Cayman Considerations forAnswerable AI

How to read it

• Five minutes. Read section 1. It is the whole standard on one page: the substance rule, the Answerable Test, the four tiers and the red lines.
• Thirty minutes. Add sections 2, 3, 5 and 6. You will know why Cayman needs its own standard, the seven principles and the law that already backs each one, what each tier requires, and how to build a clock file.
• In full. Everything, including the law table (section 4), leak-proof design (section 7), your industry (section 8), what to do when something goes wrong (section 9), the hub translation (section 10) and the Council proposal (section 11).

The tiers, the substance rule and the red lines are our recommendations; the legal duties cited keep their own scope and exceptions. Five labels are used throughout. Law is a quoted provision of an Act or Regulations, with its section. Regulator rule is a rule a Cayman regulator has issued under its statute (CIMA rules bind the entities in scope, but a breach is not itself an offence: Monetary Authority Act [2020 Revision], s.34(6)). Regulator guidance is what a regulator has said it expects. Government statement is what government has said in Parliament or in a release. Our recommendation is Perennos's position, and it carries no citation because none exists yet.

2. Why Cayman needs its own standard, and where government is today

"Paste it into a chatbot and it leaves the building. Ban it, and your team is typing letters at nine at night."

— Perennos

Trust is the product. A financial, legal, medical and insurance hub sells one thing before anything else: the confidence that what you give it stays where you put it, and that someone here answers for what is done with it. Cayman's whole regulatory architecture is built on that. Economic substance, fit and proper tests, a board that knows its business, records kept in the Islands and open to the regulator. AI does not change the promise. It changes how easy it is to break without noticing. A trust officer on Shedden Road pastes a client letter into a free assistant at 4.50 on a Friday afternoon to tidy the English. Nobody meant harm. The letter is now on a server in another country, under terms nobody read, and the firm has no record it happened.

The two bad answers are the two in the epigraph. Ban the tools and your people use them anyway, on their phones, with no record. Wave them through and the building leaks. The answer is a standard: use the tools, with substance here.

The law already applies. There is no Cayman AI Act. That does not mean AI is unregulated here. It means AI is regulated by the law we already have, which is technology-neutral and in most cases says nothing about computers at all:

• Law: the Data Protection Act [2021 Revision] applies to any AI that touches personal data, from the first principle ("Personal data shall be processed fairly", Schedule 1, Part 1) to the breach clock (s.16) and the right not to be subject to a solely automated significant decision (s.12).
• Law: the Electronic Transactions Act [2003 Revision] already says that "an electronic record shall be attributable to the originator if it was sent ... by the originator's electronic agent" (s.14(2)(b)), and an "electronic agent" is a program used to act "without review by an individual" (s.2). A message your AI sends is your message.
• Law: the Gender Equality Act [Act 21 of 2011] catches "any act, practice or policy that directly or indirectly results in discrimination ... regardless of whether the person responsible for the act, practice or policy intended to discriminate" (s.3(3)). A biased model is a practice.
• Law: for CIMA-regulated entities, the cybersecurity, corporate governance, internal controls and anti-money-laundering rules all apply to AI the moment it touches data, systems or decisions. Regulator guidance: CIMA's own words are that "FSPs are responsible for the quality of all outputs from any automated system, including those from third-party vendors" (CIMA Guidance Notes on the Prevention and Detection of ML, TF and PF in the Cayman Islands (February 2024), Section 16, Monitoring Systems, para 22).
• Law: for government, "All decisions and acts of public officials must be lawful, rational, proportionate and procedurally fair" (Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.19(1); cited below as "the Constitution"). A model does not get a public official out of that.

Where government is today. The facts, dated:

• On 6 March 2026, answering Parliamentary Question No. 83, the Premier set out "an earnest and determined aim to have a draft Artificial Intelligence legislative framework for consideration by Q2 2027" (Government statement: Parliamentary Question No. 83 (answered 6 March 2026), Artificial Intelligence Act, answer, pp.1-3), while declining to guarantee a commencement date. Whether Cayman gets a standalone AI Act or targeted amendments to the Data Protection Act and sector laws was left open. Risk classification, accountability, auditability and regulator coordination were named as decisions still to make.
• The same answer described a Civil Service AI Policy, Version 1.0, in final sign-off, under which "AI systems cannot be procured, piloted, or deployed without coordinated review for cybersecurity, legal compliance, data protection, risk assessment, and architectural alignment" (Government statement: Parliamentary Question No. 83 - civil service AI policy, answer p.2). That is the standard government set for itself. The policy text has not been published, and its adoption after March 2026 is not verified.
• The same answer said the Ministry of Health, Environment and Sustainability "is already developing legislative proposals to ensure safety, equity, informed consent, and accountable care in the use of AI" (Government statement: Parliamentary Question No. 83 (answered 6 March 2026), Artificial Intelligence Act, answer). Health may move before the rest.
• A National Digital Transformation Strategy Taskforce, appointed by the Premier and advisory, was launched in October 2025 with three pillars, one of them AI: to "Harness artificial intelligence (AI) and emerging technologies to modernise Government, empower business innovation and build a future-ready workforce" (Government statement: Government press release: Task Force to Energise Cayman's Digital Transformation (9 October 2025), press release). It was given six months. It is the only body of its kind. There is no Cayman AI governance council.
• On 2 October 2026, answering Parliamentary Question No. 36, government said "the Task Force indicated that they aim to complete the report in advance of the anticipated meeting of this Honourable House in November 2026" (Government statement: Parliamentary Question No. 36 (answered 2 October 2026), AI risks, Part 2 of answer). The report date had moved from end of May 2026 to before the November sitting. Whether it has since been delivered is not verified.
• No AI-specific statement from CIMA, the Ombudsman or OfReg was located as of 5 November 2026. Absence is not proven; it was not found.

Waiting is a choice with a cost. A draft framework in Q2 2027 is a draft. Law after that is later still. In the meantime every firm in George Town is already deciding, tool by tool, what goes in and who answers. This standard is for the meantime, and it is written so that when the framework lands it can adopt the parts that work rather than start again.

3. Principles

Seven. Each one gives the rule, why, the law that already backs it, and what good looks like at a real desk.

3.1 Substance here

Rule. Every AI use at Tier 2 or 3 has a named person in the Islands who understands it, can explain it, can stop it, and answers for it.

Why. Responsibility does not travel through an API. A vendor in another jurisdiction will not be in the room with the Ombudsman, CIMA or the client.

Law By analogy: the economic substance test asks whether an entity "is directed and managed in an appropriate manner in the Islands" (International Tax Co-operation (Economic Substance) Act [2026 Revision], s.4(2)(b)), and it is met where "the relevant entity's board of directors, as a whole, has the appropriate knowledge and expertise to discharge its duties as a board of directors" (s.4(3)(a)). Where another person does the work, the entity must be "able to monitor and control the carrying out of core income generating activities by that other person" (s.4(4)). That Act is not about AI; we borrow its test because it is a Cayman test of real control. Directly: for CIMA-regulated entities, "Regulated entities' governing bodies are ultimately responsible for cybersecurity" (CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 6.1.1), and those that outsource IT "remain ultimately responsible for outsourced IT functions and their cybersecurity" (Rule 6.5.1(a)). Regulator guidance: "regulated entities, including their Governing Body and Senior Management, remain ultimately responsible for all outsourced material functions or activities" (CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 1.3).

What good looks like. The AI register names a person, not a department. She can say in two minutes what the tool does, what it must never see, and how to switch it off. She has done it once, on purpose, to prove the switch works.

3.2 You answer for what the machine sends

Rule. A record, message, recommendation or filing produced by your AI is yours, whether or not a person read it first.

Why. Cayman law has said this since 2003. The attribution rule was written for automated systems and it fits AI exactly.

Law "an electronic record shall be attributable to the originator if it was sent -- (a) by a person who had been authorised by the originator to send the electronic record on his behalf; or (b) by the originator's electronic agent" (Electronic Transactions Act [2003 Revision], s.14(2)(b)); an electronic agent acts "without review by an individual" (s.2). An individual dealing with your electronic agent may avoid a record where the agent gave no chance to prevent or correct an error (s.15(b)). A securities licensee must "act with due skill, care and diligence in providing any service which that licensee provides or holds themselves out as willing to provide" (Securities Investment Business (Conduct of Business) Regulations [2026 Revision], reg 11(1)(b)); an AI-generated recommendation is still the licensee's recommendation. An organisation that publishes a model's false statement about a real person is the publisher, and the innocent-publication defence needs proof that the publisher exercised all reasonable care (Defamation Act [1995 Revision], s.11(5)).

What good looks like. Nothing leaves the building under your name without a person who would sign it. If a chatbot talks to clients, it says so, it gives them a way to correct an error, and a person reads the transcripts.

3.3 A person can see it and challenge it

Rule. Anyone significantly affected by an AI-shaped decision can learn that a machine was involved, get the reasons in plain English, and have a person look again.

Why. Explanation is the difference between a decision and an accident.

Law A data subject may by written notice require that "no decision taken by or on behalf of the data controller that significantly affects the data subject is based solely on the processing by automatic means of the data subject's personal data" (Data Protection Act [2021 Revision], s.12); where no notice was given and such a decision is taken, the controller must tell the person, who then has 21 days to ask for reconsideration, and the controller must reply with its steps within 21 days (s.12(2)-(3)). Where automated processing is the sole basis for a significant decision, "the data subject is entitled to be informed by the relevant data controller of the reasons for that decision" (s.8(3)). A dismissed employee who asks within fourteen days must be given a written statement of the reason within fourteen days (Labour Act [2021 Revision], s.12(1)). For public bodies, a person adversely affected has the right to request and be given written reasons (the Constitution, s.19(2)).

What good looks like. The reason a loan, a job or a renewal was refused can be written in three sentences a customer understands. "The model scored you low" is not a reason. The letter tells the person how to ask for a second look and who will do it.

3.4 What's yours stays yours

"What's yours stays yours."

— Perennos

Rule. Data goes into an AI system only for the purpose it was collected, under a lawful basis, with written terms that stop the vendor using it for anything else, and with a transfer ground before it crosses the border.

Why. The Islands' promise to clients is confidentiality. A free tool's terms of service are not a confidentiality undertaking.

Law Personal data "may be processed only if ... at least one of the conditions set out in paragraphs 1 to 6 of Schedule 2 is met" (Data Protection Act [2021 Revision], Schedule 1, Part 1, first principle); health data is sensitive personal data (s.3) and needs a Schedule 3 condition too. A controller using a processor does not comply unless there is a written contract under which the processor acts only on instructions and meets equivalent security obligations (Schedule 1, Part 2 para 3). "Personal data shall not be transferred to a country or territory unless that country or territory ensures an adequate level of protection for the rights and freedoms of data subjects" (Schedule 1, Part 1, eighth principle), subject to the Schedule 4 grounds. An attorney "shall keep the affairs of clients, former clients and potential clients ... confidential" (Cayman Islands Legal Services Code of Professional Conduct, 2026, Rule 5.1). A virtual asset service provider must "take such steps as may be necessary to protect and secure the personal data and virtual assets of its clients" (Virtual Asset (Service Providers) Act [2024 Revision], s.9(3)(b)). Regulator guidance: "A Service Provider should not use a regulated entity's proprietary information or its customers' information unless it is a necessary part of providing the contracted service and then only for the provision of such service" (CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 10.3); a VASP "should therefore be able to demonstrate that it has used Client's information only for the purpose for which it was obtained" (CIMA Rule and Statement of Guidance: Market Conduct for Virtual Asset Service Providers (February 2026) [SL 12 of 2026], para 6.7).

What good looks like. Before a tool is switched on, someone has written down: what data, whose, which Schedule 2 condition, which Schedule 3 condition if sensitive, where the servers are, which Schedule 4 ground if overseas, and the clause that says the vendor will not train on it. Our recommendation: where the data is client-confidential, patient or beneficial-owner information, run the model on systems you control, in the Islands, unless you can show why not.

3.5 Unfair outcomes are your outcomes

Rule. You test for unfair results by protected ground before go-live and at set intervals, and you fix what you find. Intent is not a defence.

Why. Models learn from the past. The past was not fair.

Law "Any act, practice or policy that directly or indirectly results in discrimination against a person on any ground specified in subsection (2), is an act of discrimination regardless of whether the person responsible for the act, practice or policy intended to discriminate" (Gender Equality Act [Act 21 of 2011], s.3(3)); it covers employment (s.4) and goods and services (s.13). "No person (whether an employer or an employee) shall discriminate with respect to any person's hire, promotion, dismissal, tenure, wages, hours or other conditions of employment" by reason of race, colour, creed, sex, pregnancy, age, disability, political belief or the exercise of rights (Labour Act [2021 Revision], s.80(1)). For government, "government shall not treat any person in a discriminatory manner in respect of the rights under this Part of the Constitution" (the Constitution, s.16(1)-(3)). Insurers and intermediaries in scope "must establish and implement policies and procedures on the fair treatment of customers that are an integral part of their business practices" (CIMA Rule: Market Conduct - Insurers, Agents and Brokers (May 2018) [SL 48 of 2018], Rule 5.3.4), and its customer-information policies "should prevent the use of customer information collected and held in a manner that results in unfair treatment" (Rule 5.4.1).

What good looks like. A CV screener was run against last year's applicants with the names and sexes swapped. The results were kept. The difference was small, written up, and the threshold was changed. The test is on the calendar for next quarter.

3.6 Every job leaves a record

"Every job gets written down. Every job leaves a record."

— Perennos

Rule. Every Tier 1 to 3 run leaves a record of inputs, system and version, output, reviewer and outcome, kept where the Islands' regulators can reach it.

Why. The clocks in section 6 are measured in days and hours. Records are the only way to meet them. They are also what the regulator will ask for, and what a court will admit.

Law CIMA may require a regulated person or a person believed to hold relevant information to provide information or produce documents, and destroying or hiding a document to avoid that requirement is an offence (Monetary Authority Act [2020 Revision], s.34(8), s.34(17)). A securities licensee "shall retain records for a minimum period of 5 years from the relevant date" (Securities Investment Business (Conduct of Business) Regulations [2026 Revision], reg 8(1) and (2)); anti-money-laundering records are kept at least five years and must be available without delay (Anti-Money Laundering Regulations [2025 Revision], reg 31). "In any civil proceedings a statement contained in a document produced by a computer is, subject to Rules of Court, admissible as evidence" if the computer was in regular use and operating properly (Evidence Act [2021 Revision], s.48); system logs meet that test more easily than unreviewed generated text does. For a public authority, prompts, outputs and logs are records, and "every person shall have a right to obtain access to a record other than an exempt record" (Freedom of Information Act [2021 Revision], s.6(1)). Regulator guidance: "Electronic information systems and the use of information technology have risks that must be effectively controlled by regulated entities to avoid disruptions to business and potential losses" (CIMA Rule and Statement of Guidance: Internal Controls for Regulated Entities (April 2023) [SL 40 of 2023], para 11.4.2), with an audit trail expected.

What good looks like. At 8.15 on a Monday morning the compliance officer can pull, for any decision last quarter, what went in, what came out, who looked, and what they did. It takes ten minutes, not a vendor ticket.

3.7 Assess before you switch it on

Rule. No AI tool goes live at Tier 1 or above without a written risk assessment, a tier decision and a named owner. Government holds itself to a review gate; so should you.

Why. The cheapest moment to find a problem is before the first real record goes in.

Law A person carrying out relevant financial business must "undertake assessment of risk prior to the launch or use of the new products and business practices, new delivery mechanisms and new or developing technologies" (Anti-Money Laundering Regulations [2025 Revision], reg 9). In-scope CIMA entities "must establish, implement, and maintain a documented cybersecurity framework" (CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 5.1.1) and "must demonstrate that data protection is part of their strategy and cybersecurity framework" (Rule 7.1); a new AI tool sits inside that framework, not beside it. "Each individual Director of the Governing Body must understand the limitations of services and reports provided by service providers, as applicable" (CIMA Rule: Corporate Governance for Regulated Entities (April 2023) [SL 39 of 2023], Rule 5.7.7). Government's own stated gate: "AI systems cannot be procured, piloted, or deployed without coordinated review for cybersecurity, legal compliance, data protection, risk assessment, and architectural alignment" (Parliamentary Question No. 83 - civil service AI policy, answer p.2).

What good looks like. The assessment is two pages, dated, signed, and attached to the register entry. It names the tier and why. It says what would make the firm turn the tool off.

4. The law as it stands

“It answers to you.”

— Perennos

This standard applies existing Cayman duties to AI use, checked to 5 November 2026; it separates those duties from recommendations and will change as the law changes.

AI does not create an empty space between you and your obligations. If you process personal data, advise a client, assess a patient or exercise public power, the existing duties still matter. The practical question is what your system does, whose interests it affects and whether you can answer for the result.

The tables use four registers: Law, Regulator rule, Regulator guidance and Government statement. Our recommendation identifies the additional safeguards proposed by this standard. A statement about a planned law is not an enacted law. A regulator’s guidance is not interchangeable with its rules.

The AI applications below are practical interpretations of the cited provisions. They are not quotations from statutes that expressly regulate AI. A dagger † marks an Act amended since its current revision; every cited section of those Acts was checked against the amending Acts on 5 November 2026 (section 12 lists the results). Where the enforcement route was not established, the table says so.

Rights and data

Electronic records and computer misuse

Employment and equality

Financial services, including CIMA rules and guidance

The cybersecurity rule excludes regulated mutual funds and private funds. Fully managed Class B, C and D insurers have special scope provisions that require careful reading. The outsourcing guidance separately excludes regulated mutual funds, private funds and private trust companies. Do not copy one entity’s obligations onto another merely because both use the same software. These exclusions do not remove other applicable obligations. (CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 3.1 and footnote 1; Rule 3.4; CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 2.1 and footnote 1.)

AML/CFT and sanctions

Legal profession

Health and insurance

The insurance market-conduct rule has a defined scope. It covers specified domestic and third-party retail business, not reinsurance generally. CIMA’s role and the Health Insurance Commission’s role are distinct. (CIMA Rule: Market Conduct - Insurers, Agents and Brokers (May 2018) [SL 48 of 2018], Rules 3.1–3.3; Health Insurance Act [2021 Revision], s.22 (disputes); s.23 (appeals).)

Government and public bodies

Five provisions deserve particular attention.

The clocks start before the investigation ends. The data-protection duty runs “without undue delay, but no longer than five days after the data controller should, with the exercise of reasonable diligence, have been aware of that breach”. CIMA’s in-scope cyber rule requires immediate written notice of a material or potentially material incident, with an outer limit of 72 hours following discovery. These are separate triggers, recipients and duties. Your clock file must support both. It is not permission to wait until either outer limit. (Data Protection Act [2021 Revision], s.16 (personal data breaches); CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 8.1.)

An unexplained score can become your problem. Cayman’s automated-decision provisions have their own structure. They include a written-notice right, duties following certain solely automated significant decisions, reconsideration procedures and exceptions. Section 8(3) also provides a reasons right in its specified circumstances. Do not replace this structure with a borrowed summary of foreign law. Our recommendation: provide meaningful human reconsideration before harm becomes difficult to reverse, even where your use falls outside those specific provisions. (Data Protection Act [2021 Revision], s.12 (rights in relation to automated decision-making); s.8(3) (reasons for solely automated decisions).)

Outsourcing does not move the responsibility out of your office. A processor contract must address instructions and security. Applicable CIMA controls retain responsibility with the regulated entity. Our recommendation: identify separately what the provider does with prompts, retrieved documents, outputs, diagnostic records and backups. A provider may process some information for you and use other information for its own purposes. One label in a contract does not answer every data question. (Data Protection Act [2021 Revision], Schedule 1, Part 1, seventh principle; Part 2 para 3 (processor contract); CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 6.5.1(a) and (c); 6.5.2.)

Your electronic agent can speak for you. The Act attributes an electronic record to its originator where the originator’s electronic agent sent it. That is narrower than a universal rule assigning every possible AI loss, but it is already enough to make careless delegation dangerous. Our recommendation: an agent that can send, order, submit or commit must have named authority and technical limits before it receives access. (Electronic Transactions Act [2003 Revision], s.14(2)(b), s.15(b) and s.2 ('electronic agent').)

A public official must still give a lawful answer. The constitutional requirement is direct: “All decisions and acts of public officials must be lawful, rational, proportionate and procedurally fair.” An adversely affected person can request written reasons. Our recommendation: retain the evidence and human reasoning needed to explain the actual decision. “The system recommended it” is not a substitute for the official’s assessment. (Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.19(1)-(2) (Lawful administrative action).)

5. The tiers in detail

"Nothing gets pasted into a chatbot. Nothing gets emailed to a stranger."

— Perennos

5.1 How to place a use

Ask the two questions and be honest.

Can the harm be undone? A draft nobody has sent can be deleted. A refused mortgage, a frozen account, a missed diagnosis, a rejected applicant who took another job: these cannot be undone by deleting a file.

Can the affected person see it and challenge it? A customer who gets a letter can argue with it. A customer whose file was quietly scored and routed never knows.

When two answers are possible, take the higher tier. A tool can sit at different tiers for different uses; register each use, not each product. Our recommendation: the tier is decided by the named owner and recorded with the reason; it is reviewed when the use changes, and at least once a year.

5.2 Tier 0: Assist

Covers. Internal drafting, search, summarising and coding on material that is public or non-personal and not client-confidential. The output is reviewed by the person who asked for it and nobody relies on it until they have.

Required controls.

• A list of approved tools, and a rule that anything else is Tier 1 until assessed.
• A named owner for the list.
• One line per tool in the AI register: system, purpose, owner.
• A plain instruction to staff: no personal data, no client, patient or beneficial-owner material, no credentials. The moment any of those goes in, it is Tier 1 or higher.

Law that bites if you get the boundary wrong. Client material in a tool is a confidentiality question for attorneys (Cayman Islands Legal Services Code of Professional Conduct, 2026, Rule 5.1). Personal data in a tool is processing under the Data Protection Act [2021 Revision] from the first principle onward.

5.3 Tier 1: Inform

Covers. Outputs a professional reviews before anyone relies on them: research memos, first drafts of client letters, meeting summaries, document extraction, translation, where personal or confidential data is or may be present.

Required controls.

• Tier 0's tool approval, ownership and register controls; personal or confidential data needs the safeguards below.
• A named owner for the use.
• The lawful basis written down: the Schedule 2 condition, and a Schedule 3 condition where the data is sensitive (Law: Data Protection Act [2021 Revision], Schedule 1, Part 1; s.3).
• A written contract with the vendor under which it acts only on your instructions and meets equivalent security obligations (Law: Data Protection Act [2021 Revision], Schedule 1, Part 2 para 3). Our recommendation: the contract says in plain words that your data is not used to train or improve the vendor's models, names sub-processors, and sets breach notice to you in hours.
• Where the vendor or its servers are outside the Islands, the transfer ground recorded before the first record goes in (Law: Data Protection Act [2021 Revision], Schedule 1, Part 1, eighth principle; Schedule 4).
• For CIMA-regulated entities, an assessment of the provider and, where the arrangement is a material outsourcing, written notice to CIMA within a reasonable timeframe of signing (Regulator guidance: CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 14.1 and 14.3; if unsure whether it is material, communicate with CIMA).
• A reviewer who signs off before reliance, named in the record.
• The clock file (section 6) completed at set-up, not after an incident.
• Records kept for as long as the underlying business record would be.

5.4 Tier 2: Decide-with

Covers. AI materially shapes a decision about a person, money, health, cover, employment or access to a service, and a person makes the final call. Credit and underwriting scores, triage suggestions, applicant rankings, transaction monitoring alerts, pricing.

Required controls.

• Everything in Tier 1.
• Human final say with real authority: the decision-maker can disagree, and overrides are recorded and counted. If nobody ever overrides, it is not human final say; it is Tier 3 wearing a badge.
• Explanation on request, in plain English, within the time the law gives (Law: Data Protection Act [2021 Revision], s.8(3); Labour Act [2021 Revision], s.12(1) for terminations; the Constitution, s.19(2) for public bodies).
• Testing for unfair outcomes by protected ground before go-live and at set intervals, results kept (Law: Gender Equality Act [Act 21 of 2011], s.3(3); Labour Act [2021 Revision], s.80(1)).
• Logged evidence per decision: inputs, system and version, output, decision-maker, outcome, override if any.
• For relevant financial business, the written risk assessment before use (Law: Anti-Money Laundering Regulations [2025 Revision], reg 9).
• For CIMA-regulated entities, the use is inside the documented cyber framework (Regulator rule: CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 5.1.1) and visible to the governing body, which "must provide oversight in respect of the design and implementation of sound risk management and internal control systems and functions" (Regulator rule: CIMA Rule: Corporate Governance for Regulated Entities (April 2023) [SL 39 of 2023], Rule 5.9.1).
• The Answerable Test rehearsed once a year: pick a decision at random and produce the seven answers against the clock.

5.5 Tier 3: Hold

Covers. A decision that is hard to reverse or invisible to the person: automated refusal of a service or application, clinical or emergency triage, fraud or anti-money-laundering flags that freeze money, hiring rejections before a person reads the application, benefit terminations, anything that acts without a person in the loop.

The name is the instruction. Hold, and check these before you go on.

Required controls.

• Tier 2's controls, with a person approving each decision unless lawful automation is justified under the safeguards below.
• Documented necessity: why a person cannot do this step, written by the owner and approved by the governing body or its equivalent.
• A named accountable person in the Islands with the authority to stop it, and a kill-switch that has been tested, with the date of the test recorded.
• Prior testing on real historical cases, results kept, including the error rate and who it falls on.
• A working appeal route the person is told about at the time of the decision, with a named reviewer and a clock. Our recommendation: acknowledge within two working days; decide within ten.
• No solely automated decision that significantly affects a person unless the conditions in s.12 of the Data Protection Act [2021 Revision] are met, and where s.12(2) applies, notice to the person after the decision, 21 days for the person to ask for reconsideration, and a written reply setting out the controller's steps within 21 days of that request.
• In anti-money-laundering work, a person, not the model, decides on the flag. The duty to disclose turns on what the person "knows or suspects or has reasonable grounds for knowing or suspecting" (Law: Proceeds of Crime Act [2025 Revision], s.136(1)); a flag is information that has come to the business, and the compliance officer keeps overall responsibility however much is delegated (Regulator rule: CIMA Rule: Effective Compliance Programme for the Prevention and Detection of ML, TF and PF for Financial Services Providers (July 2026), Rule 8.6).
• In health care, "Clinical services at every health care facility shall be managed by a person who is a registered practitioner" (Law: Health Practice Act [2026 Revision], s.8); a model advises, a practitioner decides.
• In health insurance, an approved insurer "shall not terminate, fail or refuse to renew a standard health insurance contract except where" the Act's listed grounds apply (Law: Health Insurance Act [2021 Revision], s.15(1)); an automated system cannot add grounds.
• For CIMA-regulated entities, assess materiality; outsourcing all or substantially all of a management oversight function, including compliance or risk management, is always material (Regulator guidance: CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 4.2).

5.6 Red lines

These are not tiers. They are things Cayman practice does not do.

1. Undisclosed AI impersonation of a real person. Your electronic agent's words are attributed to you (Law: Electronic Transactions Act [2003 Revision], s.14(2)(b)). Our recommendation: a person is always told they are dealing with a machine, and a machine never presents as a named human who did not write the words.
2. Training a third party's model on clients' or patients' confidential data without a lawful basis and informed agreement. Further processing must not be incompatible with the purpose the data was obtained for (Law: Data Protection Act [2021 Revision], Schedule 1, Part 1, second principle); attorneys keep client affairs confidential (Regulator rule: Cayman Islands Legal Services Code of Professional Conduct, 2026, Rule 5.1); Regulator guidance limits provider use to the contracted service (CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 10.3).
3. A model alone deciding to tip off, refuse or freeze in an anti-money-laundering context. Tipping off is an offence where a disclosure "is likely to prejudice any investigation which might be conducted following the disclosure" (Law: Proceeds of Crime Act [2025 Revision], s.139(1)); a chatbot that hints at a report creates a tipping-off risk; liability depends on the statutory elements and exceptions. The firm is "responsible for the quality of all outputs from any automated system" (Regulator guidance: CIMA Guidance Notes on the Prevention and Detection of ML, TF and PF in the Cayman Islands (February 2024), Section 16, Monitoring Systems, para 22).
4. Fully automated decisions with legal or similarly significant effect where the Data Protection Act gives the person rights against them, unless the Act's conditions are met (Law: Data Protection Act [2021 Revision], s.12).
5. AI-generated evidence or filings presented to a court or regulator without disclosure and human verification. An attorney shall "not knowingly or recklessly mislead or attempt to mislead the Court or permit the Court to be misled" (Regulator rule: Cayman Islands Legal Services Code of Professional Conduct, 2026, Rule 2.1), and breach of the Code may amount to professional misconduct (Law: Legal Services Act [Act 57 of 2020], s.37(5)). Hiding a document from CIMA is an offence (Law: Monetary Authority Act [2020 Revision], s.34(17)).

Section 10 maps these tiers and red lines to the EU AI Act, NIST and ISO/IEC 42001 for firms with overseas exposure.

6. The Answerable Test and the clock file

6.1 The test

Could you answer for this AI use inside the clocks Cayman law already sets?

Law Five days for a personal data breach, counted from when you "should, with the exercise of reasonable diligence, have been aware of that breach" (Data Protection Act [2021 Revision], s.16), with notice to the Ombudsman and the people affected, and a fine of $100,000 for failure (s.16(2)). Immediate written notice to CIMA for a material or potentially material incident, "no later than 72 hours following the discovery of said incident" (CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 8.1), for entities in scope; regulated mutual funds and private funds are excluded (Rule 3.1, footnote 1), though their administrators and managers may not be. Ten days by email to the regulator for "any substantive issues which could materially affect the regulated entity" (CIMA Rule: Corporate Governance for Regulated Entities (April 2023) [SL 39 of 2023], Rule 5.6.1(a)). A breach of a CIMA rule is dealt with by CIMA under its supervisory powers and administrative fines; it "shall not constitute an offence" of itself (Monetary Authority Act [2020 Revision], s.34(6)). The Data Protection Act's five days is a statutory duty with a criminal penalty. Do not blend the two.

Note the different starts. CIMA's clock runs from discovery. The Ombudsman's runs from when reasonable diligence would have found it. If your logs would have shown the leak on Monday and you looked on Thursday, the five days began on Monday. Regulator guidance: missing investigation details do not justify waiting for a final report, and a processor's incident still requires the controller to act (Personal data breaches, Ombudsman guidance, undated, checked 4 November 2026).

The test is not a form you fill in after the event. It is a file you keep so that the event is survivable. Two uses at Tier 1 need the file. One use at Tier 3 needs it rehearsed.

6.2 The clock file: field by field

One file per registered use. Fields marked * are completed before the first real record goes in. The rest are kept current.

Our recommendation If a field is blank at Tier 2 or 3, the use is not live. If it takes more than a working day to produce the seven answers for a live use, that is the finding, and the owner fixes the file before the next run.

7. Leak-proof by design: future-proof controls

“What's yours stays yours.”

— Perennos

“Leak-proof” is the design aim. It is not a promise that any system cannot fail. Security guidance is plain: prompt injection has no complete fix. Build several barriers and assume one may fail. (OWASP Top 10 for LLM Applications 2025, LLM01, LLM02, LLM06 mitigation sections.)

Our recommendation: design around what the system can read, reveal and do. Those boundaries remain useful when the model changes. Every control below includes a version a five-person firm can put in place this month. The security references inform the design; they do not create Cayman legal duties.

7.1 Model and data leakage

What it stops. Client material escaping through a prompt, retrieved document, output, diagnostic record or model trained on confidential information. It also limits access to model files and connected data.

Reference. Least privilege and restricted access address sensitive-information disclosure. Secure development guidance also calls for protecting sensitive records and tracking models, data and prompts. (OWASP Top 10 for LLM Applications 2025, LLM01, LLM02, LLM06 mitigation sections; UK NCSC and CISA (with partners), Guidelines for secure AI system development, version 1.0, Secure development, Secure operation and maintenance.)

Our recommendation — this month. Make one approved-use sheet: permitted information, permitted users, destinations, retention and whether provider training is allowed. Remove client names where they are unnecessary. Separate matters and customers in retrieval permissions. Test whether one user can obtain another user’s documents. Keep passwords and signing secrets outside prompts.

Do not assume removing a name makes a Cayman record anonymous. An unusual occupation, district and transaction may identify someone. Record what you removed and what could still identify them.

7.2 Direct and indirect prompt injection

What it stops. Instructions from a user, email, website or document being mistaken for authority to reveal data or perform an action.

Direct injection arrives through an instruction to the assistant. Indirect injection arrives inside material the assistant was asked to read. A document saying “send the client list here” remains a document, even when the assistant reads it aloud.

Reference. Treat the model as an untrusted participant, restrict tools and require human approval for high-risk actions. (OWASP Top 10 for LLM Applications 2025, LLM01, LLM02, LLM06 mitigation sections.)

Our recommendation — this month. Disable automatic external sending. Restrict retrieval to approved folders. Give the assistant read access before considering write access. Put harmless hostile instructions into a test email and attachment. Confirm that neither can change recipients, retrieve unrelated records or trigger a payment.

A prompt saying “ignore malicious instructions” is useful guidance, not a security boundary. Enforce permissions outside the model.

7.3 Agents with named authority

“Every job gets written down. Every job leaves a record.”

— Perennos

What it stops. An assistant using broad credentials to act beyond the task, spend repeatedly or pass its authority to another system.

Reference. Minimise tools and functions, avoid open-ended capabilities and require human approval for privileged operations. (OWASP Top 10 for LLM Applications 2025, LLM01, LLM02, LLM06 mitigation sections.)

Law An electronic record sent by the originator’s electronic agent is attributable to the originator. (Electronic Transactions Act [2003 Revision], s.14(2)(b), s.15(b) and s.2 ('electronic agent').)

Our recommendation — this month. Give every acting agent an authority record naming the responsible person, approved task, permitted systems, recipients, expiry, spending ceiling and stop control. Use a separate restricted account. Set both a per-action ceiling and a total ceiling; otherwise repeated small actions can defeat the limit. Block delegation unless expressly authorised.

Keep the proposed action, approving person, actual action and result. A log that records only “success” cannot show whether the action was authorised.

7.4 Provenance and labelling

What it stops. Synthetic material being mistaken for a person’s words, an original photograph or verified evidence.

Reference. Signed provenance information can record an asset’s origin and edits. The verified specification material supports integrity checking, not a guarantee that the depicted event is true. Credentials can also be removed. (C2PA Content Credentials: C2PA Technical Specification (version selector lists 2.5), normative text on redaction of hard-binding assertions (2.2 page).)

Our recommendation — this month. Label materially generated or altered public content in language the reader can see. Keep the original, edited version, purpose, reviewer and publication record. Use signed provenance where supported, but retain visible disclosure when content is copied or printed.

Never use undisclosed AI impersonation of a real person. Never present generated court or regulatory material as verified evidence without disclosure and human verification. A provenance record establishes a history; it does not establish truth.

7.5 Local and sovereign compute

What it stops. Some unnecessary transfers and dependencies on external operators. It can make access, retention and interruption easier to control.

Reference. Secure development guidance supports supply-chain assessment and protection of assets. The regional task-force report identifies sovereign data infrastructure as a policy direction. Neither establishes that local processing is automatically safe. (UK NCSC and CISA (with partners), Guidelines for secure AI system development, version 1.0, Secure development, Secure operation and maintenance; Caribbean Telecommunications Union, Caribbean AI Task Force Final Report: Toward Harmonized AI Policies and Recommendations for the Caribbean (2026), front matter.)

Our recommendation — this month. Trace one real workflow. Identify the inference location, administrator access, backups, update route and diagnostic destinations. Test that the private route stays private during errors and outages. Record any external fallback and keep it disabled until approved.

Local equipment can still leak, fail or be remotely administered. A system abroad may have stronger operational controls than a neglected machine in your office. Choose the arrangement you can evidence and govern. Location alone does not establish data-protection compliance.

7.6 Testing and red-teaming

What it stops. Familiar failures reaching people because a convincing demonstration was mistaken for an assessment.

Reference. Adversarial testing and treating the model as untrusted are explicit controls. Secure operation includes monitoring and records for investigation and remediation. (OWASP Top 10 for LLM Applications 2025, LLM01, LLM02, LLM06 mitigation sections; UK NCSC and CISA (with partners), Guidelines for secure AI system development, version 1.0, Secure development, Secure operation and maintenance.)

Our recommendation — this month. Build twenty short cases from your work using fictional or properly protected data. Include wrong identities, missing facts, contradictory documents, hostile instructions, unfair outcomes and interrupted tasks. Have someone other than the person who configured the system run them.

Record expected and actual behaviour. Retest after changes to the model, retrieval sources, tools or permissions. For Tier 2 and Tier 3, test the human review too: can the reviewer detect the error, refuse the recommendation and stop the action?

7.7 Sandboxes

What it stops. A trial becoming an uncontrolled live service.

Reference. Restricted tools and human approval support technical containment. The supplied official report describes supervision of sandbox licensees in the virtual-asset context; it does not establish a general Cayman AI sandbox. (OWASP Top 10 for LLM Applications 2025, LLM01, LLM02, LLM06 mitigation sections; CIMA Annual Report 2025 (tabled at parliament.ky, June 2026) - technology outlook and VASP and Fintech Innovation Unit, 2026 Outlook list, printed near p.76; Structure section.)

Our recommendation — this month. Use a separate environment, fictional records, no live payment credentials and no ability to send externally. State the question being tested, duration, owner, success measures and stop conditions. Record what must change before live use.

A sandbox is a controlled test, not an exemption. Calling something a pilot does not suspend confidentiality, professional duties or applicable regulation.

7.8 Incident sharing without a second disclosure

What it stops. The same failure recurring across firms, while preventing incident reports from exposing clients again.

Reference. Secure operation guidance supports monitoring, investigation and remediation and treats diagnostic records as sensitive. (UK NCSC and CISA (with partners), Guidelines for secure AI system development, version 1.0, Secure development, Secure operation and maintenance.)

Our recommendation — this month. Prepare two incident forms. The restricted form holds identities and evidence for authorised responders. The shared form describes the failure pattern, affected capability, control that failed and repair, without personal or confidential information.

Do not attach raw prompts to a group email. Have one person check the shared version for identification risk. Mandatory notifications follow their own legal routes and clocks; voluntary learning must neither delay nor replace them.

8. Industry guidelines

Our recommendation Classify the use, not the industry. A bank can have Tier 0 drafting. A small hotel can have Tier 3 automated exclusion. Higher tiers add controls:

• Tier 0 — Assist: public or non-personal material, reversible internal use and a named owner.
• Tier 1 — Inform: approved data use, source checks, competent review and records before reliance.
• Tier 2 — Decide-with: meaningful human authority, testing for unfair outcomes, reasons and recorded decisions.
• Tier 3 — Hold: no live use until necessity, named accountability, prior testing, a working challenge route and a stop control are documented.

A human signature does not lower the tier if the person cannot understand or change the result. Red lines remain red lines at every tier. The examples below are illustrative, not reports of Cayman incidents.

9. When something goes wrong: incidents and complaints

Only routes and clocks that the law or a regulator has set are given here. Where the law is silent, the route is marked as our recommendation.

9.1 If the AI leaks, loses or wrongly exposes personal data

Law Notify the Ombudsman and the data subjects "without undue delay, but no longer than five days after the data controller should, with the exercise of reasonable diligence, have been aware of that breach", describing the nature, consequences, measures taken and recommended mitigation (Data Protection Act [2021 Revision], s.16). Failure is an offence with a fine of $100,000 (s.16(2)). Directors and officers are liable where an offence is attributable to their consent, connivance or neglect (s.58). Regulator guidance: a vendor's incident is your incident; do not wait for the vendor's final report (Personal data breaches, Ombudsman guidance, undated, checked 4 November 2026).

Law If you are a CIMA-regulated entity in scope of the cyber rule, immediate written notice to CIMA for a material or potentially material incident, and no later than 72 hours after discovery (CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 8.1); affected persons are told where non-public information is breached (Rule 8.4). Regulated mutual funds and private funds are outside that rule (Rule 3.1, footnote 1).

Law For any substantive issue that could materially affect a regulated entity, the governing body ensures the regulator is notified by email within ten days (CIMA Rule: Corporate Governance for Regulated Entities (April 2023) [SL 39 of 2023], Rule 5.6.1(a)).

Our recommendation Start all three clocks at the earliest moment, which is when the logs would have shown it. Write the first notice from the clock file; that is what it is for.

9.2 If an AI flag or output touches a suspicion of crime

Law Where a person in the regulated sector "knows or suspects or has reasonable grounds for knowing or suspecting that another person is engaged in criminal conduct", the required disclosure goes to the nominated officer or the Financial Reporting Authority as soon as practicable (Proceeds of Crime Act [2025 Revision], s.136(1)). The model's flag is information that came to the business; a person decides what it means. Nobody, and no system, makes a disclosure "likely to prejudice any investigation which might be conducted following the disclosure" (s.139(1)): check every customer-facing AI message for tipping off before it is sent.

Law If an AI agent or scraper has secured access to a computer, or modified its contents, without authority, offences may arise under the Computer Misuse Act [2015 Revision] where the required intent, knowledge and other elements are present (s.3(1); s.5(1)). Our recommendation: treat it as a potential offence from the first hour, preserve the logs, take advice, and report it.

9.3 If a person wants to challenge an AI-shaped decision

The route depends on who decided and about what. Tell the person the route; the law in several places requires it.

Law Any organisation, personal data. A person may by written notice require that no significant decision about them be based solely on automated processing (Data Protection Act [2021 Revision], s.12). After a solely automated significant decision taken without such a notice, the controller must tell the person, who has 21 days to ask for reconsideration; the controller replies with its steps within 21 days (s.12(2)-(3)). The person is entitled to the reasons (s.8(3)). A controller answering a subject request must tell the requester of the right to complain to the Ombudsman (Data Protection Regulations [SL 17 of 2019], reg 5). The Ombudsman may order a controller to take steps or to stop processing (s.45), failure to comply is an offence (s.46), and a serious contravention likely to cause substantial damage or distress can draw a monetary penalty of up to $250,000 (s.55).

Law Employment. A dismissed employee who asks within fourteen days must be given the written reason within fourteen days (Labour Act [2021 Revision], s.12(1)). Discrimination in hire, promotion, dismissal and conditions is an offence (s.80(1)); indirect discrimination on the Gender Equality Act's grounds is discrimination regardless of intent (Gender Equality Act [Act 21 of 2011], s.3(3)).

Law Health insurance. A dispute over a standard health insurance contract goes to the Health Insurance Commission in the first instance, with an appeal to the Grand Court within 90 days (Health Insurance Act [2021 Revision], s.22; s.23).

Law Virtual asset service providers. Complaints go through the provider's complaints-handling framework: written acknowledgement, timely, fair and consistent handling, a log of complaints and resolutions, and reporting to CIMA of complaints that represent a material risk to clients (CIMA Rule and Statement of Guidance: Market Conduct for Virtual Asset Service Providers (February 2026) [SL 12 of 2026], para 11.1, 11.3-11.6).

Law Government and public officials. A person adversely affected by a public official's decision has the right to request and be given written reasons (the Constitution, s.19(2)). It is unlawful for a public official "to make a decision or to act in a way that is incompatible with the Bill of Rights" (s.24), and an application may be made to the Grand Court (s.26). Records of public-sector AI use can be requested (Freedom of Information Act [2021 Revision], s.6(1)).

Law Health care. Negligence or incompetence by a registered practitioner, including in the use of a tool, is a matter for the practitioner's Council, which may censure, condition, suspend or strike off (Health Practice Act [2026 Revision], s.36(1)(d)).

Law Legal services. Failure to comply with the Code of Professional Conduct may amount to professional misconduct (Legal Services Act [Act 57 of 2020], s.37(5)).

Our recommendation The firm's own appeal route in the clock file (field H5) runs alongside the statutory routes and must never delay or restrict them: a named person, two working days to acknowledge, ten to decide, in writing, in plain English.

9.4 The order of work in the first 72 hours

Our recommendation, built from the clocks above.

• Hour 0 to 2. Pull the kill-switch if the harm is continuing. Open the clock file. Fix the time the logs would first have shown the problem; that is day one for the Ombudsman.
• Immediately for CIMA notice; within 24 hours for the record. Write the seven answers as they stand. Where the data is personal, draft the s.16 notice. Where you are CIMA-regulated and in scope, send the Rule 8.1 notice; "potential to become a material incident" is enough.
• Without undue delay. Tell the people affected in plain words: what happened, what it means for them, what you have done, what they can do.
• Without undue delay, inside the five-day limit. Ombudsman notice filed. Reversal plan running: wrong outputs withdrawn, decisions remade by a person.
• Day 10. Governing-body notice to the regulator where Rule 5.6.1(a) applies. Root cause written. File updated. Rehearsal date set.

10. Hub translation and what we borrowed

You should not need nine separate evidence files to explain one system. You do need to distinguish the duties that apply to you from frameworks you choose to use.

Our recommendation: use the clock file as a common evidence pack. Then add the requirements of each relevant jurisdiction or framework. The mapping below identifies useful connections. It does not establish equivalence, certification or a substitute for a foreign legal assessment.

The obvious alternative is a renamed global principles checklist. We reject it because it does not tell a Cayman firm who must act when the clock starts.

Mapping the standard

The EU dates matter

The verified source records Regulation (EU) 2026/1744 as entering into force on 27 July 2026. It moves application for Annex III high-risk systems to 2 December 2027, and Annex I high-risk systems to 2 August 2028. The general application date remains 2 August 2026, with earlier phased provisions. Do not repeat the superseded high-risk dates. (Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, Art. 113 application dates.)

The verified amendment also changes the AI-literacy wording to measures supporting the development of literacy. For specified systems already on the market before 2 August 2026, the machine-readable marking requirement has a 2 December 2026 compliance date. The supplied source also records new prohibitions concerning non-consensual intimate imagery and child sexual abuse material generation applying from 2 December 2026. These dates address different provisions. (Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, Art. 4 AI literacy; Regulation (EU) 2024/1689, Art. 50(1), (2), (4) transparency; Art. 5(1)(a)-(h) prohibited practices (headings, paraphrased).)

Cayman location alone does not settle EU scope. The verified provision reaches specified third-country providers and deployers where system outputs are used in the Union. Assess the actual activity and role. (Regulation (EU) 2024/1689 (AI Act), Art. 2(1)(a)-(c).)

What the smaller jurisdictions teach us

Jersey’s financial-services guidance keeps responsibility with firms and does not create new requirements. Guernsey’s policy statement works through existing principles and points firms towards established risk and security references. Bermuda’s verified paper is a consultation. These are distinct approaches, not evidence that every financial centre has adopted the same rulebook. (Jersey Financial Services Commission, Guidance on the use of AI in Jersey's financial services sector (July 2026), executive summary; JFSC industry update 14 July 2026; Guernsey Financial Services Commission, Policy Statement - Use of Artificial Intelligence (29 January 2026), statement text; Bermuda Monetary Authority, Consultation Paper: The Responsible Use of Artificial Intelligence in Financial Services in Bermuda (14 August 2026), paras 2 and 43.)

The Council of Europe Convention opened for signature on 5 September 2024, and the UK signed that day. As of 4 November 2026 the Treaty Office records it as not in force, with two ratifications; the UK has not ratified and no territorial application is recorded, so it does not currently extend to the Cayman Islands. It is therefore not presented here as a Cayman obligation. (Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, coe.int convention page; UK Government press release: UK signs first international treaty addressing risks of artificial intelligence, press release, 5 September 2024.)

What this standard brings together for Cayman

That heading describes the proposed Cayman combination. The evidence does not establish worldwide exclusivity for any individual mechanism.

1. Substance in the Islands. A named person here must understand, explain, stop and answer for consequential AI use. The economic-substance wording provides the analogy, not an AI obligation: “is directed and managed in an appropriate manner in the Islands in relation to that relevant activity”. (International Tax Co-operation (Economic Substance) Act [2026 Revision], s.4(2)(b).)

1. One file built for separate local clocks. The Answerable Test connects everyday evidence to the DPA’s five-day duty and CIMA’s immediate notification duty with its 72-hour outer limit. It preserves both triggers rather than replacing them with one deadline. (Data Protection Act [2021 Revision], s.16 (personal data breaches); CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 8.1.)

1. Our recommendation: reversibility and challenge determine scrutiny. Ask whether the harm can be undone and whether the affected person can see and challenge it. That catches a hidden hotel blacklist as readily as a financial decision.

1. Our recommendation: a small-firm version of every control. The five-person office gets a workable starting point. The large institution gets the same questions with deeper evidence. Neither gets to substitute a policy document for actual control.

11. A Cayman AI Governance Council: a proposal

“Spread love. Help people. Build things that turn people into enterprises.”

— Perennos

Cayman already has a government-appointed National Digital Transformation Strategy Task Force with an AI-related remit. Its October 2026 parliamentary update described a report intended ahead of the anticipated November sitting. Public records do not yet show it delivered or tabled. Nor do they show that a standing AI council already exists. If either exists, it had not been published when we checked. (Government press release: Task Force to Energise Cayman's Digital Transformation (9 Oct 2025), press release; launch event 6 Oct 2025; Parliamentary Question No. 36 (answered 2 October 2026) - Task Force status and Minister, Part 2 of answer.)

Our recommendation: establish a standing advisory Council only after checking the task force’s final remit and recommendations. Carry its work forward. Do not create a competing centre of authority.

Three organisational choices

All three options below are Our recommendation for consideration, not descriptions of existing institutions.

Mandate

Our recommendation The Council should maintain a public-interest standard, organise shared learning and advise government on gaps that existing powers cannot address.

Its work should include reusable clock-file templates, sector examples, public-sector impact assessments, evaluation methods and plain-language explanations of individual rights. It should identify where regulation overlaps or leaves uncertainty and send those questions to the competent bodies.

It should publish evidence about adoption and harm, not just activity counts. A useful measure is whether a small firm can identify a faulty decision, explain it and correct it. Another is whether a person can reach someone with authority to reconsider an outcome.

The Council should support Caymanian capability. Commission practical teaching material, include local technologists in technical review and make participation possible beyond firms with dedicated policy staff.

Membership by seat

Our recommendation Appoint members by published criteria and role. Name no automatic beneficiary.

Our recommendation Use staggered terms and rotate industry and technical seats. Publish the selection criteria, affiliations, relevant experience and reasons for appointments. Require declarations of financial interests, provider relationships and paid advisory work. Record recusals.

Regulatory participation should not imply endorsement of every Council publication. Members should be able to record dissent. Bring in additional expertise for a defined question without turning every specialist into a permanent member.

What it must not do

Our recommendation The Council must not become a licensing authority through informal practice.

It should not approve individual products, certify that a firm complies with Cayman law, direct a regulator’s enforcement decisions or hear appeals that belong to courts, regulators or statutory bodies. It should not offer a shortcut around professional obligations.

It should not collect raw client, patient or employee records to maintain a national incident database. Learning can usually begin with carefully reduced descriptions. Where identifiable evidence is necessary for an investigation, it belongs with the properly authorised body.

Membership must not confer preferred-supplier status. Council material should remain usable without buying services from any member.

Transparency and the first-year plan

Our recommendation Publish agendas, minutes, attendance, recommendations, dissent and a conflicts register. Where material is withheld, publish the reason and review the restriction. Keep commercially sensitive consultations separate from decisions about the standard.

Our recommendation Each quarter should end with usable material and named responsibility for unfinished work. Participation totals and meeting counts are secondary. Publish whether recommendations were adopted, deferred or rejected, and why.

Reporting to Parliament

Our recommendation The responsible Minister should arrange for the annual report to be laid before Parliament, subject to the proper parliamentary procedure. The report should distinguish Council advice, government decisions and regulator positions.

It should set out expenditure, participation, material dissent, practical results, unresolved gaps and proposed legislative questions. Publish it at the same time as tabling, with necessary confidential material separated. Invite scrutiny through the appropriate parliamentary process.

The Council should recommend a response timetable for government consideration of its proposals. That timetable would be a governance commitment, not an invented statutory deadline.

Your test for the Council should be the same test applied to the systems it discusses: who is responsible, what evidence supports the decision, who can challenge it and what changes when something goes wrong?

12. Review cadence, versioning and corrections

Cadence. This document is reviewed every three months. Version 1.0 is current to 5 November 2026. The next scheduled review is 5 February 2027. A review is also triggered, out of cycle, by any of these: publication of the Task Force report; publication or adoption of the Civil Service AI Policy; a draft AI legislative framework from government; any AI-specific statement from CIMA, the Ombudsman or OfReg; a new revision of any statute cited in section 4; a reported Cayman court decision on AI.

Versioning. Whole-number versions change the standard's rules (the tiers, the test, the red lines). Decimal versions change citations, dates, wording or examples without changing a rule. Every version carries the date it is current to and a change note saying what moved and why. Superseded versions stay available.

Known open checks. We publish what we have not been able to verify, so that you do not rely on it blind. On 5 November 2026 we read the amending Acts against every section this standard cites. The Monetary Authority (Amendment) Act 2026 (Act 22 of 2026) replaces "Minister charged with responsibility for Financial Services" with "Minister" throughout the principal Act, including s.34(1) and (5). The Health Practice (Amendment) Act 2026 (Act 14 of 2026) replaced s.35(1) with new subsections (1), (1A) and (1B): Cabinet may prescribe standards of professional practice for every registered practitioner, and those standards prevail over a Council's code where they conflict. The 2026 amendments to the Health Insurance, Proceeds of Crime, Mutual Funds, Private Funds and Virtual Asset (Service Providers) Acts, and the Insurance Act amendments since 2010, do not touch the sections cited. Still open: the in-force date of Part I of the Constitution was not confirmed from an official source, although the Part is operating law; no AI-specific statement from CIMA, the Ombudsman or OfReg was located, and absence is not proven.

Corrections. If a citation is wrong, a quote is not verbatim, a section has moved, or a rule here would not survive a Cayman desk on a Monday morning, write to talktome@perennos.com with the section number and the source. Corrections to a citation are published within the next decimal version. Disagreements with a rule are answered in writing and, where we change our mind, credited.

13. Glossary

Plain words for the terms this standard uses.

• AI-assisted decision. Any decision where a model's output was an input, whether or not a person made the final call.
• AI register. Your list of every AI use: system, purpose, tier, owner. One line per use.
• Answerable Test. The one question: could you answer for this AI use, with the seven answers, inside the five-day and 72-hour clocks?
• Controller / processor. Terms from the Data Protection Act [2021 Revision]. The controller decides why and how personal data is processed; a processor processes it on the controller's behalf, under a written contract.
• Data subject. The person the personal data is about.
• Electronic agent. In the Electronic Transactions Act [2003 Revision], s.2: a program or other automated means, set up by a person, used to initiate or respond to an electronic record "without review by an individual". An AI that sends messages or takes steps on its own is one.
• Clock file. The record, kept per AI use, that holds the seven answers before they are needed. Section 6.
• Governing body. The board of directors, general partner, LLC manager or board of trustees of a regulated entity, as CIMA's rules use the term.
• Human final say. A person with the authority and the information to disagree with the model, who sometimes does, and whose overrides are recorded.
• Kill-switch. A tested way to stop an AI use at once, by a named person, without the vendor.
• Lawful basis. The Schedule 2 condition (and, for sensitive personal data, the Schedule 3 condition) under the Data Protection Act [2021 Revision] that allows the processing.
• Material outsourcing. A function that, if disrupted, could significantly affect operations, reputation or profitability, as described in CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 3.1.3 and 4.1.
• Named owner. The person in the Islands who answers for an AI use. A role is not enough; a name.
• Personal data breach. In the Data Protection Act [2021 Revision], a breach of security leading to accidental or unlawful destruction, loss, alteration or unauthorised disclosure of, or access to, personal data.
• Red line. A use this standard says Cayman practice does not do, at any tier.
• Reversibility. Whether the harm from a wrong output can be undone by a person in reasonable time. The first of the two tier questions.
• Sensitive personal data. The categories in s.3 of the Data Protection Act [2021 Revision], including health, racial or ethnic origin, political opinions, religious beliefs and offence data.
• Solely automated decision. A decision "based solely on the processing by automatic means of the data subject's personal data" (Data Protection Act [2021 Revision], s.12), with no person deciding.
• Substance. Borrowed, as an analogy, from the International Tax Co-operation (Economic Substance) Act [2026 Revision]: real direction and control in the Islands, by people with the knowledge to exercise it, not nominal control.
• The Constitution. Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities. Its rights run against government and public officials, not directly against private businesses.
• Tier. The level of control a use needs, from 0 (Assist) to 3 (Hold), set by reversibility and visibility, not by sector.
• Transfer ground. A reason the Data Protection Act [2021 Revision] accepts for sending personal data to a country without adequate protection: Schedule 4.
• Vendor. Any outside party whose model, service or servers your AI use relies on, wherever they are.

14. Find your decision: the Cayman Considerations

The practical desk behind this standard. Ten decisions Cayman firms face first, each with the source and one thing to do. This is the same material published at perennos.com/cayman-considerations/, kept as written.

Personal Data and Lawful Conditions

Law

Under the Data Protection Act (2021 Revision), entering personal data into an automated tool remains processing, and a supplier arrangement does not remove the controller's duties. The Act sets distinct conditions for ordinary and sensitive personal data, security safeguards and individual rights. Health information is sensitive personal data. Our recommendation: identify the lawful condition and data sensitivity before staff place personal records into any prompt or workflow.

Put it into practice. Map each automated workflow to a documented lawful processing condition and block sensitive personal data where no condition applies.

Sources: Data Protection Act (2021 Revision) https://www.ombudsman.ky/images/pdf/laws_regs/Data_Protection_Act_2021_Rev.pdf (2021-03-31; checked 4 November 2026); Cayman Islands legislation index by subject https://legislation.gov.ky/cms/legislation/index/by-subject.html (2026-11-04; checked 4 November 2026)

Processors and Supplier Contracts

Regulator guidance

Ombudsman guidance states that controller-to-processor arrangements require written contract terms covering controller instructions and security. It also discusses additional provisions for subprocessors, breach reporting and audit or return of data according to the processing and risk. A supplier may act as a controller for some activities, such as model training, rather than a processor throughout. Our recommendation: verify whether a vendor reuses inputs for its own purposes before treating it as a processor.

Put it into practice. Review vendor terms to confirm the supplier's role, disable provider training on business inputs, and document subprocessor and return obligations.

Sources: Contracts between data controllers and data processors https://ombudsman.ky/data-protection-organisation/contracts-between-data-controllers-and-data-processors (Undated; checked 4 November 2026); Data Protection Act (2021 Revision) https://www.ombudsman.ky/images/pdf/laws_regs/Data_Protection_Act_2021_Rev.pdf (2021-03-31; checked 4 November 2026)

Overseas Transfers and Hosted Paths

Mixed

Under the eighth data protection principle, controllers must verify destination adequacy or a Schedule 4 condition and review recipient roles. For in-scope entities, CIMA's April 2023 cybersecurity guidance treats cloud computing as outsourcing and addresses overseas-jurisdiction risk, data isolation, recovery and exit plans. Our recommendation: trace where prompts, uploads, logs, backups and support access reside before enabling a hosted assistant.

Put it into practice. Document the geographic path of prompts, logs and backups, and verify transfer safeguards and exit plans before live deployment.

Sources: Eighth data protection principle: international transfers https://ombudsman.ky/data-protection-organisation/data-protection-principles/eighth-data-protection-principle-international-transfers (Undated; checked 4 November 2026); Statement of Guidance: Cybersecurity for Regulated Entities https://www.cima.ky/upimages/regulatorymeasures/SOG-CybersecurityforRegulatedEntities_1681937002.pdf (2023-04; checked 4 November 2026)

Separate Breach and Incident Clocks

Mixed

Cayman entities can face two distinct reporting clocks. Under the Data Protection Act, controllers must report a personal data breach to the Ombudsman and affected individuals without undue delay and no longer than five days after they should, with reasonable diligence, have been aware. Missing details do not justify waiting. Under CIMA's April 2023 Cybersecurity Rule 8.1, in-scope entities must immediately notify CIMA in writing of a material or potentially material incident, and no later than 72 hours after discovery. Our recommendation: track both requirements independently in your incident playbook.

Put it into practice. Maintain separate escalation triggers for immediate written CIMA notification, with its 72-hour outer limit after discovery, and the five-day Ombudsman deadline, counted from when the controller should reasonably have been aware.

Sources: Personal data breaches https://ombudsman.ky/data-protection-organisation/personal-data-breaches (Undated; checked 4 November 2026); Rule: Cybersecurity for Regulated Entities https://www.cima.ky/upimages/regulatorymeasures/Rule-CybersecurityforRegulatedEntities_1681936712.pdf (2023-04; checked 4 November 2026)

Banking, Funds and Outsourcing Scope

Mixed

CIMA's April 2023 Cybersecurity Rule requires documented cybersecurity controls and oversight for in-scope entities. Its Outsourcing Statement of Guidance sets expectations for oversight and materiality assessments. Both instruments exclude regulated mutual funds and private funds, and the Outsourcing guidance also excludes private trust companies. Those exclusions do not remove funds' separate governance, data protection or applicable AML responsibilities. Fully managed Class B, C and D insurers have special scope under Cybersecurity Rule 3.4. Our recommendation: assess an automated tool against the entity's actual licence and service materiality rather than assuming every tool is material outsourcing.

Put it into practice. Check entity licence scope and exemptions first, then assess service materiality and administrator or manager coverage in writing.

Sources: Rule: Cybersecurity for Regulated Entities https://www.cima.ky/upimages/regulatorymeasures/Rule-CybersecurityforRegulatedEntities_1681936712.pdf (2023-04; checked 4 November 2026); Statement of Guidance: Outsourcing, Regulated Entities https://www.cima.ky/upimages/regulatorymeasures/SOG-OutsourcingRegulatedEntities_1681935949.pdf (2023-04; checked 4 November 2026); Statement of Guidance: Cybersecurity for Regulated Entities https://www.cima.ky/upimages/regulatorymeasures/SOG-CybersecurityforRegulatedEntities_1681937002.pdf (2023-04; checked 4 November 2026); Banking Services Regulatory Measures https://www.cima.ky/banking-services-regulatory-measures (Undated; checked 4 November 2026); Statement of Guidance: Corporate Governance—Mutual Funds and Private Funds https://www.cima.ky/upimages/regulatorymeasures/SOG-Corp.Gov.forMFsandPFs_1682541877.pdf (2023-04; checked 4 November 2026)

AML Programmes and Financial Sanctions

Regulator rule

CIMA's Compliance Programme Rule and Financial Sanctions Rule took effect on 18 September 2026. They operate within the wider AML framework, including the Anti-Money Laundering Regulations (2025 Revision). Delegating tasks to third parties does not transfer ultimate compliance responsibility, and the AMLCO retains responsibility for the compliance programme despite delegating duties. Customer due diligence requires reliable independent sources, and sanctions screening requires documented rationale and re-screening after list updates. Our recommendation: never let a model summary or score replace primary records, official sanctions lists or documented officer review.

Put it into practice. Require human verification against primary documents and current official sanctions lists before clearing any onboarding or screening match.

Sources: Rule: Effective Compliance Programme for the Prevention and Detection of ML, TF, PF for Financial Services Providers https://www.cima.ky/upimages/regulatorymeasures/Rule-EffectiveComplianceProgrammeforMLTFPF_1784899845.pdf (2026-09-18; checked 4 November 2026); Rule: Compliance with Financial Sanctions and Targeted Financial Sanctions https://www.cima.ky/upimages/regulatorymeasures/Rule-CompliancewithFinancialSanctionsandTargetedFinancialSanctions_1784899917.pdf (2026-09-18; checked 4 November 2026); CIMA AML/CFT frequently asked questions https://www.cima.ky/aml-cft-faqs (2026-09-18; checked 4 November 2026); Cayman Islands legislation index by subject https://legislation.gov.ky/cms/legislation/index/by-subject.html (2026-11-04; checked 4 November 2026)

Legal Practice Confidentiality and Supervision

Regulator guidance

The official Guidance Notes to the Cayman Islands Legal Services Code of Professional Conduct, published on 15 April 2026, address client confidentiality, privilege, employee training, technology and outsourcing risks, competence and supervision. Our recommendation: approve tools before staff disclose client material and require a qualified lawyer to verify every cited authority and filing.

Put it into practice. Require partner approval before placing client records into external tools and verify all generated authorities against primary law reports.

Sources: The Cayman Islands Legal Services Code of Professional Conduct, 2026 https://legalservicescouncil.ky/cayman-islands-legal-services-code-of-professional-conduct-2026/ (2026-04-26; checked 4 November 2026); Guidance Notes to the Cayman Islands Legal Services Code of Professional Conduct, 2026 https://gov.ky/documents/35692/0/Guidance%2BNotes%2BTo%2BThe%2BCayman%2BIslands%2BLegal%2BServices%2BCode%2BOf%2BProfessional%2BConduct,%2B2026..pdf/875bb32d-06b7-5a03-32a5-9ffdcc586217?t=1777061969891 (2026-04-15; checked 4 November 2026)

Patient Safety and Clinical Responsibility

Law

The Health Practice Act (2026 Revision), published on 5 February 2026, sets the practitioner and facility framework, with the legislation index listing Amendment Act 14 of 2026 (in force) and Amendment (No. 2) Act 25 of 2026 (marked not in force). Health records are sensitive personal data under the Data Protection Act. Our recommendation: separate administrative drafting from clinical judgment, keep a registered clinician accountable for care decisions, and apply strict data protection controls to patient records.

Put it into practice. Restrict automated tools to approved administrative tasks unless a registered clinician reviews and takes responsibility for clinical outputs.

Sources: Health Practice Act (2026 Revision) https://gov.ky/documents/35692/0/Health%2BPractice%2BAct%2B(2026%2BRevision),%2B.pdf/9f7911fd-313e-1c89-fea3-aca8dd150c1e?t=1770319174266 (2026-02-05; checked 4 November 2026); Data Protection Act (2021 Revision) https://www.ombudsman.ky/images/pdf/laws_regs/Data_Protection_Act_2021_Rev.pdf (2021-03-31; checked 4 November 2026); Cayman Islands legislation index by subject https://legislation.gov.ky/cms/legislation/index/by-subject.html (2026-11-04; checked 4 November 2026)

Insurance Conduct and Benefit Decisions

Mixed

CIMA's insurance regulatory measures link conduct, governance, internal controls, cybersecurity, outsourcing and AML or sanctions duties, while Cybersecurity Rule 3.4 sets specific scope for fully managed Class B, C and D insurers. The Health Insurance Commission separately regulates health insurance and monitors approved insurers. Our recommendation: map obligations to the specific insurance licence, manager arrangement and personal-data footprint, and assign an accountable reviewer for any automated output affecting cover, underwriting, benefit decisions or policyholder notices.

Put it into practice. Designate a named reviewer for automated underwriting, cover or benefit outputs and verify how managed insurer rules apply.

Sources: CIMA Insurance Regulatory Measures https://www.cima.ky/insurance-regulatory-measures (Undated; checked 4 November 2026); Rule: Cybersecurity for Regulated Entities https://www.cima.ky/upimages/regulatorymeasures/Rule-CybersecurityforRegulatedEntities_1681936712.pdf (2023-04; checked 4 November 2026); Health Insurance Commission https://gov.ky/web/dhrs/hic (Undated; checked 4 November 2026)

OfReg ICT and Sector Remit

Regulator guidance

OfReg's Information and Communications Technology remit covers telecommunications and broadcasting licensing, .ky domain administration, duties under the Electronic Transactions Act and work with government and stakeholders on national cybersecurity. It is not described as a general licensor for automated software tools. Our recommendation: raise ICT licensing, domain and applicable electronic-transactions questions with OfReg while directing data protection, financial, legal and health governance to the relevant sector regulator.

Put it into practice. Match each compliance question to the regulator with statutory authority over your sector and data processing activity.

Sources: OfReg Information and Communications Technology https://www.ofreg.ky/ict (Undated; checked 4 November 2026)

15. For a Cayman Firm: from the weekly brief

Each week the brief takes real, dated AI security and safety stories and says what they mean for a Cayman firm. This standard carries the latest edition's boxes; the brief itself stays current at perennos.com/briefing/. Latest edition: 4 November 2026, “Patch Windows, Agent Containment, Monitoring and Privacy Testbeds”.

Faster Exploitation Narrows Patch Windows and Exposes Unofficial Tools

UK National Cyber Security Centre · 3 November 2026

Cayman law firms, banks, administrators, insurers and clinics can share dependencies on internet-facing software and hosted suppliers with firms internationally. When vulnerability exploitation accelerates, firms need clear authority for emergency patching and visibility into which unapproved tools staff use with client or patient records.

What to do. Assign an owner for urgent patching, inventory internet-facing systems and critical suppliers, and give staff an approved route for handling business data.

These are UK threat assessments, not confirmation of a breach at a Cayman Islands organisation.

Read the primary source https://www.ncsc.gov.uk/collection/annual-review-2026/understanding-threat/ai-impact-cyber-threat

UK Plans Consultation on Managing Agent Loss of Control

UK Government, DSIT/DCMS · 3 November 2026

Cayman organisations connecting software agents to email, document stores or operational workflows face practical containment risks now. An assistant given broad credentials can send unauthorised messages, expose confidential records or alter workflows before staff intervene.

What to do. Grant agents minimum necessary permissions, require human approval for consequential external actions, and rehearse revoking access while preserving audit logs.

This paper sets out a proposed UK policy consultation, not a Cayman Islands statutory or regulatory requirement.

Read the primary source https://www.gov.uk/government/publications/cyber-security-and-resilience-network-and-information-systems-bill-factsheets/agentic-ai-resilience-measures

OpenAI Details Multi-Stage Monitoring After Evaluation Gap

OpenAI · 29 October 2026

For Cayman professional practices and regulated institutions, a supplier's safety summary is a prompt to verify operational controls. Polite outputs do not prove an assistant stayed within bounds; firms need clear action logs, alert ownership and a reliable stop mechanism.

What to do. Require an auditable action log, named escalation owners and a manual stop switch, and test a harmless out-of-scope prompt before live connection.

This is a vendor self-report, and its replay of an earlier incident is a simulation rather than proof that updated controls prevented it.

Read the primary source https://openai.com/index/how-we-detect-frontier-model-misalignment/

NIST Releases Federated Learning Testbed and Soybean Dataset

US National Institute of Standards and Technology · 29 October 2026

Cayman healthcare providers and insurers exploring joint analytics can examine architectures that keep raw records at their source and measure privacy leakage. Even with federated methods, controllers must still assess lawful processing conditions and international transfer rules under the Data Protection Act.

What to do. Test collaborative analytics on synthetic or public data first, measure privacy leakage and accuracy, and document lawful processing before touching personal records.

These are research evaluation tools built around a soybean genomic dataset, not a clinical patient product or a regulatory approval.

Read the primary source https://www.nist.gov/news-events/news/2026/10/privacy-preserving-federated-learning-ppfl-testbed-and-curated-dataset

Part Two: How Perennos can help

Part One is the standard. It sells nothing. This part is separate, and it is the only place we talk about our own work.

A conversation first. Tell us where AI already sits in your firm: the tools staff use, the vendors who process your data, the decisions it touches. We will tell you, plainly, which tier each use sits in and what the standard asks of it.

A governance review against this standard. We build your AI register, tier every use, write your clock file, check your vendor terms against the Data Protection Act and, where you are CIMA-regulated, the cyber, outsourcing and AML rules. Then we test the switch with you, on purpose, so you know it works.

Building the controls. Perennos Tempo runs on a computer in your office and works on your own files. It answers to you. Online helpers see public work only, after a privacy check, and a person approves every online answer. One click shows a record of every job and where it went. Pilots are open; still in testing are profiles, the desktop installer and the network guard. Where you need an assistant that answers only from your approved documents and hands off to your staff, we build it as a custom project, quoted after a fit call.

Training. For boards, partners and the people at the desk: what the standard asks, what the law already says, and how to answer the seven questions on a bad day.

Built and maintained in the Cayman Islands. Email talktome@perennos.com.

Downloads

• Cayman Considerations for Answerable AI, version 1.0 (PDF)
• Cayman Considerations for Answerable AI, version 1.0 (Word, .docx)

Sources

Every law, rule, guidance note, government statement and framework this standard relies on, with a working link. Each was opened and read on the date shown. Where a publisher's site blocked access, an archived copy of the official page is given.

Cayman law: rights, data, technology, employment, health

• Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.1(2)(c) and s.1(3) (checked 5 November 2026)
• Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.9(1) (checked 5 November 2026)
• Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.11(1) (checked 5 November 2026)
• Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.16(1)-(3) (Non-discrimination) (checked 5 November 2026)
• Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.7(1) (Fair trial) (checked 5 November 2026)
• Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.19(1)-(2) (Lawful administrative action) (checked 5 November 2026)
• Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.24 (checked 5 November 2026)
• Cayman Islands Constitution Order 2009 (UK SI 2009/1379), Schedule 2 (the Constitution), Part I Bill of Rights, Freedoms and Responsibilities, s.122 (Freedom of information) (checked 5 November 2026)
• Data Protection Act [2021 Revision], Schedule 1, Part 1 (first to eighth data protection principles); s.5(1) (checked 5 November 2026)
• Data Protection Act [2021 Revision], s.3 (sensitive personal data) and Schedule 3 para 8 (medical purposes) (checked 5 November 2026)
• Data Protection Act [2021 Revision], s.12 (rights in relation to automated decision-making) (checked 5 November 2026)
• Data Protection Act [2021 Revision], s.8(3) (reasons for solely automated decisions) (checked 5 November 2026)
• Data Protection Act [2021 Revision], Schedule 1, Part 1, seventh principle; Part 2 para 3 (processor contract) (checked 5 November 2026)
• Data Protection Act [2021 Revision], s.16 (personal data breaches) (checked 5 November 2026)
• Data Protection Act [2021 Revision], Schedule 1, Part 1, eighth principle; Part 2 paras 4-5; Schedule 4 (checked 5 November 2026)
• Data Protection Act [2021 Revision], s.45 (enforcement orders) and s.46 (failure to comply) (checked 5 November 2026)
• Data Protection Act [2021 Revision], s.55 (monetary penalty), s.54 (unlawful obtaining), s.58 (liability of officers) (checked 5 November 2026)
• Data Protection Regulations [SL 17 of 2019] (printed title: The Data Protection Regulations, 2018, made under the Data Protection Law, 2017), reg.7 (health exemption); reg.3-5 (subject requests); reg.8 (education exemption) (checked 5 November 2026)
• Freedom of Information Act [2021 Revision], s.2 ('record', 'public authority'); s.6(1) (checked 5 November 2026)
• Electronic Transactions Act [2003 Revision], s.14(2)(b), s.15(b) and s.2 ('electronic agent') (checked 5 November 2026)
• Electronic Transactions Act [2003 Revision], s.6 and s.12 (checked 5 November 2026)
• Computer Misuse Act [2015 Revision], s.3(1) and s.3(3)-(4) (checked 5 November 2026)
• Computer Misuse Act [2015 Revision], s.5(1)-(2) (unauthorised modification) and s.7(1) (causing a computer to cease to function) (checked 5 November 2026)
• Evidence Act [2021 Revision], s.29 (criminal) and s.48 (civil): Computers (checked 5 November 2026)
• Information and Communications Technology Authority Act [2026 Revision], s.9(3)(ha)-(hd); s.75; s.76 (checked 5 November 2026)
• Utility Regulation and Competition Act [2024 Revision], s.6(1) (principal functions) (checked 5 November 2026)
• Labour Act [2021 Revision], Operative statute; s.12(1); s.49 and s.51; s.55; s.80(1) (checked 5 November 2026)
• Gender Equality Act [Act 21 of 2011], s.3(1)-(3) (indirect discrimination); s.4(1)-(2) (employment); s.13 (goods and services); s.14; s.16 (checked 5 November 2026)
• Health Practice Act [2026 Revision], s.5(1), s.8, s.15(1)-(2), s.35(1)-(5), s.36(1)(d) (checked 5 November 2026)
• Health Insurance Act [2021 Revision], s.15(1) (termination); s.15A (reduction of benefits); s.21 (direct payment); s.22 (disputes); s.23 (appeals) (checked 5 November 2026)
• Mental Health Act [2023 Revision], s.16 (restrictions on access to post and electronic networks) (checked 5 November 2026)
• Education Act [2024 Revision], whole Act (keyword scan) (checked 5 November 2026)
• Defamation Act [1995 Revision], s.11 (unintentional defamation, offer of amends) and s.11(5) (checked 5 November 2026)
• Copyright (Cayman Islands) Order 2015 (UK SI 2015/795), as amended by the Copyright (Cayman Islands) (Amendment) Order 2016 (UK SI 2016/370), art.2; Schedule Part 1 (excluded provisions) and Part 2 (modifications); applied CDPA 1988 s.9(3) and s.178 (checked 5 November 2026)
• Copyright, Designs and Patents Act 1988 (UK), as extended to Cayman, s.9(3) and s.178 definition of 'computer-generated' (checked 5 November 2026)

Cayman law and regulation: financial services, AML/CFT, legal profession, economic substance

• Monetary Authority Act [2020 Revision], s.34(1)(a) (checked 5 November 2026)
• Monetary Authority Act [2020 Revision], s.34(6) and (7) (checked 5 November 2026)
• Monetary Authority Act [2020 Revision], s.42A(1); s.42B(3)-(4) (checked 5 November 2026)
• Monetary Authority Act [2020 Revision], s.34(8), s.34(17) (checked 5 November 2026)
• Monetary Authority Act [2020 Revision], s.50(1) (checked 5 November 2026)
• CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 3.1 and footnote 1; Rule 3.4 (checked 5 November 2026)
• CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 3.3 (checked 5 November 2026)
• CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 5.1.1 and 5.1.2(a) (checked 5 November 2026)
• CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 6.1.1 (checked 5 November 2026)
• CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 6.5.1(a) and (c); 6.5.2 (checked 5 November 2026)
• CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 7.1 (checked 5 November 2026)
• CIMA Rule: Cybersecurity for Regulated Entities (April 2023) [SL 38 of 2023], Rule 8.1 (checked 5 November 2026)
• CIMA Statement of Guidance: Cybersecurity for Regulated Entities (April 2023), para 7.2.2(a); para 13.1 (checked 5 November 2026)
• CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 1.3 (checked 5 November 2026)
• CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 3.1.3; para 4.1; para 4.2 (checked 5 November 2026)
• CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 14.1 and 14.3 (checked 5 November 2026)
• CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 9.6, 9.7, 9.8 (checked 5 November 2026)
• CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 10.3; para 10.4 (checked 5 November 2026)
• CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 5.8, 5.9, 7.4, 14.1(c) (checked 5 November 2026)
• CIMA Statement of Guidance: Outsourcing - Regulated Entities (April 2023), para 12.1 and 12.2 (checked 5 November 2026)
• CIMA Rule: Corporate Governance for Regulated Entities (April 2023) [SL 39 of 2023], Rule 5.7.7 (checked 5 November 2026)
• CIMA Rule: Corporate Governance for Regulated Entities (April 2023) [SL 39 of 2023], Rule 5.7.4 (checked 5 November 2026)
• CIMA Rule: Corporate Governance for Regulated Entities (April 2023) [SL 39 of 2023], Rule 5.1.1; 5.4.2; 5.9.1 (checked 5 November 2026)
• CIMA Rule: Corporate Governance for Regulated Entities (April 2023) [SL 39 of 2023], Rule 5.6.1(a); 5.6.1(d) (checked 5 November 2026)
• CIMA Rule and Statement of Guidance: Internal Controls for Regulated Entities (April 2023) [SL 40 of 2023], para 8.1; para 11.4.2 (checked 5 November 2026)
• Banks and Trust Companies Act [2025 Revision], s.3 (checked 5 November 2026)
• Banks and Trust Companies Act [2025 Revision], s.18(1)(b) (checked 5 November 2026)
• Insurance Act [Act 32 of 2010], s.8(2)(g) (checked 5 November 2026)
• Insurance Act [Act 32 of 2010], s.4(5)(d) (checked 5 November 2026)
• CIMA Rule: Market Conduct - Insurers, Agents and Brokers (May 2018) [SL 48 of 2018], Rule 5.3.4; 5.3.1; 5.3.6 (checked 5 November 2026)
• CIMA Rule: Market Conduct - Insurers, Agents and Brokers (May 2018) [SL 48 of 2018], Rule 5.4.1; 5.5.4; 5.5.8 (checked 5 November 2026)
• Mutual Funds Act [2025 Revision], s.30(1)(d) (checked 5 November 2026)
• CIMA Rule: Calculation of Asset Values - Regulated Mutual Funds (July 2020) [SL 101 of 2020], Rule 5.6.1; Rule 5.14 (checked 5 November 2026)
• Private Funds Act [2025 Revision], s.16(1); s.16(4)(b); s.15 (checked 5 November 2026)
• CIMA Rule: Calculation of Net Asset Values - Registered Private Funds (July 2020) [SL 102 of 2020], Rule 5.7.1; Rule 5.11 (checked 5 November 2026)
• Securities Investment Business (Conduct of Business) Regulations [2026 Revision], reg 11(1)(a),(b),(d); reg 11(2) (checked 5 November 2026)
• Securities Investment Business (Conduct of Business) Regulations [2026 Revision], reg 8(1) and (2) (checked 5 November 2026)
• Securities Investment Business Act [2020 Revision], s.17(1)(b) (checked 5 November 2026)
• Virtual Asset (Service Providers) Act [2024 Revision], s.9(3)(b) and (c) (checked 5 November 2026)
• CIMA Rule and Statement of Guidance: Market Conduct for Virtual Asset Service Providers (February 2026) [SL 12 of 2026], para 6.7; para 6.2, 6.3, 6.6 (checked 5 November 2026)
• CIMA Rule and Statement of Guidance: Market Conduct for Virtual Asset Service Providers (February 2026) [SL 12 of 2026], para 11.1, 11.3-11.6 (checked 5 November 2026)
• Proceeds of Crime Act [2025 Revision], s.136(1) (checked 5 November 2026)
• Proceeds of Crime Act [2025 Revision], s.139(1) (checked 5 November 2026)
• Anti-Money Laundering Regulations [2025 Revision], reg 9 (checked 5 November 2026)
• Anti-Money Laundering Regulations [2025 Revision], reg 5(1)(a)(viiia); reg 12(1)(e) (checked 5 November 2026)
• Anti-Money Laundering Regulations [2025 Revision], reg 24 and reg 25(2) (reliance on others); reg 31 (checked 5 November 2026)
• CIMA Rule: Effective Compliance Programme for the Prevention and Detection of ML, TF and PF for Financial Services Providers (July 2026), Rule 10.3(f); Rule 10.4; Rule 8.6 (checked 5 November 2026)
• CIMA Rule: Effective Compliance Programme for the Prevention and Detection of ML, TF and PF for Financial Services Providers (July 2026), Rule 7.1; Rule 10.2; Rule 13.2 (checked 5 November 2026)
• CIMA Rule: Compliance with Financial Sanctions and Targeted Financial Sanctions (July 2026), Rule 7.1; 7.11; 7.12 (checked 5 November 2026)
• CIMA Guidance Notes on the Prevention and Detection of ML, TF and PF in the Cayman Islands (February 2024), Section 16, Monitoring Systems, para 22 (p.118 of 245) (checked 5 November 2026)
• Legal Services Act [Act 57 of 2020], s.37(4) and (5) (checked 5 November 2026)
• Cayman Islands Legal Services Code of Professional Conduct, 2026, Rule 5.1 (checked 5 November 2026)
• Cayman Islands Legal Services Code of Professional Conduct, 2026, Rule 1.1; Rule 8.5; Rule 8.6 (checked 5 November 2026)
• Cayman Islands Legal Services Code of Professional Conduct, 2026, Rule 2.1; Rule 2.4(a); Rule 4.1(c)(ii) (checked 5 November 2026)
• Guidance Notes to the Cayman Islands Legal Services Code of Professional Conduct, 2026, Guidance Note G.5.4 (numbering interleaved in extraction) (checked 5 November 2026)
• Directors Registration and Licensing Act [Act 10 of 2014], s.11(2)(a); s.12 (checked 5 November 2026)
• International Tax Co-operation (Economic Substance) Act [2026 Revision], s.4(2)(b) (checked 5 November 2026)
• International Tax Co-operation (Economic Substance) Act [2026 Revision], s.4(3)(a)-(e) (checked 5 November 2026)
• International Tax Co-operation (Economic Substance) Act [2026 Revision], s.4(4) (checked 5 November 2026)

Cayman government statements and international frameworks

• Parliamentary Question No. 83 (answered 6 March 2026), Artificial Intelligence Act, answer, pp.1-3 (checked 5 November 2026)
• Parliamentary Question No. 83 - civil service AI policy, answer p.2 (checked 5 November 2026)
• Parliamentary Question No. 36 (answered 2 October 2026), AI risks, Part 2 of answer (checked 5 November 2026)
• Government press release: Task Force to Energise Cayman's Digital Transformation (9 October 2025), press release (checked 5 November 2026)
• Cayman Compass, 10 March 2026: Government targets 2027 for Cayman Islands AI legislation (news reporting), article (archived copy: https://web.archive.org/web/20260311113733/https://www.caymancompass.com/2026/03/10/government-targets-2027-for-cayman-islands-ai-legislation/) (checked 5 November 2026)
• Regulation (EU) 2024/1689 (AI Act), Art. 2(1)(a)-(c) (Official Journal text via EU Publications Office; eur-lex.europa.eu blocks automated fetch) (checked 5 November 2026)
• Regulation (EU) 2024/1689, Art. 3(1) definition of AI system (checked 5 November 2026)
• Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, Art. 4 AI literacy (checked 5 November 2026)
• Regulation (EU) 2024/1689, Art. 5(1)(a)-(h) prohibited practices (headings, paraphrased) (checked 5 November 2026)
• Regulation (EU) 2024/1689, Art. 6(2) and Annex III points 3, 4, 5(b), 5(c) (checked 5 November 2026)
• Regulation (EU) 2024/1689, Art. 9(1)-(2) risk management (checked 5 November 2026)
• Regulation (EU) 2024/1689, Art. 14(1) human oversight (checked 5 November 2026)
• Regulation (EU) 2024/1689, Art. 26(1), (2), (5), (6), (7) deployer obligations (checked 5 November 2026)
• Regulation (EU) 2024/1689, Art. 50(1), (2), (4) transparency (checked 5 November 2026)
• Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, Art. 113 application dates (checked 5 November 2026)
• NIST AI Risk Management Framework 1.0 (NIST AI 100-1), Part 1 and Part 2 (Core) (checked 5 November 2026)
• NIST AI 600-1, AI RMF: Generative Artificial Intelligence Profile, cover; risk list (checked 5 November 2026)
• OECD Recommendation of the Council on Artificial Intelligence (OECD/LEGAL/0449), Principles 1.1-1.5 (checked 5 November 2026)
• UK: A pro-innovation approach to AI regulation (white paper, Command Paper 815), principles (checked 5 November 2026)
• UK: A pro-innovation approach to AI regulation: government response (6 February 2024), paras 1-11 (checked 5 November 2026)
• Bermuda Monetary Authority, Consultation Paper: The Responsible Use of Artificial Intelligence in Financial Services in Bermuda (14 August 2026), paras 2 and 43 (checked 5 November 2026)

International standards and treaties (archived official pages)

• ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system, iso.org standard page (General information) (archived copy: https://web.archive.org/web/20261023081751/https://www.iso.org/standard/42001) (checked 5 November 2026)
• Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, coe.int convention page (archived copy: https://web.archive.org/web/20261016195158/https://www.coe.int/en/web/artificial-intelligence/the-framework-convention-on-artificial-intelligence) (checked 5 November 2026)
• UK Government press release: UK signs first international treaty addressing risks of artificial intelligence, press release, 5 September 2024 (checked 5 November 2026)

What other jurisdictions do

• Model Artificial Intelligence Governance Framework, Second Edition (IMDA/PDPC, Singapore), paras 3.14-3.17 (pp.30-31) (checked 5 November 2026)
• Model Artificial Intelligence Governance Framework, Second Edition - ISAGO self-assessment guide, Foreword (ISAGO paragraph), p.7 area of text (checked 5 November 2026)
• Model AI Governance Framework for Generative AI (IMDA and AI Verify Foundation, Singapore), May 2024 (file copy dated 19 June 2024), Dimension 'Trusted Development and Deployment', 'Disclosure -- Food Labels' (p.13); Dimension 'Accountability', 'Ex Ante / Ex Post' (p.7) (checked 5 November 2026)
• Model AI Governance Framework for Generative AI - Incident Reporting dimension, Incident Reporting, p.17 (checked 5 November 2026)
• Model AI Governance Framework for Agentic AI, Version 1.5 (IMDA, Singapore), published 20 May 2026, updated 5 June 2026, Section 2.2.2 'Design for meaningful human oversight', 'Setting up a system for effective human supervision' (PDF p.29) (checked 5 November 2026)
• Model AI Governance Framework for Agentic AI, Version 1.5 - Authorisation and system-level controls, Section 2.1.2 (PDF p.14 and p.24) (checked 5 November 2026)
• AI Verify Testing Framework and AI Verify toolkit (AI Verify Foundation, Singapore), AI Verify Foundation page 'AI Verify Testing Framework' (checked 5 November 2026)
• Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT) in the Use of Artificial Intelligence and Data Analytics in Singapore's Financial Sector (MAS, 2018), Accountability principles 7-11; Transparency principles 12-14 (PDF p.6) (checked 5 November 2026)
• Veritas Initiative (MAS-led consortium) - FEAT assessment methodology and Veritas Toolkit, MAS page 'Veritas Initiative' (checked 5 November 2026)
• Guidelines on Artificial Intelligence Risk Management (MAS), final, 7 October 2026, para 4.12 (risk materiality: impact, complexity, reliance); para 4.5 (AI inventory); transition per Response to Feedback para 13.5 (checked 5 November 2026)
• Guidelines on Artificial Intelligence Risk Management (MAS), final, 7 October 2026 - contingency plans and kill switches; transition periods, para 5.3; Response to Feedback paras 13.4-13.6 (7 October 2026) (checked 5 November 2026)
• Consultation Paper P017-2025 on Guidelines on Artificial Intelligence Risk Management (MAS), November 2025, Preface; Response to Feedback para 1.1-1.3 (checked 5 November 2026)
• Directive on Automated Decision-Making (Treasury Board of Canada Secretariat), Appendix B Impact Assessment Levels and Appendix C Impact Level Requirements; date modified 2025-06-24, Appendix B; Appendix C; s.6.1.1; s.1.2.1 (checked 5 November 2026)
• Directive on Automated Decision-Making - peer review and explanation (Appendix C), Appendix C rows 'Peer review (s.6.3.7)' and 'Explanation (s.6.2.3)' (checked 5 November 2026)
• Bill C-27, Digital Charter Implementation Act, 2022 (Artificial Intelligence and Data Act, Part 3) - Parliament of Canada LEGISinfo, LEGISinfo page for C-27 (44-1) (checked 5 November 2026)
• General-Purpose AI Code of Practice - Safety and Security chapter (European Commission AI Office), published 10 July 2025, Commitment 1 (Measure 1.1); Commitment 4 (acceptance, safety margin); Commitment 9 (serious incident reporting); Commitment 10 (checked 5 November 2026)
• General-Purpose AI Code of Practice - Transparency chapter (European Commission AI Office), 10 July 2025, Measure 1.1 and Model Documentation Form (checked 5 November 2026)
• General-Purpose AI Code of Practice - Copyright chapter (European Commission AI Office), 10 July 2025, Measure 1.2 (lawful access, paywalls), Measure 1.3 (robots.txt) and Measure 1.5 (point of contact and complaints) (checked 5 November 2026)
• Recommendation on the Ethics of Artificial Intelligence (UNESCO), adopted November 2021, and Readiness Assessment Methodology (RAM) incl. RAM 2.0, UNESCO Global AI Ethics and Governance Observatory, 'AI Readiness Assessment Methodology' page (checked 5 November 2026)
• Ethical impact assessment: a tool of the Recommendation on the Ethics of Artificial Intelligence (UNESCO, 2023), UNESCO publication page, 28 August 2023 (checked 5 November 2026)
• Hiroshima Process International Code of Conduct for Organizations Developing Advanced AI Systems (G7), 30 October 2023, Actions 2, 4, 7 (checked 5 November 2026)
• Hiroshima AI Process (HAIP) Reporting Framework (OECD), launched 7 February 2025, OECD.AI 'Hiroshima AI Reporting Framework' overview page (checked 5 November 2026)
• Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), Vilnius 5 September 2024, Art. 16(1)-(4) Risk and impact management framework; Art. 26 Effective oversight mechanisms (checked 5 November 2026)
• CETS No. 225 - status, UK signature and territorial application (Council of Europe Treaty Office), Treaty Office chart of signatures and ratifications, status as of 04/11/2026; Art. 30 and Art. 32 of the Convention (checked 5 November 2026)
• Artificial Intelligence: Model Personal Data Protection Framework (Office of the Privacy Commissioner for Personal Data, Hong Kong), June 2024 (published 11 June 2024), Part II para 30 and section 2.2 (printed p.28); Part III Figure 18 AI Incident Response Plan (printed p.~44-46) (checked 5 November 2026)
• Consumer Protection in respect of Use of Generative Artificial Intelligence (HKMA circular to all authorized institutions), 19 August 2024; extends the 5 November 2019 Big Data Analytics and AI guiding principles, circular pp.2-4, principles 1-4 (checked 5 November 2026)
• Data Protection Regulations, Consolidated Version No. 2 (in force 1 September 2023) - Regulation 10 'Personal Data Processed Through Autonomous and Semi-Autonomous Systems' (DIFC), Reg 10.2.2(b)(ii); Reg 10.3.2(a); Reg 10.3.3(d) (checked 5 November 2026)
• DIFC Regulation 10 Accreditation and Certification Framework; Consultation Paper No. 3 (June 2026) Updated Data Protection Regulations (DIFC Authority / Commissioner of Data Protection), Consultation Paper paras 1, 6-8, 15 (checked 5 November 2026)
• ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system (Edition 1, December 2023), ISO catalogue page (iso.org/standard/42001) (checked 5 November 2026)
• ISO/IEC 23894:2023 Information technology - Artificial intelligence - Guidance on risk management (Edition 1, February 2023), ISO catalogue page (iso.org/standard/77304.html) (checked 5 November 2026)
• OMB Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust (3 April 2025) - CURRENT, Section 4(a)-(b) (PDF pp.13-17) incl. practices i-vii; Section 5 definition; Section 6 presumed categories (checked 5 November 2026)
• OMB Memorandum M-24-10, Advancing Governance, Innovation, and Risk Management for Agency Use of Artificial Intelligence (28 March 2024) - RESCINDED by M-25-21, Section 5(c) minimum practices; Section 6 categories (checked 5 November 2026)

Small-jurisdiction peers, security references and Cayman bodies

• Government press release: Task Force to Energise Cayman's Digital Transformation (9 Oct 2025), press release; launch event 6 Oct 2025 (checked 5 November 2026)
• Parliamentary Question No. 83 (answered 6 March 2026) - membership and remit of the National Digital Transformation Taskforce, answer p.1-2 (checked 5 November 2026)
• Parliamentary Question No. 36 (answered 2 October 2026) - Task Force status and Minister, Part 2 of answer (checked 5 November 2026)
• Minister responsible for innovation / digital (Hon. Isaac Rankine MP), gov.ky press release 9 Oct 2025; PQ 36 (2 Oct 2026) (checked 5 November 2026)
• gov.ky press release: Digital Transformation Strategy Task Force launches national survey (11 May 2026), press release (checked 5 November 2026)
• CIMA Annual Report 2025 (tabled at parliament.ky, June 2026) - technology outlook and VASP and Fintech Innovation Unit, 2026 Outlook list, printed near p.76; Structure section (checked 5 November 2026)
• Cayman Islands Chamber of Commerce Economic Forum 2026 (Radio Cayman notice, 20 Jan 2026), news notice (checked 5 November 2026)
• Caymanian Times, 27 Apr 2026: Cayman Islands AI Strategy Showcased at Major U.S. Political Science Conference, article (checked 5 November 2026)
• Jersey Financial Services Commission, Guidance on the use of AI in Jersey's financial services sector (July 2026), executive summary; JFSC industry update 14 July 2026 (checked 5 November 2026)
• Guernsey Financial Services Commission, Policy Statement - Use of Artificial Intelligence (29 January 2026), statement text (checked 5 November 2026)
• Caribbean Telecommunications Union, Caribbean AI Task Force Final Report: Toward Harmonized AI Policies and Recommendations for the Caribbean (2026), front matter (checked 5 November 2026)
• OWASP Top 10 for LLM Applications 2025, LLM01, LLM02, LLM06 mitigation sections (checked 5 November 2026)
• OWASP Agentic AI - Threats and Mitigations (Agentic Security Initiative, v1.0 guide), resource page (checked 5 November 2026)
• OWASP Top 10 for Agentic Applications for 2026, resource page, 9 Dec 2025 (checked 5 November 2026)
• MITRE ATLAS (Adversarial Threat Landscape for AI Systems), repository README; data release ATLAS-2026.10 (checked 5 November 2026)
• NIST AI 100-2 E2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, Abstract (checked 5 November 2026)
• C2PA Content Credentials: C2PA Technical Specification (version selector lists 2.5), normative text on redaction of hard-binding assertions (2.2 page) (checked 5 November 2026)
• UK NCSC and CISA (with partners), Guidelines for secure AI system development, version 1.0, Secure development, Secure operation and maintenance (checked 5 November 2026)
• UK AI Security Institute, and Inspect evaluation framework, home page; Inspect welcome page (checked 5 November 2026)