UNUS London · ISO/IEC 42001:2023 Finance Series · ART-AIMS-FIN-G004 v1.0 · FCA · MLR 2017 · Consumer Duty · ISO/IEC 27001:2022
The question that exposes two years of unassessed risk
It is day five of the Harcastle questionnaire. Marcus has resolved the ChatGPT data breach discovery — ChatGPT is suspended, legal advice has been sought, the ICO notification assessment is underway. He has sent the AI System Register to Harcastle. Question three is marked provisionally satisfied.
Harcastle's legal team sends question four.
Harcastle Group
"Please provide your AI Risk Assessment covering each AI system listed in your register. We would expect this to include for each system: the key risks identified, the likelihood and impact scores applied, the controls in place to manage each risk, and who holds accountability for risk treatment. We would also expect the assessment to have been reviewed and accepted by your board or senior management."
Marcus Osei
He reads the question once. Then he opens the AI System Register. AIMS-SYS-003 — LexisNexis AML screening. In use for eighteen months. Generates alerts that the compliance team investigates and, in some cases, files as Suspicious Activity Reports to the National Crime Agency. No risk assessment exists for it. If the model has a bias — if it flags certain transaction types or client profiles disproportionately — eighteen months of AML decisions may be built on flawed data.
Marcus Osei
AIMS-SYS-004 — Experian credit risk scoring. In use for two years. Informs lending decisions for Harcastle's underlying clients. No risk assessment for that either. He picks up the phone to Diana.
Diana Whitmore
"Marcus, we've been using the LexisNexis platform since we launched the AML service. It's a market-leading product. Are you telling me we've never formally assessed the risk of the AI inside it?"
Marcus Osei
"That's exactly what I'm telling you. The product is reputable. The AI inside it has never been assessed by us."
This is the distinction at the heart of Gap Four. A reputable vendor does not eliminate risk. It means the vendor has managed their risks. The firm using that vendor still has its own obligation to assess the AI risk that the system introduces into its regulated decision-making processes.
The LexisNexis platform may be excellent. The Experian model may be highly accurate. But Marcus does not know that. He has never assessed it. He has no documentation. He has no risk register entry. He has no treatment plan. And when the FCA asks — as Harcastle just did — the answer is silence.
⚠
Warning
ISO/IEC 42001:2023 Clause 6.1 requires organisations to plan actions to address AI risks and opportunities. Under the FCA's Senior Managers & Certification Regime, the Compliance SMF is personally accountable for the firm's risk framework. Where AI systems are influencing regulated decisions — AML alerts, credit risk scores, suitability assessments — and no AI risk assessment has been conducted, the Compliance SMF cannot demonstrate that they have identified and managed the material risks within their area of accountability. Under SM&CR, undocumented risk is personal liability without a treatment plan.
What ISO/IEC 42001:2023 requires from Clause 6.1
ISO/IEC 42001:2023 — Clause 6.1 · Actions to Address Risks and Opportunities
When planning for the AI management system, the organisation shall consider the issues referred to in Clause 4.1 and the requirements referred to in Clause 4.2 and determine the risks and opportunities that need to be addressed to give assurance that the AI management system can achieve its intended outcome; prevent, or reduce, undesired effects; achieve continual improvement. The organisation shall plan actions to address these risks and opportunities and how to integrate and implement the actions into its AI management system processes.
Source: ISO/IEC 42001:2023 Clause 6.1. British English retained per ISO/IEC Directives Part 2.
The clause is clear: risks shall be determined, and actions to address them shall be planned. Not noted. Not acknowledged. Assessed, documented, owned, and treated.
ISO/IEC TR 24028:2020 — the technical report on AI trustworthiness referenced in ISO 42001's Annex B — identifies the methodology for AI risk assessment. It establishes four categories of AI-specific risk that every organisation using AI in regulated contexts must address.
Risk Category 1
Hallucination & Incorrect Output
The AI produces confident but factually incorrect outputs. In finance: an AI-generated suitability report contains incorrect regulatory references. A credit narrative states incorrect income figures. An AML alert narrative describes a transaction pattern that doesn't match the underlying data.
Finance example → ChatGPT (AIMS-SYS-001) generating client reports with fabricated performance figures that Jade didn't catch in review
Risk Category 2
Model Bias & Discriminatory Output
The AI produces systematically different outputs for different groups — demographic, geographic, or behavioural — in ways that are not justified by the underlying data. In finance: an AML model that flags transactions from certain ethnic communities at higher rates. A credit model that scores clients from certain postcodes differently.
Finance example → LexisNexis AML platform (AIMS-SYS-003) — eighteen months of alerts with no bias assessment
Risk Category 3
Model Drift & Performance Degradation
The AI's accuracy degrades over time as the real-world data it encounters diverges from its training data. In finance: a credit scoring model trained on pre-2023 economic conditions that performs less accurately in a higher interest rate environment. A fraud detection model that misses new fraud typologies that emerged after its training cutoff.
Finance example → Experian credit scoring API (AIMS-SYS-004) — two years of credit decisions with no drift monitoring
Risk Category 4
Supply Chain & Vendor Dependency
The AI vendor changes their model, their terms, or their service — and the firm has no governance controls to assess or respond to the change. In finance: LexisNexis updates the scoring thresholds in their AML model without notifying the firm. Experian deprecates an API version the firm's systems depend on. The vendor is acquired and data processing terms change.
Finance example → all four vendor-dependent systems — no change notification controls in place (addressed further in F-G9)
◈
ITIL 4 — Risk Management Practice · Secondary: Change Enablement
In ITIL 4, the Risk Management practice requires risks to be identified, assessed, and treated through a structured process — and that treatment decisions are documented with named owners and review dates. REG-AIMS-FIN-RISK-001 functions as the AI-specific risk register within this practice, with each entry mapped to a risk owner from RACI-AIMS-FIN-001. The Change Enablement practice provides a secondary control: any change to an AI system — a model update, a change in data inputs, a vendor terms amendment — shall be assessed against the existing risk register before being implemented. This prevents new risks from entering the environment undocumented.
Section 3 — The FCA & MLR 2017 Position
Why the AML risk is the most acute
Every AI system in the register carries risk. But for Whitmore & Associates, the AML screening platform (AIMS-SYS-003) is the one that concentrates risk most acutely — because its outputs directly inform decisions that have regulatory and legal consequences.
Money Laundering Regulations 2017 and the AML AI
The MLR 2017 requires FCA-regulated firms to have adequate systems and controls to prevent being used to facilitate money laundering. The AML screening platform is a central component of that control framework. Under Regulation 19 of the MLR 2017, firms must monitor their customers and transactions in a manner that is calibrated to their risk appetite — and must be able to demonstrate that monitoring is effective.
If the LexisNexis AI model has a systematic bias — if it is generating false positives for certain transaction types, causing unnecessary client friction and NCA Suspicious Activity Reports — the firm may be filing inaccurate SARs. If it is generating false negatives — missing genuine money laundering because its detection model has drifted from current typologies — the firm may be facilitating financial crime without knowing it.
Neither scenario is hypothetical. Both are foreseeable. Neither has been assessed at Whitmore & Associates. And the MLR 2017 places the obligation to assess squarely on the firm — not on the vendor.
Consumer Duty and the credit scoring risk
The Experian credit scoring model (AIMS-SYS-004) informs credit decisions for Harcastle's underlying clients. If the model has drifted — producing credit scores that are systematically too high or too low compared to actual creditworthiness — clients may be receiving inappropriate credit outcomes. Under the FCA's Consumer Duty (PS22/9), firms must deliver good outcomes for clients. An AI model that produces poor outputs, undetected, is not delivering good outcomes — and the firm cannot demonstrate it is, because it has never assessed the model's performance.
◆
Caution
A common response to this gap is to request the vendor's own model validation documentation and treat it as the firm's risk assessment. This is insufficient. ISO 42001 Cl. 6.1 requires the organisation to assess risks to its own AI management system — including the risks introduced by using a vendor's AI model in its specific regulatory context. A vendor's model card or validation report documents the model's general performance. The firm's risk assessment must document the risk to this firm's clients in this firm's regulated activities. These are different documents, and only one of them satisfies Clause 6.1.
The anatomy of this nonconformance
Gap F-G4 — Nonconformance Taxonomy
Gap IDF-G4
ISO ClauseISO/IEC 42001:2023 Clause 6.1 — Actions to Address Risks and Opportunities
SeverityCritical — regulated AI decisions (AML, credit, suitability) made without documented risk assessment; MLR 2017, Consumer Duty, and SM&CR personal liability exposure simultaneously active
Current StateNo AI risk assessment exists for any of the six systems in REG-AIMS-FIN-SYS-001. Risk categories (hallucination, bias, drift, supply chain) have not been identified, scored, or treated for any system. No risk register. No treatment plan. No board-level risk acceptance.
Regulatory ParallelMLR 2017 Reg. 19 — AML monitoring calibration obligation; FCA Consumer Duty PS22/9 — good client outcomes obligation; FCA SYSC 13 — operational risk management; SM&CR — Compliance SMF personal accountability for risk framework
Commercial RiskHarcastle questionnaire question four cannot be answered. Institutional mandate renewal at risk. FCA supervision finding if AI risk is examined during visit.
Required ArtefactREG-AIMS-FIN-RISK-001 — AI Risk Register for FCA-Regulated Financial Services Firms
How Marcus built the risk register — system by system
Marcus cannot complete a full AI risk assessment for six systems in the time remaining before the Harcastle response deadline. He does something better: he prioritises by risk impact, completes the two highest-risk systems first, and documents the remaining four as "assessment in progress" with committed completion dates. Harcastle receives a partial risk register with a credible remediation plan — which is significantly more satisfying to an institutional due diligence team than silence.
The methodology Marcus uses is based on ISO/IEC TR 24028:2020 and integrated with the firm's existing risk scoring approach.
The risk scoring methodology
Likelihood Scale (1–5)
1Rare — controls prevent occurrence in almost all cases
2Unlikely — occurs occasionally despite controls
3Possible — occurs under some circumstances
4Likely — occurs regularly without intervention
5Almost Certain — expected to occur in most circumstances
Impact Scale (1–5)
1Negligible — minimal operational effect
2Minor — limited client or regulatory effect
3Moderate — material client harm or FCA finding
4Significant — regulatory action or client loss
5Critical — FCA enforcement, personal liability, mandate loss
Risk Score & Appetite
1–4Low — Accept with annual monitoring
5–9Medium — Mitigate within 90 days
10–14High — Mitigate within 30 days or escalate
15–19Critical — Immediate action; board sign-off required
20–25Unacceptable — Suspend system use until mitigated
The risk register — first two systems assessed
| Risk ID |
System |
Risk Description |
L |
I |
Score |
Treatment |
Owner |
| RISK-001 |
AIMS-SYS-003 LexisNexis AML |
Model bias — AI flags certain transaction profiles at disproportionate rates due to historical training data skew, leading to inaccurate SARs and potential discrimination |
3 |
5 |
15 |
Mitigate |
M. Osei |
| RISK-002 |
AIMS-SYS-003 LexisNexis AML |
Model drift — AML model accuracy degrades as new financial crime typologies emerge post-training, leading to false negatives and MLR 2017 monitoring failures |
3 |
5 |
15 |
Mitigate |
M. Osei |
| RISK-003 |
AIMS-SYS-004 Experian Credit |
Model drift — credit scoring accuracy degrades in post-2023 high interest rate environment, producing scores that do not reflect current creditworthiness |
3 |
4 |
12 |
Mitigate |
M. Osei |
| RISK-004 |
AIMS-SYS-004 Experian Credit |
Model bias — credit scoring algorithm produces systematically lower scores for clients from certain geographic areas, violating Consumer Duty fairness obligations |
2 |
5 |
10 |
Mitigate |
M. Osei |
| RISK-005 |
AIMS-SYS-002 MS Copilot |
Hallucination — Copilot generates inaccurate internal regulatory summaries that staff treat as authoritative, leading to misinformed compliance decisions |
3 |
3 |
9 |
Mitigate |
IT Lead |
1
Establish the methodology before assessing any individual system
Marcus documents the risk scoring methodology first — likelihood scale, impact scale, risk appetite thresholds, and treatment categories — in a one-page methodology statement attached to REG-AIMS-FIN-RISK-001. This ensures every system is assessed consistently and that Harcastle (and the FCA) can see that the scores are not arbitrary. The methodology is derived from ISO/IEC TR 24028:2020 and aligned with the firm's existing ISO/IEC 27001 risk approach. One methodology. Applied uniformly. Documented before any scores are assigned.
2
Prioritise by regulatory impact — AML and credit first
Marcus does not assess all six systems simultaneously. He prioritises AIMS-SYS-003 (AML) and AIMS-SYS-004 (credit) because their outputs directly inform regulated decisions with the highest consequence — SAR filings and credit outcomes. These two systems are assessed first, producing risk register entries with scores, treatment plans, and named owners. The remaining four systems are entered with placeholder rows marked "Assessment in progress — completion date: [date]." A partial register with a committed schedule is a compliant response. A blank register is not.
3
Assess all four risk categories for each system
For each system assessed, Marcus works through all four categories: hallucination risk (does this system produce confident incorrect outputs?), bias risk (does this system produce systematically different outputs for different groups?), drift risk (could this system's accuracy degrade over time?), and supply chain risk (what happens if the vendor changes the model?). Not every category applies to every system — the LexisNexis platform does not generate text, so hallucination is low-scored. But every category must be considered and the assessment must be documented, even where the conclusion is low risk.
4
Document treatment plans with deadlines and named owners
Every risk with a score above the acceptance threshold requires a treatment plan. For RISK-001 (AML model bias), the treatment plan is: request LexisNexis's model fairness documentation, establish quarterly alert pattern reviews to detect demographic skew, escalate to Diana for board-level awareness. Each treatment has a completion date and a named owner from the RACI. A risk register without treatment plans is an inventory of problems. A risk register with treatment plans is a governance document.
5
Get board-level sign-off on residual risk acceptance
Harcastle's questionnaire specifically asked whether the risk assessment had been "reviewed and accepted by your board or senior management." ISO 42001 Cl. 6.1 requires the organisation's top management to be involved in risk decisions. Marcus presents the completed risk register to Diana at a documented management review meeting. Diana reviews the Critical-rated risks and formally accepts the residual risk position — or escalates for further treatment. That acceptance is recorded in the meeting minutes. Those minutes are the evidence Harcastle is asking for.
◉
Note
The AI Risk Register (F-G4) and the AI Impact Assessment (F-G5) are related but distinct. The risk register asks: what risks does this AI system create, how likely are they, and what controls manage them? The impact assessment asks: if an AI output is wrong, who is harmed, how severely, and what is the firm's obligation to those individuals? Both are required by ISO 42001. F-G5 covers the impact assessment in the next article in this series. The risk register is the prerequisite — you cannot assess impact without first knowing what the risks are.
REG-AIMS-FIN-RISK-001 — what the fix document contains
REG-AIMS-FIN-RISK-001 — Document Control Block
Document IDREG-AIMS-FIN-RISK-001
TitleAI Risk Register — FCA-Regulated Financial Services Firms
Version1.0
Standard RefsISO/IEC 42001:2023 Cl. 6.1 · ISO/IEC TR 24028:2020 · MLR 2017 Reg. 19 · FCA Consumer Duty PS22/9 · FCA SYSC 13 · ISO 9001:2015 Cl. 6.1
ContentsRisk assessment methodology statement · Risk register template with 12-column structure per risk entry · Four risk category assessment guides (hallucination, bias, drift, supply chain) · Treatment plan template · Board risk acceptance record template · Review schedule guidance
OwnerAI Risk Owner (from RACI-AIMS-FIN-001)
Approval AuthorityChief Executive (SMF1) — board-level risk acceptance required
Review CycleAnnual minimum; triggered upon any new AI system, material change to existing system, or regulatory development
Related DocsREG-AIMS-FIN-SYS-001 · RACI-AIMS-FIN-001 · ASSESS-AIMS-FIN-IMP-001 (F-G5)
Fix Document — Ungated Download
REG-AIMS-FIN-RISK-001 — AI Risk Register Template
ISO/IEC 42001:2023 Cl. 6.1 · ISO/IEC TR 24028:2020 · FCA · MLR 2017 · v1.0
Download Template →
After F-G4: the impact question
Marcus sends the partial risk register to Harcastle — two systems fully assessed, four in progress, methodology documented, board sign-off recorded at the management review meeting. Harcastle's legal team acknowledges it and marks question four as provisionally satisfied.
Question five arrives: "For each AI system that influences decisions affecting individuals — clients, counterparties, or third parties — please provide an AI Impact Assessment documenting the potential consequences for those individuals if an AI output is incorrect, and the controls in place to protect them."
The risk register told Marcus what could go wrong. The impact assessment must tell Harcastle what happens to a real person when it does. Those are different questions — and the second one is harder.
If the LexisNexis AML model falsely flags a Harcastle director as a high-risk individual, that director may be subject to enhanced due diligence, delayed transactions, or a Suspicious Activity Report filed in their name with the National Crime Agency — without their knowledge. What is Whitmore's obligation to that individual? How was the AI's role disclosed? Who reviewed the alert before it became a SAR? Those are the questions the impact assessment must answer.
That is Gap Five.
Remediation Sequence — Finance Series
✓ F-G1AI Policy — POL-AIMS-FIN-001 · Complete
✓ F-G2AI Accountability in SoRs — RACI-AIMS-FIN-001 · Complete
✓ F-G3AI System Register — REG-AIMS-FIN-SYS-001 · Complete
✓ F-G4AI Risk Assessment — REG-AIMS-FIN-RISK-001 · This article
→ F-G5AI Impact Assessment — ASSESS-AIMS-FIN-IMP-001
F-G6AI Data Governance — REG-AIMS-FIN-DATA-001
F-G7Client Disclosure — DOC-AIMS-FIN-DISC-001
F-G8Human Oversight — PROC-AIMS-FIN-HITL-001
F-G9AI Supplier Governance — REG-AIMS-FIN-SUP-001
Section 8 — Quality Gate Record
Quality gate record
The following gates were checked before publication. All 10 gates pass. This record satisfies ISO 9001:2015 Clause 7.5.3 for the article series.
- G1All claims carry named sources: FATF Guidance on AI in AML/CFT 2023 (40% statistic); FCA Enforcement Annual Report 2023 (£568M penalties); ISO/IEC TR 24028:2020 (four risk categories); MLR 2017 Reg. 19; FCA Consumer Duty PS22/9. No fabricated statistics.PASS
- G2ISO/IEC 42001:2023 Clause 6.1 correctly cited and characterised as a mandatory requirement. ISO/IEC TR 24028:2020 correctly characterised as the technical reference for AI trustworthiness methodology. Prescriptive language ("shall") applied correctly throughout.PASS
- G3Scenario consistent with REF-AIMS-FIN-SCEN-001. Marcus Osei (SMF16), Diana Whitmore (SMF1/CEO), Harcastle Group (£28M AUM). Day 5 of questionnaire consistent with series timeline. AML alert indicator in scenario box consistent with LexisNexis system being the primary risk focus. All system IDs consistent with F-G3 register.PASS
- G4MLR 2017 characterised accurately — Reg. 19 monitoring calibration obligation is a legal requirement, not guidance. FCA Consumer Duty characterised as a mandatory FCA requirement. Vendor documentation characterised as insufficient (Caution notice) — not prohibited, but explicitly noted as not substituting for the firm's own assessment.PASS
- G5REG-AIMS-FIN-RISK-001 linked ungated. All document IDs F-G1 through F-G9 consistent with series registry. F-G5 impact assessment correctly identified as a separate document from the risk register.PASS
- G6Article links to predecessor (F-G3) and successor (F-G5). Series navigation present. Dependency between risk register and impact assessment explicitly documented in Section 5 Note and Section 7.PASS
- G7Disclaimer present as standalone section. All named parties identified as fictional. Article noted as educational only, not regulatory, legal, or AML compliance advice.PASS
- G8REG-AIMS-FIN-RISK-001 document control block present with all mandatory fields: ID, version, title, status, standard refs, contents summary, owner, approval authority, review cycle, related documents.PASS
- G9H1 includes primary keyword "risk-assessed" in direct reference to AML AI. Stat bar within first screenful. Schema.org Article JSON-LD with isPartOf series present. Canonical URL set.PASS
- G10IEC 82079-1:2012 hierarchy correct. DANGER not used. WARNING for SM&CR/MLR 2017 personal liability consequence. CAUTION for vendor documentation being insufficient substitute for firm's own assessment. NOTE for risk register vs. impact assessment distinction. ITIL 4 notice covers Risk Management and Change Enablement — both substantively mapped, not generic.PASS