Marcus had been methodical. Six gaps closed in twelve days. AI policy written. Accountability mapped. Systems registered. Risks assessed. Data governance documented. Each answer to the Harcastle questionnaire had been hard-won, but they had been won.
He had two questions left — then he opened the wrong folder.
It was Jade's Q3 portfolio review for Harcastle Group. Sent three weeks ago. Before the ChatGPT suspension. He skimmed the executive summary, the asset allocation commentary, the ESG screening narrative. Well-structured. Confident. Professionally presented.
Then he searched the document for the word "AI."
Zero results.
He searched for "automated." Zero. "Generated." Zero. "Tool." Zero.
A pause on the Teams call. The kind that tells you the answer before the words arrive.
Marcus looked back at the Harcastle questionnaire. Question 7.
He could answer nine of ten questions now. But question 7 was not about what he had built in the last twelve days. It was about what had already been sent.
The Gap: What ISO 42001 Annex A.2 Requires
ISO/IEC 42001:2023 Annex A includes a set of controls that responsible AI system operators are expected to implement. Control A.2 — Responsible AI by Design — contains a disclosure requirement that most firms overlook because it does not appear in the main clause structure. It sits in the annex. But it is not optional.
The disclosure obligation under Annex A.2 has three components:
| Component | Requirement | Obligation Level | FCA Parallel |
|---|---|---|---|
| A.2.2 | Affected parties shall be informed when AI has been used in processes that materially affect them | SHALL | Consumer Duty — Acting in Good Faith (PRIN 12) |
| A.2.3 | The form and timing of disclosure shall be documented in the AI management system | SHALL | SYSC 3.2 — Record-keeping obligations |
| A.2.6 | Where AI-generated content is presented to clients, the AI origin shall be identifiable or otherwise disclosed | SHALL | COBS 4.2 — Fair, clear, not misleading |
| A.6.1 | Clients should be given opportunity to request human review of AI-assisted outputs that affect their interests | SHOULD | Consumer Duty — Products & Services / Price & Value |
Whitmore & Associates had none of these. No disclosure language in their report templates. No documented process for informing clients when AI had been used. No record of who had decided the current approach was acceptable, because no one had ever consciously decided it.
Why This Matters Beyond Compliance
The disclosure requirement exists for a reason that goes beyond box-ticking. AI systems can introduce errors, biases, and confident-sounding inaccuracies. A client who does not know AI was involved cannot ask the right questions. They cannot request human review. They cannot apply appropriate scrutiny to commentary that may have been generated, not considered.
Jade's Q3 report for Harcastle contained ChatGPT-generated commentary on ESG screening criteria. That commentary referenced a regulatory standard — SFDR Article 8 classification thresholds — with complete confidence and marginal inaccuracy. The numbers were close, but not current. Jade had edited the prose but not checked the regulatory reference.
Harcastle had read it. Relied on it. Made no complaint, because they had no reason to know the source.
The Before and After: What a Disclosure Looks Like
Marcus spent forty-five minutes on Day 13 pulling every client-facing document template Whitmore used: the quarterly portfolio review, the annual suitability report, the rebalancing recommendation letter, and the ESG screening summary. None of them contained disclosure language.
He built a disclosure clause. Then he compared what Harcastle had received against what they would receive from Day 14 onward.
Executive Summary
Your portfolio delivered a return of 3.2% in Q3, outperforming the benchmark by 0.8%. The ESG screening applied across your holdings continues to align with SFDR Article 8 classification thresholds, with exposure to excluded sectors at 0.4% — within tolerance…
[No AI disclosure. No reference to tools used. No client rights statement.]
Prepared by: Jade Nwosu, Relationship Manager
Reviewed by: [blank]
Date: 14 September 2024
Executive Summary
Your portfolio delivered a return of 3.2% in Q3, outperforming the benchmark by 0.8%. The ESG screening applied across your holdings continues to align with SFDR Article 8 classification thresholds, with exposure to excluded sectors at 0.4% — within tolerance…
Sections of this report were drafted with the assistance of AI language tools (Microsoft Copilot, AIMS-SYS-002). All AI-assisted content was reviewed and approved by the named Relationship Manager prior to issue. If you would like any section reviewed by a qualified adviser without AI assistance, please contact us directly.
Authorised by: Marcus Osei, Head of Compliance (SMF16)
Building DOC-AIMS-FIN-DISC-001: The Client Disclosure Framework
Marcus did not just update the Q3 template. He built a disclosure framework that could be applied consistently across all client-facing outputs, regardless of which AI system was used or which member of staff had prepared the document.
The framework defined three elements: the disclosure template language, the threshold at which disclosure is required, and the sign-off process. Each is documented below as it was embedded into Whitmore's templates.
What Harcastle's Question 7 Actually Asked
Question 7 had three sub-parts. Marcus addressed each one.
| Sub-question | Marcus's Response | Evidence Document |
|---|---|---|
| Do you disclose AI use to clients? | Yes — from Day 14, all client documents use the disclosure clause in DOC-AIMS-FIN-DISC-001. Prior reports are under review. | DOC-AIMS-FIN-DISC-001 v1.0 |
| What form does disclosure take? | A standardised disclosure block embedded in the document footer section, identifying the AI system by AIMS-SYS reference, the sections affected, and the human reviewer. | Template library v1.0 |
| Can clients object to AI-assisted work? | Yes — the disclosure clause includes a client opt-out provision. Clients may request a non-AI version of any document. Response SLA: 5 business days. | PROC-AIMS-FIN-HITL-001 (forthcoming) |
The third sub-answer referenced a document that did not yet exist: PROC-AIMS-FIN-HITL-001, the Human-in-the-Loop oversight procedure. Marcus had written a placeholder. He had eight days left and two gaps to close.
Marcus sent the updated framework to Diana at 11:42pm on Day 13. No preamble. Just the document.
She replied at 7:08am.
Marcus had already drafted the note. A brief, factual letter to Harcastle acknowledging that their Q3 review had included AI-assisted content, confirming that no material inaccuracies had been identified upon human review, and confirming that all future reports would carry full disclosure.
He had not sent it. He wanted legal sign-off first. But the draft existed.
One problem at a time. Two gaps left. Seven days on the clock.