Research-led vision and strategy, 29 October 2026. A proposal for discussion, not adopted government policy. This develops the initial research brief, registry catalogue and joint Compuconomy–MANX research. The legal map distinguishes original documents, versions and outstanding gaps. Reading coverage records the limits of the folder review: indexing both archives is not exhaustive reading.
Vision and strategic choice
The Isle of Man should aim to become a leading specialist jurisdiction for accountable agentic enterprise: a place where businesses delegate useful work to AI under identifiable responsibility, clear authority, governed data rights and enforceable remedies, while residents share in the resulting prosperity.
An agentic business uses software that pursues objectives through tools and delegated actions: gathering documents, reconciling records, seeking quotations or preparing transactions. The economic value lies in accepted work completed. The institutional questions are who authorises the work, what the system may do, whose information it uses, who bears losses and how authority can be withdrawn.
The Island should develop services that make those answers credible across borders. Its registration businesses offer relevant experience in identifying entities and assets, maintaining records, dealing with representatives, recording interests and serving international customers. That is a plausible foundation for a new export business. Whether customers will pay for the proposed agent services remains to be tested.
Start with maritime and aviation administration, company-service workflows and licensed data products. Establish demand and measured reliability before creating a statutory agent register. Keep responsibility in existing legal persons. The initial programme does not require AI legal personhood, and the reviewed laws do not establish it.
What the data tells us
Retain dates and definitions: fleet scale, inspection performance and commercial awards measure different things.
| Evidence | Strategic meaning and limit |
|---|---|
| The Ship Registry's 2025 annual report records 800 vessels: 301 merchant ships, 318 pleasure yachts, 30 commercial yachts, 59 fishing vessels and 92 small ships. It reports 11.94 million gross tons and 20th globally by gross tonnage. | A dated, registry-reported size ranking and international customer base; not an independently audited ranking or 800 prospective AI customers. Annual report, printed p 9 |
| The official 2024 enterprise plan described the ship register as 19th globally by tonnage using Clarksons. | A dated size indicator, not a verified 2026 rank. Department plan |
| UK statistics place the combined Crown Dependencies 17th by trading-fleet deadweight at end-2025; Isle of Man accounts for over 99.9% of that deadweight. | A close grouped proxy for Manx scale, not a separately published exact Manx rank. DfT statistics |
| Paris MoU's 2025 list places Isle of Man 17th, with 362 inspections and six detentions in 2023–2025; effective 1 July 2026. | A defined regional inspection-performance comparison, distinct from fleet size and a universal quality score. Table, method and effective date |
| Isle of Man in Numbers 2025 records 284 aircraft on the register at 31 March 2025. | A dated fleet baseline. Cumulative registrations and a historical sixth-largest description cannot establish a current global rank. Official statistics, printed p 21 |
| The aircraft registry identifies new mortgage rules effective 1 October 2026. | Continuing development of legal services around assets; not a general power to register AI or data mortgages. Registry notice |
| Tynwald records Royal Assent for the Foundations (Amendment) Act 2026. An October Chamber report following an event with Digital Isle of Man targets launch in Q1 2027, with January the working aim. | Enactment is distinct from commencement and opening. The event report says regulations are complete and fees await final approval; final made instruments and a live register have not been verified. Parliamentary record, October report |
Most corporate, domestic and regulatory registers have no reliable global league table established here. The opportunity is a service proposition, not a composite ranking. Existing registry success also does not prove spare staff capacity, modern interfaces, insurer appetite or willingness to buy agent credentials. These are first-stage research questions.
Translate registry competence into agent services
The Merchant Shipping Registration Act 1991 addresses representatives, registration, removal, mortgages and specialised registers. Section 60A also provides for ships under construction. This illustrates how law can distinguish objects, interests and lifecycle stages; it does not extend those rights to software. Registry-published Act
| Existing capability | Proposed application | Boundary to preserve |
|---|---|---|
| Corporate identity and representatives | Link an agent deployment to a verified legal principal and accountable contact. | Company registration does not authorise every business activity. |
| Asset identity and change records | Identify a deployment, version, mandate and material changes. | A model name or software hash does not describe all runtime behaviour. |
| Recorded interests and searches | Make relevant authority and permission claims checkable. | An entry creates title, priority or a security interest only where applicable law provides that effect. |
| Technical expertise and supervision | Assess defined tasks and operating configurations independently. | Registration, assurance and permission to conduct a regulated activity remain distinct. |
| Suspension, removal and remedies | Withdraw credentials, publish status and support redress. | Removing an entry cannot stop a copied agent; actual system access must be revoked. |
| International customer relationships | Recruit domain buyers and test counterparty acceptance. | A Manx credential has no automatic international legal recognition. |
Domestic civil records remain purpose-limited public infrastructure. Their inclusion in the legal inventory is not a proposal to pool birth, health, electoral, beneficial-ownership or other restricted information into commercial AI training data.
Five linked records for accountable delegation
Build an interoperable service around five objects. This is a proposed design, not an existing statutory agent register.
- The legal principal. Identify the company or other permitted legal person, jurisdiction, responsible officers, service contact and relevant licences. Use existing registries lawfully. Establish duties and liability for the chosen form and activity; a foundation must not be assumed to be an unrestricted trading company.
- The agent deployment. Identify the software supplier, declared model and tool dependencies, deployment identifier, operator, material version and change history. Distinguish reusable software from a customer's running instance.
- The mandate. Specify permitted tasks, counterparties, data, financial limits, expiry, approvals and escalation. Bind credentials to the mandate. Keep keys and sensitive contracts outside public records. A signature demonstrates control of a key under stated assumptions; it does not alone prove legal authority or the truth of an action.
- The data permissions. Identify the data product, authorised licensor, permitted purposes, restrictions, provenance, update obligations and withdrawal process. Distinguish training, retrieval, analysis and onward distribution permissions.
- The assurance and lifecycle record. State who assessed which configuration, for what tasks, with what limitations and expiry. Record material incidents, complaints and current status. Keep sensitive evidence under controlled access while publishing a useful verification route.
A portable credential could carry these claims. W3C Verifiable Credentials Data Model 2.0 offers a published technical standard to evaluate. Using it does not grant legal recognition or certify safety. Test interoperability between suppliers. W3C recommendation
Operational controls must enforce the mandate: separate tool credentials, limited permissions, transaction limits, specified human approvals, monitoring, recovery and tested revocation. A model, tool, prompt-policy or data-source change can invalidate an assessment. Record the assessed configuration and define triggers for reassessment. Logs provide evidence of recorded events, not proof that every underlying assertion was true.
The data proposition, explained
Useful data services require several layers that are often confused.
| Layer | Example | Required evidence |
|---|---|---|
| Underlying information | Maintenance observations, equipment identifiers, manuals and invoices. | Accuracy, provenance, confidentiality, personal information and third-party rights. |
| Defined data product | A maintained collection with documented boundaries, quality and version. | Contents, exclusions, update process and usefulness to a buyer. |
| Rights and governance | Permission to retrieve selected fields for administration, excluding training and resale. | Licensor authority, lawful processing, contractual terms and enforcement. |
| Registry metadata | Identifier, responsible entity, description, permission summary and current status. | Legal effect of an entry, publication limits, correction and dispute procedures. |
| Agent use | A particular deployment reads permitted records and drafts a work order. | Whether this deployment, purpose and action fit the permission and mandate. |
The registry can help a buyer discover and verify a product while the underlying dataset remains with its holder. Controlled access can supply the permitted content. Metadata may itself be sensitive and needs a publication policy. A recorded claim cannot remedy a missing licence, remove privacy obligations, establish ownership of raw facts or guarantee usefulness.
For example, a vessel manager authorises an agent to compare maintenance records and supplier quotations. The manager supplies authorised records, a publisher licenses manual content and a supplier supplies its quotation. The agent identifies missing information and prepares a purchase recommendation. A named person approves an order above an agreed limit. The system records sources, permissions, recommendation, approval and execution, and escalates inconsistent information or expired permissions. It does not autonomously certify seaworthiness or override a safety authority.
This creates a testable service: lower administrative cost, improved traceability and fewer avoidable errors. The economic asset might be a maintained licensed product, contractual receipts or other legally recognised rights. Financing requires separate examination of transferability, enforceability, valuation, priority and lender demand. Registration cannot guarantee collateral acceptance, accounting recognition, token rights or a liquid data market.
The Data Asset Foundation initiative could complement this design. Before relying on statutory effects, obtain the exact enacted 2026 amendment, commencement instruments, final regulations and current registrar guidance. The downloaded March 2026 white paper explicitly identifies itself as a consultation draft; it explains proposals, not operative rights. Controller/processor registration is a separate compliance system. Consultation draft
Participation economics also need testing. The July 2026 consultation proposes initial fees of £5,000–£10,000 and an annual £15,000 fee per registered asset, excluding entity administration, assurance, legal and technical costs. These are provisional figures, not a confirmed tariff. Early pilots must establish whether the service creates enough additional value to cover full participation costs; the case for low-cost voluntary agent credentials is separate from the economics of statutory data registration. Original fee proposal
Three initial markets
Registry and company-service administration. Develop assistants that explain published requirements with dated citations, identify missing filing information and prepare drafts for authorised review. Start with retrieval and document preparation. Filing acceptance, legal advice, identity determinations and statutory decisions need the relevant authorised people and processes. The MANX navigator is a development starting point; its development results do not establish independent accuracy, staff savings or live government deployment. Development record
Maritime and aviation administration. Recruit operators, managers and service firms for bounded record reconciliation, expiry checks and procurement preparation. Identify a buyer and workflow for each pilot. Keep safety-critical operations and regulatory certification outside the initial autonomous action scope. Test whether identity and mandate verification actually reduces a counterparty's costs.
Governed data products. Help a few rights holders package documented, permissioned datasets for paying users. Begin with commercial information whose rights and sensitivity can be resolved. Measure time to establish rights, quality, buyer acceptance and recurring receipts. Government personal information should not be the default raw material.
These markets can generate demand for legal, compliance, cybersecurity, evaluation, insurance and stewardship services. Large compute investment should follow demonstrated confidentiality, resilience, latency and cost requirements, assessed against external services and selective local provision. Include power, connectivity, switching and recovery costs.
Legal development and external acceptance
Start through existing entities, lawful contracts, voluntary credentials and bounded procurement. Commission a written gap analysis covering agency and automated contracting, corporate duties, consumer protection, negligence, insurance, electronic evidence, privacy, intellectual property, cybersecurity, insolvency and sector obligations. Contracts cannot confer a statutory power that an operator lacks.
UNCITRAL's Model Law on Automated Contracting 2024 is a comparator for attribution and automated transactions; it is not evidence of Manx adoption. Singapore's Model AI Governance Framework for Agentic AI offers a governance comparator for responsibility and bounded deployment; the downloaded copy is version 1.5, published 20 May and updated 5 June 2026. It is guidance, not Manx law. UNCITRAL model, IMDA source, preserved version
Test acceptance with customers, assessors, insurers, payment providers and relevant authorities. State the precise scope of any agreement. Describe portable credentials as verified information rather than a global regulatory passport. EU AI Act territorial scope can reach certain third-country providers and deployments; Manx registration provides no exemption from applicable foreign law. EU AI Act, Article 2
Propose new legislation only after identifying a concrete gap: for example, a necessary registrar power, evidential effect, correction mechanism or enforceable reporting duty. Consult on burdens and recognition. A compulsory register could otherwise add expense without improving control over agents.
Delivery and decision gates
The timetable and numerical gates below are proposals, not funded commitments or measured results.
| Period from approval | Work | Evidence required to proceed |
|---|---|---|
| Months 0–3 | A designated sponsor brings together registry operators, Digital Isle of Man, relevant regulators, legal expertise and customer representatives. Establish authority, capped budget, procurement route and independent evaluation. | Reconciled core law; baseline costs; 20 structured buyer/counterparty interviews; at least three credible pilot buyers; rights and access reviews. Narrow or stop if buyers cannot identify a valued problem. |
| Months 3–6 | Select pilot delivery through an authorised route. Publish a voluntary credential specification, access policy and assessment method. | Demonstrated mandates, revocation and supplier exit; independent test sets; agreed responsibility, complaints and incident processes. No external autonomous action before its required controls work. |
| Months 6–12 | Run up to three controlled pilots with customer owners and independent evaluators, comparing each with the existing process. | Accepted quality, full cost, supervision and correction data; paid demand; recovery exercises. Define unacceptable errors in advance and suspend affected workflows when they occur. |
| Months 12–24 | Scale successful services, pursue justified legal changes and external acceptance, expand paid skills pathways. | Renewals, viable recurring costs, independently assessed reliability, portability and value retained on the Island. Publish failures and expenditure alongside successes. |
Strategy approval does not itself confer power to procure, spend, regulate or sign contracts. Refresh the MANX authority and funding research against current records. National priorities must remain independent of a particular supplier or proposed appointment.
Separate promotion, credential issuance, independent assessment, regulatory decisions and complaints oversight. A provider's product test must not silently become government endorsement. Name accountable delivery and escalation owners, publish conflicts and test deception, unauthorised tools, stale information and misleading credentials.
Economics and participation
Maintain four distinct accounts: the service operator's margin, the customer's improvement, public expenditure and receipts, and residents' labour and capital income. Registrations and transaction volumes are activity measures. Adding them to revenues does not calculate national value created.
For each workflow, estimate annual net benefit as the value of accepted additional output and avoided existing costs, less additional implementation, compute, licensing, supervision, assurance, correction, security and expected-loss costs. Avoid counting both time saved and additional output generated from those same hours. Report realised cash separately from estimated opportunity value, and show uncertainty.
For a registry or assurance service, test whether paying customers multiplied by realised receipts per customer cover servicing, assessment, legal, security, support and renewal costs, plus a reasonable share of initial investment. Run low-demand and high-supervision cases. Fees, professional exports and useful local operations are possible income channels; customer evidence and costs must precede revenue forecasts.
Resident benefit requires deliberate routes: paid work and progression in operations, stewardship and assurance; useful tools for local small businesses; better public services; and ownership where a viable legal and financial structure exists. A resident dividend requires funded net receipts and enforceable entitlements. A data foundation or AI-access entitlement alone does not provide them. This follows Compuconomy's distinction between productive capability and rights to its returns. Revised Agentic Economy chapter
What leadership would mean
The 30 October 2026 international comparison changes the competitive baseline. China has launched a national data-property registration trial alongside local data-IP schemes; Korea and Japan protect qualifying commercial data; Vietnam has a Data Law and exchange rules; the EU has actual registered intermediary organisations. The Island should differentiate through the precise additional legal right, effective governance, independent assurance and useful cross-border transactions. A global “first” or “only” claim is not established by this research. Manx commencement and operational registration remain subject to the separate legal-status audit.
Within two years, assess a chosen specialist market: counterparties accept the credentials, customers renew, assessed tasks meet published standards, limits and revocation work, complaints receive remedies, firms can change suppliers and residents retain more value and capability. Report denominators, periods and assessors.
Use a compact dashboard: accepted task success and serious errors; full cost per accepted task; paying and renewing customers; time to verify authority and permissions; revocation and recovery results; incidents and redress; local skilled work and progression; resident income and ownership receipts; and net public cost. A global number-one claim would require a comparable international dataset.
The immediate decision is to authorise an evidence-building programme with a capped budget, named responsibilities and stop gates. The larger vision is a jurisdiction that makes delegated AI commercially useful and institutionally accountable, and connects its gains to the people who live there.