Part 1 answers the questions in paragraphs 14 and 48 of the Consultation Paper. Part 2 proposes wording on eight points of the proposed Guidance Note: authority to restrict or suspend a use case; evidence that the ability to suspend works; contracts with AI providers; group-operated systems; records of overrides, suspensions and versions; re-examining decisions made while an error went undetected; insurance and reinsurance; and a short annex of illustrative evidence (proposed Annex G).
Submitted by email to policy@bma.bm.
Consultation Paper: Bermuda Monetary Authority, Consultation Paper, 14 August 2026
The full text of our submission follows.
Full text of the submission
Comments on the Consultation Paper "The Responsible Use of Artificial Intelligence in Financial Services in Bermuda" (14 August 2026) and the proposed Guidance Note
To: Bermuda Monetary Authority (policy@bma.bm)
From: Fijishi
Date: 19 October 2026
Fijishi is an independent organisation based in Pune, India, that works on the governance of AI decisions.
These comments respond to the questions in paragraphs 14 and 48 of the Consultation Paper and propose wording for the proposed Guidance Note. Because the two Parts are numbered separately, "CP para" refers to Part A and "GN para" to Part B. Where we propose to amend existing text, the existing words are quoted as they stand and the words we propose to add are underlined.
Our comments concern one question the Note raises in several places: what a regulated entity could show, as evidence, that its human oversight of a material AI use case works in practice.
Part 1 - Answers to the Authority's questions
1. Whether the proposed GN appropriately explains how existing requirements apply to AI (CP para 48(a); CP para 14, clarity)
The proposed GN explains clearly how existing frameworks apply, and its use-case approach (GN para 9) gives entities a workable starting point. One area could be made clearer. The GN refers to the ability to restrict or suspend use (GN paras 50, 58 and 61) and to "appropriate and effective override points and kill-switches for high-risk systems" (GN para 34(g)). It says less about what an entity could show to demonstrate that these controls work. GN para 115 lists "human oversight arrangements" among the information the Authority may request, without illustrating what that evidence could contain.
The proposals in Part 2 offer illustrative, non-mandatory examples of such evidence, consistent with GN paras 4 and 5.
2. Whether the proposed materiality and risk assessment approach is workable across sectors (CP para 48(b); CP para 14, proportionality)
The approach is workable, and evidence of oversight can be scaled to the same tiers that GN para 12 describes:
- For a lower-risk use, such as an internal productivity tool, a named owner and a record of the use, at category or platform level where GN para 33 allows.
- For a material decision-support tool, in addition: documented authority to restrict or suspend the use case, the ability to do so without halting unrelated processes, and a record of overrides and suspensions.
- For a higher-impact or more autonomous use case, in addition: a suspension time that has been tested, provider contract terms or compensating controls, and a way to re-examine decisions affected by an undetected error.
3. Whether further sector-specific examples would be useful, and whether any sector-specific or emerging risks need further consideration (CP para 48(c); CP para 14)
A reinsurance example in Annex D would be useful, given the place of reinsurance in Bermuda's market (CP para 10). We propose one in Part 2, point 7.
Two emerging risks could receive further consideration:
- Changes to underlying models by providers. CP para 31 notes that "common vendors may change underlying models without adequate notice". Such changes affect an entity's ability to link a decision to the version that produced it, and to know when retesting is needed (GN para 55). Part 2, points 3 and 5, address this.
- Group-operated systems. Where an AI system is operated at group level (CP para 26; GN para 95), a Bermuda entity may need to stop its own use of the system without depending on decisions taken elsewhere in the group. Part 2, point 4, addresses this.
Point 1. Authority to restrict or suspend a use case - GN paras 32 and 34(g)
GN para 30(b) asks entities to "assign accountable ownership", and GN para 32(a) lists "the purpose and owner" among the information relevant to material or higher-risk use cases. Ownership alone may not show who can stop a use case quickly, including when the owner is unavailable.
Proposed new item in GN para 32, after item (a):
Proposed wording for GN para 34(g):"the person, and a deputy, with documented authority to restrict or suspend the use case;"
Point 2. Evidence that the ability to suspend works - GN paras 50, 58 and 60"appropriate and effective override points and kill-switches for high-risk systems, and the persons with authority to use them (see paragraph 32)."
GN para 50 refers to "the ability to restrict or suspend use where outcomes appear unreliable, outside approved use or inconsistent with applicable obligations", and GN para 58 lists "suspension rights" among the forms of oversight. GN para 60 provides that oversight relied upon as a control may include "appropriate authority, competence, information, intervention time, escalation routes and awareness of automation bias or undue reliance on AI outputs." An ability to suspend is a capability, and it can best be shown by exercising it.
Proposed new sentence at the end of GN para 60:
Point 3. Contracts with AI providers - GN para 90, with CP paras 31 and 33"Where the ability to restrict or suspend use is relied upon as a control for a material or higher-risk use case, an entity could evidence it by showing that it can suspend one use case, or one type of decision, without halting unrelated processes; that it has set a target time for a suspension to take effect; and that it has tested the suspension at an interval it sets, for example at least annually for higher-impact use cases, with the result recorded."
CP para 31 notes that "common vendors may change underlying models without adequate notice", and CP para 33 recognises that smaller entities and Bermuda subsidiaries "may have limited negotiating leverage". GN para 90 already lists terms that contracts may address "where proportionate and reasonably achievable". Three additions would support the ability to suspend use and to link decisions to versions, while keeping that qualifier and the compensating controls in GN para 92.
Proposed wording for the first sentence of GN para 90:
Point 4. Group-operated systems - GN para 95"Contracts and assurance arrangements may, where proportionate and reasonably achievable, address permitted use, service levels, data access and security, the entity's right and practical means to restrict or suspend its use of the system without the provider's consent, advance notice of material model or feature changes, incident notification, audit or assurance rights, regulatory access, access to records of the system's outputs and changes relevant to the entity, during the contract and after it ends, subcontracting, continuity, portability, termination and exit."
Where a system is operated at group level, a Bermuda entity may need to stop its own use without halting the system for the rest of the group.
Proposed wording for GN para 95:
Point 5. Records of overrides, suspensions and versions - GN para 111, with GN para 55"Where an AI system is developed, governed or operated at group level, the BMA-regulated entity should retain sufficient local oversight to understand the purpose, limitations and risk profile of the AI use case and to assess whether its use is consistent with the entity's regulatory obligations, risk appetite and business context, and, for material use cases, the ability to restrict or suspend the entity's own use of the system without halting the system's use elsewhere in the group."
GN para 111 provides that relevant records may include information to "investigate incidents or unexpected outcomes" and "reconstruct material decisions", including information on "human review, overrides, reliance and the basis for the final decision or action". A decision can be reconstructed only if the version of the system that made it is known, and a record is useful for investigation only if later changes to it can be seen. GN para 55 already treats changes to "the model, data, provider, configuration" as relevant changes.
Proposed wording for GN para 111:
Point 6. Re-examining decisions made while an error went undetected - GN para 85"For material or higher-risk AI use cases, and as appropriate under existing obligations, relevant records may include information sufficient to understand system behaviour, investigate incidents or unexpected outcomes, support internal assurance, evidence escalation or reconstruct material decisions including, where relevant, information on material inputs and outputs, the model or system version in use, human review, overrides, suspensions and restarts (who, when, why and which decisions were affected), reliance and the basis for the final decision or action. Such records could be kept so that any later alteration or deletion is itself recorded, and could be supported by a dated record of model, version and configuration changes, consistent with paragraph 55, so that each material decision can be linked to the version that produced it."
GN para 85 provides that root-cause analysis "may consider AI-specific factors such as data drift, model or vendor updates, emergent behaviour, human over-reliance, changes in prompts or retrieval sources, and interaction between connected systems." Where such a factor went undetected for a period, decisions made during that period may also have been affected, including for parties who have not raised a complaint.
Proposed new sentence at the end of GN para 85:
The version record proposed in point 5 is what makes such identification practical."Where an incident reveals an error that went undetected for a period, the entity should consider how decisions made during that period can be identified and, where appropriate, re-examined, having regard to their effect on customers, policyholders, counterparties or other affected parties."
Point 7. Insurance and reinsurance - CP para 24 and Annex D
CP para 24 notes that AI may influence activities such as underwriting, catastrophe modelling, reserving and risk transfer, and that "institutional counterparties, boards, auditors and supervisors must retain sufficient information to independently challenge those outcomes". The insurance and reinsurance row of Annex D lists, among possible heightened considerations, "treatment of policyholders, cedants, beneficiaries and counterparties" and "reliance on third-party data, models or platforms".
Proposed additions to the "Possible heightened considerations" column of that row:
"For a reinsurer's own use of AI, for example in underwriting support, catastrophe and exposure modelling or reserving, the evidence of oversight illustrated in Annex G, scaled to the materiality of the use case."
The second addition concerns the reinsurer's own assessment of the risk it accepts, in the spirit of CP para 24. It does not suggest that the GN should apply to the reinsured insurer."Where AI used by a reinsured insurer materially shapes the risk ceded, the reinsurer could consider asking for proportionate information about which decisions the AI shapes, who can restrict or suspend its use, and how changes to it are recorded."
Point 8. A short annex of evidence examples - new Annex G, and GN para 115
GN para 115 lists the information the Authority may request, including "human oversight arrangements". A short annex of illustrative evidence, tiered in the same way as GN para 12 and Annex C, would help entities of different sizes see what proportionate evidence could look like.
Proposed wording in GN para 115:
The proposed Annex G follows. All of its text is new."... monitoring and drift reports, human oversight arrangements (illustrative examples of evidence are set out in Annex G), incident records ..."
Proposed Annex G - Illustrative evidence of human oversight
The examples below are illustrative and non-exhaustive, consistent with GN paras 4 and 5. A regulated entity may use other evidence that achieves the same outcome, and may use its own tiering in place of the one shown (GN para 40).
| Use-case intensity (see GN para 12 and Annex C) | Evidence an entity could hold |
| Lower-risk, for example an internal productivity tool | A named owner; approved-use boundaries; a record of the use case, at category or platform level where GN para 33 allows. |
| Material, for example a decision-support tool affecting regulated outcomes | As above, plus: a person and a deputy with documented authority to restrict or suspend the use case; the ability to suspend it without halting unrelated processes; a record of overrides, suspensions and restarts; a dated record of model, version and configuration changes. |
| Higher-impact or more autonomous | As above, plus: a target time for a suspension to take effect, tested at an interval the entity sets (for example at least annually), with the result recorded; provider contract terms on suspension, advance notice of changes and access to records, or compensating controls under GN para 92; a way to identify and re-examine decisions made while an error went undetected. |
Akhil Sharma and Preethi Sharma, Fijishi
Fijishi Enterprises Private Limited
protocol@fijishi.com