UNUS London · ISO/IEC 42001:2023 Finance Series · ART-AIMS-FIN-G005 v1.0 · FCA · UK GDPR · MLR 2017 · Consumer Duty
The question that makes the risk register feel abstract
Day seven. Marcus has been working through the Harcastle questionnaire for over a week. He has four documents in Harcastle's hands: the AI Policy, the RACI, the AI System Register, and a partial AI Risk Register with the two highest-risk systems fully assessed. Harcastle has provisionally accepted each one.
Question five lands differently from the others.
Harcastle Group
"For each AI system in your register that influences decisions affecting individuals — including clients, counterparties, or any third parties — please provide an AI Impact Assessment documenting: the categories of individuals affected, the types of harm they could experience if an AI output is incorrect, the severity and reversibility of those harms, and the controls in place to protect them."
Marcus Osei
He reads it twice. The risk register told him what could go wrong with the LexisNexis AML platform — model bias, false positives, inflated alert rates. It scored the risk. It assigned a treatment owner. But it didn't ask what happens to a specific person when that risk materialises.
Marcus Osei
He thinks about it carefully. The AML platform had flagged three alerts this quarter. Two were investigated and cleared. One had been escalated to the MLRO and filed as a Suspicious Activity Report with the National Crime Agency. The individual named in that SAR was a director of a Harcastle subsidiary. If the model had produced a false positive — if the alert was generated by a biased pattern rather than genuine suspicious behaviour — a real person's name was now in an NCA database. Without their knowledge. Without any recourse. Without a word of explanation from anyone at Whitmore.
Diana Whitmore
"Marcus — are you telling me we may have filed a SAR based on an AI flag that nobody formally assessed for bias?"
Marcus Osei
"I'm telling you we don't have a document that proves we did."
This is what the impact assessment gap looks like when the stakes become concrete. The risk register identified the possibility. The impact assessment should have documented what happens to an actual human being when that possibility becomes reality — and what the firm's obligations to that person are.
The risk register is the firm's internal view of risk. The impact assessment is the individual's view of harm. Both are required. They are not the same document.
⚠
Warning
Under UK GDPR Article 22, individuals have rights in relation to solely automated decisions that produce significant effects on them. Where an AI-generated AML alert leads to a Suspicious Activity Report being filed — even following human review — and that alert was generated by a model with unassessed bias, the firm may be unable to demonstrate compliance with the Article 22 requirement to ensure appropriate safeguards for individuals subject to automated processing. ISO/IEC 42001:2023 Clause 6.1.2 requires the impact assessment to be completed before AI systems are deployed in contexts that affect individuals — not after a SAR has already been filed.
What ISO/IEC 42001:2023 requires: the individual at the centre
Clause 6.1.2 is the clause that puts people — not processes — at the centre of AI governance. It requires organisations to consider, before deploying AI in contexts that affect individuals, what the consequences of AI errors would be for those people.
ISO/IEC 42001:2023 — Clause 6.1.2 · AI System Impact Assessment
The organisation shall conduct an impact assessment to understand the potential impacts of the AI system on individuals, groups of individuals, and society. The assessment shall address the types of impact, the severity of impact, and the reversibility of impact. The organisation shall document the results of the impact assessment and maintain this documented information.
Source: ISO/IEC 42001:2023 Clause 6.1.2. British English retained per ISO/IEC Directives Part 2. The clause is supplemented by ISO/IEC TR 24028:2020 Annex D — Harm Taxonomy for AI Systems.
Three dimensions of impact are required: type, severity, and reversibility. This third dimension — reversibility — is what makes the AML SAR scenario so acute. A credit decision that is wrong can, in principle, be reversed. A SAR filed with the National Crime Agency cannot. Once filed, it enters a secure database. The subject is never told. The subject cannot challenge it. The subject has no right of erasure under UK GDPR — because SAR records are exempt from data subject rights to protect the integrity of financial crime investigations.
The irreversibility of the harm is precisely what the impact assessment must document — before the AI system is deployed in that context, not after a filing has already occurred.
ISO 42001 Annex A — the harm taxonomy
ISO/IEC 42001 Annex A, supplemented by ISO/IEC TR 24028:2020, identifies five categories of harm that AI systems can cause to individuals. All five are relevant to FCA-regulated AI systems:
Financial Harm
The individual suffers economic loss as a direct result of an incorrect AI output. A credit scoring model that underscores a creditworthy individual causes them to be denied credit — or to receive credit at a higher interest rate than their profile warrants. A false AML flag that delays a legitimate transaction causes a client financial disruption. Financial harm is often the most quantifiable and the easiest to demonstrate — but not always the most severe.
Reputational Harm
The individual's standing — professional, social, or commercial — is damaged by an incorrect AI output. A false SAR naming a corporate director as a person of interest in a financial crime investigation creates reputational harm that persists even if the investigation is never pursued. The individual may never know the SAR was filed, but if it surfaces during a background check or a future regulatory application, the damage is real and — given SAR exemption from data subject rights — largely irreversible.
Psychological Harm
The individual experiences distress, anxiety, or loss of autonomy as a result of an AI-influenced decision. A client who is denied a loan because an AI credit model incorrectly assessed their risk profile, and who receives no explanation, may experience significant psychological harm — particularly if the denial affects a major life event such as a home purchase. Under the Consumer Duty, the FCA requires firms to consider the emotional impact of their communications and decisions on clients.
Discrimination
A group of individuals with a protected characteristic — ethnicity, age, disability, gender — receives systematically worse outcomes because an AI model has encoded historical bias. In AML screening, this may manifest as higher alert rates for certain nationalities or transaction origins. In credit scoring, it may manifest as lower scores for certain postcodes that correlate with demographic characteristics rather than actual credit behaviour. Discrimination is simultaneously an equality law concern, a Consumer Duty concern, and an ISO 42001 impact concern.
Loss of Rights or Legal Exposure
The most severe category. The individual loses access to a right — the right to access financial services, the right to conduct transactions, the right to a fair credit decision — as a result of an AI output. In AML contexts, the SAR scenario represents loss of rights in its most acute form: the individual is effectively placed under suspicion by an automated process, without knowledge, without recourse, and without the firm having assessed whether the AI that generated the flag was operating correctly.
◈
ITIL 4 — Relationship Management Practice · Secondary: Service Design
In ITIL 4, the Relationship Management practice requires organisations to understand and manage the expectations and experiences of stakeholders — including the individuals affected by services. The AI Impact Assessment is the instrument that makes individual stakeholders visible in the governance framework. Service Design provides a secondary lens: AI systems that affect individuals should be designed with safeguards from the outset, not retrofitted with impact assessments after deployment. ASSESS-AIMS-FIN-IMP-001 includes a pre-deployment checklist aligned to the Service Design practice for use when any new AI system is added to the register.
Section 3 — The UK GDPR & FCA Position
Two frameworks, one absent document
UK GDPR Article 22 — automated decision-making
Article 22 of the UK GDPR provides individuals with the right not to be subject to a decision based solely on automated processing — including profiling — that produces a significant effect on them. Where automated processing is used, the data controller must implement appropriate safeguards and ensure the individual has a meaningful way to obtain human intervention, express their point of view, and contest the decision.
AML screening via the LexisNexis platform involves automated risk scoring — a form of automated processing. Where a high-risk score leads to a SAR filing, the effect on the individual is significant. The question the impact assessment must answer is: has the firm implemented appropriate safeguards, and does the individual have a meaningful right to contest? In most firms, the answer to both is: they have never considered it.
◆
Caution
The SAR exemption from data subject rights — which prevents AML subjects from knowing a SAR has been filed — does not exempt the firm from its obligation to conduct an impact assessment before deploying the AI system. The exemption governs the subject's rights after the fact. The impact assessment obligation governs the firm's governance before the system is deployed. The two provisions operate independently. Pointing to the SAR exemption does not discharge the Article 22 safeguarding obligation or the ISO 42001 Clause 6.1.2 impact assessment obligation.
FCA Consumer Duty — harm prevention obligations
The Consumer Duty's Principle 12 requires firms to act to deliver good outcomes for retail clients across four outcome areas: products and services, price and value, consumer understanding, and consumer support. Where an AI system is influencing product access decisions (credit scoring), the Consumer Duty requires the firm to have understood and mitigated the risk that the AI produces poor outcomes. An impact assessment is the document that demonstrates this understanding.
The FCA's expectations under the Consumer Duty have moved from principle to scrutiny. Firms are now asked, in supervisory visits and data requests, to evidence how they have tested for and monitored AI-influenced consumer outcomes. An impact assessment is not a nice-to-have. It is the evidence the FCA is asking to see.
The anatomy of this nonconformance
Gap F-G5 — Nonconformance Taxonomy
Gap IDF-G5
ISO ClausesISO/IEC 42001:2023 Clause 6.1.2 (AI System Impact Assessment) and Annex A (AI system harm categories)
SeverityCritical — AI systems influencing regulated decisions affecting individuals with no documented impact assessment; UK GDPR Article 22 safeguarding obligation potentially unmet; FCA Consumer Duty evidence gap
Current StateNo AI Impact Assessment exists for any system in REG-AIMS-FIN-SYS-001. The AML screening platform has generated alerts that became SARs — with no prior assessment of the harm that a false positive SAR would cause to the named individual. The credit scoring platform has informed credit decisions with no assessment of the harm to individuals incorrectly scored.
Regulatory ParallelUK GDPR Art. 22 — automated processing safeguards; FCA Consumer Duty PS22/9 Principle 12; FCA SYSC 4 — adequate arrangements to prevent harm; Equality Act 2010 — indirect discrimination from biased AI
Commercial RiskHarcastle questionnaire question five unanswerable. Institutional mandate at risk. FCA Consumer Duty examination finding if AI-influenced consumer harm is discovered without prior assessment.
Required ArtefactASSESS-AIMS-FIN-IMP-001 — AI Impact Assessment Framework for FCA-Regulated Financial Services Firms
What Marcus built — the AML impact assessment
Marcus starts with the highest-consequence system: AIMS-SYS-003, the LexisNexis AML platform. He works through the five harm categories systematically, documenting the assessment for the firm's specific use of this system, with these clients, in this regulatory context.
The completed impact assessment — AIMS-SYS-003
System
AIMS-SYS-003 — LexisNexis AML AI Screening Platform
Affected Individuals
Clients of Whitmore & Associates and their associated parties (directors, beneficial owners, counterparties) subject to AML transaction monitoring and screening
Harm Types
Financial
Reputational
Psychological
Discrimination
Loss of Rights
Primary Harm Scenario
False positive SAR filing: AI model generates an incorrect high-risk alert based on biased pattern recognition. Human reviewer escalates. MLRO files SAR with NCA. Named individual enters financial crime intelligence database without knowledge, without recourse, and without right of erasure. Reputational and legal harm: irreversible.
Secondary Harm Scenario
Transaction delay / account freeze: AI alert triggers enhanced due diligence, delaying legitimate transactions. Financial and psychological harm. Reversible with remediation — but not before the client experiences disruption.
Severity
Critical (5/5) for false SAR scenario — irreversible harm to named individual in a law enforcement database. Moderate (3/5) for transaction delay scenario — reversible with action.
Reversibility
SAR filing: Irreversible — SAR records exempt from UK GDPR data subject rights (POCA 2002 Sch.7). Transaction delay: Reversible — cleared through MLRO review and client communication.
UK GDPR Art. 22 Status
AI alert plus human MLRO review constitutes a partially automated decision process. Human review gate is an Article 22 safeguard — but only if the MLRO is genuinely reviewing alert quality, not rubber-stamping AI outputs. Current status: safeguard exists in procedure; not evidenced in practice. Action: implement MLRO review quality log (see F-G8 — Human Oversight).
Protective Controls
MLRO human review gate before any SAR filing; quarterly alert pattern review (RISK-001 treatment plan); request for LexisNexis fairness documentation (in progress)
Residual Harm Accepted By
Diana Whitmore — CEO / SMF1 — at management review meeting 14 August 2026
1
Identify every AI system that makes or influences decisions affecting individuals
Not all AI systems require an impact assessment. Microsoft Copilot helping Jade draft an internal email does not affect an individual beyond Jade herself. LexisNexis AML screening, the Experian credit model, and any AI system generating client-facing communications all influence decisions that have real consequences for identifiable individuals. The first step is identifying which systems in REG-AIMS-FIN-SYS-001 require an impact assessment — and for which ones a lighter-touch assessment is sufficient. Prioritise by potential harm severity, consistent with the approach used for the risk register in F-G4.
2
Document who is affected — specifically, not generically
The standard requires assessment of impact on individuals, groups, and society. This means naming the categories of people the AI touches: clients, their associated parties (directors, beneficial owners), counterparties, third parties named in transactions. "Clients" is insufficient — the AML platform also screens counterparties who are not Whitmore's clients. Specificity matters: a vague impact assessment is as useless as no impact assessment when a regulator or institutional client asks to see it.
3
Map every harm scenario against the five harm categories
For each affected individual category, work through all five harm types: financial, reputational, psychological, discrimination, loss of rights. For each type, ask: could an incorrect AI output cause this harm? How severe would it be? Can it be reversed? The irreversibility question is often the most important — and the most neglected. An impact assessment that categorises all harms as "reversible" without examining each one has not been conducted rigorously and will not withstand scrutiny from Harcastle or the FCA.
4
Map existing controls against each harm scenario
Existing controls — the MLRO review gate, the quarterly alert pattern review committed in the risk register treatment plan — reduce the likelihood or severity of harms. Each control is documented against the harm it mitigates, with an honest assessment of whether the control is currently operating effectively. Where it is not — where the MLRO review is in procedure but not evidenced in practice — this is noted as a gap and connected to the relevant gap in the series (F-G8 — Human Oversight). The impact assessment is not a tick-box exercise. It creates accountability chains across the entire AIMS framework.
5
Document residual harm acceptance and build the pre-deployment gate
After controls are applied, some residual harm possibility remains. Diana must formally accept this position for the systems she is responsible for as SMF1. That acceptance is recorded, dated, and signed. And — most importantly for the future — the impact assessment template is established as a mandatory pre-deployment requirement. From this point, no AI system enters the register without a completed impact assessment. The harm that the AML platform created, undocumented, for eighteen months is now structurally impossible for any future system.
◉
Note — The relationship between F-G5 and F-G8
The impact assessment for the AML platform identified a specific control gap: the MLRO human review gate exists in procedure but is not evidenced in practice. This gap is closed by F-G8 — Human Oversight, which builds the human review log and quality record that turns a procedural control into documented evidence. The two gaps are interdependent. F-G5 identifies what human oversight should be protecting against. F-G8 builds the mechanism that makes that protection visible. Neither is complete without the other.
ASSESS-AIMS-FIN-IMP-001 — what the fix document contains
ASSESS-AIMS-FIN-IMP-001 — Document Control Block
Document IDASSESS-AIMS-FIN-IMP-001
TitleAI Impact Assessment Framework — FCA-Regulated Financial Services Firms
Version1.0
Standard RefsISO/IEC 42001:2023 Cl. 6.1.2 & Annex A · ISO/IEC TR 24028:2020 Annex D · UK GDPR Art. 22 · FCA Consumer Duty PS22/9 · Equality Act 2010 · MLR 2017
ContentsSystem selection guide (which systems require full vs. light-touch assessment) · Individual category mapping template · Five harm category assessment matrix · Reversibility classification guide · Control mapping template · Residual harm acceptance record · Pre-deployment checklist for new AI systems
OwnerAI Ethics Reviewer (from RACI-AIMS-FIN-001)
Approval AuthorityChief Executive (SMF1) — residual harm acceptance required
Review CycleAnnual; triggered upon new AI deployment, material system change, or regulatory development in AI/data protection
Related DocsREG-AIMS-FIN-SYS-001 · REG-AIMS-FIN-RISK-001 · PROC-AIMS-FIN-HITL-001 (F-G8)
Fix Document — Ungated Download
ASSESS-AIMS-FIN-IMP-001 — AI Impact Assessment Framework
ISO/IEC 42001:2023 Cl. 6.1.2 · UK GDPR Art. 22 · FCA Consumer Duty · v1.0
Download Template →
After F-G5: Harcastle's sixth question
Marcus sends the completed impact assessments for the AML and credit scoring systems to Harcastle, along with the framework document. He is transparent about the control gap identified — the MLRO review is procedural but not yet evidenced — and notes it as a committed remediation action.
Harcastle marks question five as provisionally satisfied. Eight days into the questionnaire. Five gaps identified, five documents built. Four gaps remaining. Twelve days left.
Question six arrives: "How does your firm ensure that client personal and financial data processed through AI systems is governed in accordance with your data protection obligations? Please provide your AI data governance framework."
Marcus thinks about Jade's ChatGPT use. He thinks about the data that has been flowing through the AML platform and the credit scoring API. He thinks about the impact assessment entry that noted the Experian model processes client financial data — and the fact that there is still no data governance document for any of the AI systems in the register.
That is Gap Six.
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 · Complete
✓ F-G5AI Impact Assessment — ASSESS-AIMS-FIN-IMP-001 · This article
→ 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: FCA Financial Lives Survey 2022 / HM Treasury de-banking review 2023 (cost statistic); UK GDPR Art. 22 (automated decision rights); UK GDPR Art. 83(4) / ICO enforcement guidance 2024 (£17.5M fine maximum); ISO/IEC TR 24028:2020 Annex D (harm taxonomy). No fabricated statistics.PASS
- G2ISO/IEC 42001:2023 Clause 6.1.2 and Annex A correctly cited and characterised. UK GDPR Article 22 accurately characterised — automated processing with human review constitutes a partially automated decision; Article 22 safeguarding obligation applies. SAR exemption correctly characterised as applying to data subject rights after filing, not to the firm's pre-deployment assessment obligation.PASS
- G3Scenario consistent with REF-AIMS-FIN-SCEN-001. Marcus Osei (SMF16), Diana Whitmore (SMF1), Harcastle Group (£28M AUM). Day 7 consistent with series timeline. SAR filed against Harcastle director — new narrative element consistent with LexisNexis AML platform (AIMS-SYS-003) as previously established in F-G3 and F-G4. All system IDs consistent.PASS
- G4UK GDPR Art. 22 characterised accurately — right applies to solely automated decisions; human review constitutes a safeguard. Caution notice correctly distinguishes SAR exemption (data subject rights) from impact assessment obligation (firm's pre-deployment duty). Consumer Duty characterised as mandatory FCA requirement. Equality Act characterised as applicable to discrimination harm — factually correct.PASS
- G5ASSESS-AIMS-FIN-IMP-001 linked ungated. All document IDs F-G1 through F-G9 present and consistent with series registry. F-G8 dependency correctly identified in Note notice.PASS
- G6Article links to predecessor (F-G4) and successor (F-G6). Series navigation present. F-G5 / F-G8 interdependency explicitly documented in Note notice — preventing the reader from treating F-G5 as self-contained.PASS
- G7Disclaimer present as standalone section. All named parties identified as fictional. References to LexisNexis and Experian are illustrative. Article noted as educational only, not legal, regulatory, data protection, or AML advice.PASS
- G8ASSESS-AIMS-FIN-IMP-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 "Who Is Harmed" — primary keyword phrase for impact assessment content. 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 UK GDPR Art. 22 pre-deployment obligation. CAUTION for SAR exemption misconception. NOTE for F-G5/F-G8 interdependency. ITIL 4 notice covers Relationship Management and Service Design — both substantively mapped to the impact assessment obligation and pre-deployment gate.PASS