misunderstood gap is a gap that will not be closed.
Module 02 — Preview
Module 02 moves from understanding to building. Halstead & Cole will produce its first three controlled documents: the AI Policy (POL-AIMS-001), the Governance RACI Matrix (RACI-AIMS-001), and the AI System Register (REG-AIMS-SYS-001). By the end of Module 02, the firm will have issued a signed Leadership Statement formally declaring the AIMS established — and will be able to answer the first four questions on the panel client's AI-governance questionnaire.
Module 01 — Complete
Halstead & Cole has now mapped its full AI governance gap landscape: the four cast members, the two triggers, the five critical gaps, the four high-severity gaps, the master document register, and the 7-step method that structures every remediation. The firm can answer Pemberton Capital's three questions. Proceed to Module 03 when every item on the completion checklist has been confirmed by the C-Suite.