Three Regulators, One Technology: Comparing Supervisory Approaches to Artificial Intelligence in Insurance in Switzerland, the United States and Ghana

1. Introduction

Artificial intelligence has become the first genuinely simultaneous supervisory problem in the history of insurance regulation. Earlier waves of insurance technology diffused slowly enough that supervisors could learn from one another in sequence: motor rating bureaus, computerised administration, catastrophe models and internet distribution each reached different markets years or decades apart. The current wave is different. The same foundation models, the same vendor platforms and often the same consulting playbooks arrived in Zurich, in Des Moines and in Accra within a span of months. Insurers in mature and emerging markets alike now deploy machine learning in pricing and underwriting, generative systems in claims and customer interaction, and increasingly agentic systems in actuarial and compliance work. Every insurance supervisor on earth is, at this moment, deciding what to do about the same technology.

They are not, however, deciding from the same place. A supervisory response to artificial intelligence is not written on a blank page; it is written into an existing institutional endowment, a legal tradition, a stock of supervisory capacity, an enforcement reputation and a market structure. This paper exploits that variation deliberately. It compares three supervisory responses chosen for maximum institutional contrast rather than similarity: Switzerland, where FINMA has addressed artificial intelligence through Guidance 08/2024 within a consciously principle-based and technology-neutral framework (FINMA, 2024); the United States, where the National Association of Insurance Commissioners has coordinated a model bulletin subsequently adopted across states and is piloting an AI Systems Evaluation Tool for use in examinations (National Association of Insurance Commissioners [NAIC], 2023, 2026); and Ghana, where the National Insurance Commission supervises a rapidly formalising market under the Insurance Act 2021 (Republic of Ghana, 2021) and where the national artificial intelligence strategy assigns the financial sector a dedicated implementation pillar (Republic of Ghana, 2025), but where no insurance-specific artificial intelligence instrument yet exists. The Ghanaian configuration deserves stating precisely at the outset, because the paper's later argument turns on it: as of mid-2026, and to the author's knowledge, the Commission possesses the statutory capacity to act and a national strategy that legitimises acting, and has not yet acted, a combination this paper will read as an open design brief rather than a deficit.

The comparison is motivated by a question that is practical for regulators and theoretical for scholars of regulation. Practically: what should a supervisor that has not yet issued artificial intelligence guidance, of which there are many, and most of them in emerging markets, actually do, given that the two most visible templates on offer emerge from institutional settings unlike its own?

The paper's research question can be stated in one sentence: when institutionally dissimilar insurance supervisors confront the same technology, do their responses converge on the technology's properties, as functionalist accounts of regulation would predict, or diverge along institutional lines, as institutionalist accounts would predict, and what does the answer imply for the design choices of supervisors that have not yet acted? The paper argues, and the evidence supports, the institutionalist reading: the three responses studied here are not early, middle and late points on a single maturity curve but three distinct equilibria, each internally coherent and each rational given its endowment. That finding has a liberating implication for late movers, developed in the final sections: the task is not to catch up along someone else's curve but to design for one's own endowment, and lateness itself carries a specific, exploitable advantage.

The paper contributes in three ways. First, it provides, to the author's knowledge as of mid-2026, the first structured comparison of insurance-supervisory artificial intelligence instruments that includes a Sub-Saharan African regulator as a full analytical case rather than as an afterthought, treating the absence of an instrument as an institutional datum to be explained rather than a deficiency to be deplored; the Ghanaian case, moreover, does not merely add the Global South to the sample but reverses the expected hierarchy, since the jurisdiction without an artificial intelligence instrument turns out to possess the strongest statutory named-person accountability infrastructure of the three.

Second, it extends the framework of institutionally contingent technology adoption hierarchies, developed in the author's earlier work to explain why firms in different institutional environments rationally adopt technology in different sequences, from the regulated to the regulator: supervisors, too, adopt supervisory technology in institutionally contingent hierarchies. Third, it converts the comparison into design guidance, culminating in a proposed architecture for a National Insurance Commission guideline that is deliberately calibrated to Ghanaian supervisory capacity and legal form rather than transplanted from Basel, Bern or Kansas City.

The remainder of the paper proceeds as follows. Section 2 develops the analytical framework and the six comparative dimensions. Section 3 sets out the methodology and case selection logic. Section 4, Section 5 and Section 6 present the three cases, Switzerland, the United States and Ghana, in a common structure. Section 7 conducts the structured comparison. Section 8 develops the theoretical argument that the cases represent distinct institutional equilibria. Section 9 translates the analysis into design guidance for late-moving supervisors and outlines the proposed NIC guideline architecture. Section 10 acknowledges limitations, and Section 11 concludes.

2. Analytical Framework

2.1. From Firm-Level to Supervisor-Level Contingency

The framework of institutionally contingent technology adoption hierarchies (Botchey, 2026b) holds that the sequence in which organisations adopt technologies is not determined by the technologies' intrinsic merits but by the institutional environment in which the organisation operates: the availability of complementary infrastructure, the credibility of enforcement, the depth of skilled labour markets and the structure of trust between transacting parties. In environments with institutional voids (Khanna & Palepu, 1997), the rational adoption hierarchy differs systematically from the hierarchy observed in institutionally dense environments, and behaviour that looks like lag from the outside is often optimisation from the inside.

Existing cross-jurisdictional treatments of artificial intelligence regulation in finance, including international stocktakes and application papers by standard-setting bodies (Financial Stability Board [FSB], 2024; International Association of Insurance Supervisors [IAIS], 2025) and the comparative guides produced by consultancies and law firms, perform valuable cataloguing work: they list instruments, summarise requirements and frequently arrange jurisdictions along an implicit maturity continuum from silence through guidance to binding rules. What they generally do not do is map instruments to the institutional endowments that make them workable, which is precisely the step that determines whether an instrument can travel. This paper's framework is designed to supply that step.

Two established literatures make that step tractable. Classic regulatory studies has long treated instrument choice as conditioned by regulatory capacity, enforcement strategy and the reputational resources of the regulator, rather than by the intrinsic merits of instruments (Ayres & Braithwaite, 1992; Lodge & Wegrich, 2012). And the institutional-voids stream has developed from diagnosis of what emerging markets lack into a design programme for building strategy around what they have (Khanna & Palepu, 1997, 2010). This paper joins the two: it reads supervisory responses to artificial intelligence through the instrument-choice lens, applies the voids logic to the supervisor rather than the firm, and derives from the combination a concrete design space for late movers.

This paper's central analytical move is to apply the same logic to supervisors. A regulatory instrument is itself a technology, a designed artifact intended to change behaviour at acceptable cost, and a supervisor choosing among instrument types, binding circular, non-binding guidance, model law, examination tool, moral suasion, silence, is performing technology adoption under institutional constraint. A principle-based guidance note presupposes supervised institutions with the sophistication to translate principles into controls and a supervisor with the credibility to make vague expectations bite. An examination tool presupposes a large examination workforce and a legalistic compliance culture. A licensing condition presupposes a licensing bottleneck the supervisor actually controls. None of these presuppositions is universal, and an instrument transplanted without its presuppositions does not merely underperform; it can actively mislead, generating paper compliance that consumes scarce supervisory attention while leaving risk untouched.

2.2. Six Comparative Dimensions

The cases are compared along six dimensions, selected to capture both the instrument and its institutional carrier. First, legal form and bindingness: what kind of legal object is the instrument, and what does non-compliance trigger? Second, regulatory philosophy: is the approach principle-based and technology-neutral, or rule-based and technology-specific, and where does it sit on the spectrum from entity regulation to activity regulation? Third, institutional carrier: which body issues, maintains and operationalises the instrument, and how is coordination achieved across jurisdictional fragmentation where it exists? Fourth, accountability locus: on whom does the instrument place responsibility, the institution, the board, a named function, a named individual? Fifth, enforcement mechanism: through what channel does the instrument acquire force, examination, licensing, litigation, reputation, market discipline? Sixth, market context: the structure, sophistication and artificial-intelligence exposure of the supervised market itself. The six dimensions are applied to each case in Section 4, Section 5 and Section 6 and assembled comparatively in Section 7.

A word on why these six, rather than the available alternatives. Classic instrument typologies in regulatory studies classify legal form and coerciveness but abstract from the body that must operate the instrument; artificial intelligence governance taxonomies classify the content of obligations, risk tiers, documentation duties, oversight requirements, but are largely silent on enforcement machinery and market structure. The research question here concerns transferability, whether an instrument can travel between institutional settings, and transferability turns exactly on the variables those frameworks omit: who carries the instrument, through which enforcement channel it acquires force, and against what market it must operate. The six dimensions therefore combine two instrument-facing dimensions (legal form, philosophy) with two carrier-facing dimensions (institutional carrier, enforcement mechanism) and two environment-facing dimensions (accountability locus, market context), which is the minimum set on which the transplantation question can be posed. Terminologically, instrument is used throughout as the generic category, with guidance, circular, bulletin, directive and guideline reserved for the specific legal forms those words name in their respective jurisdictions.

3. Methodology and Case Selection

3.1. Design and Case Selection

The study is a structured, focused comparison in the tradition of qualitative comparative case analysis (George & Bennett, 2005): the same six analytical dimensions are applied to each case, and the cases are selected on a most-different-systems logic. Switzerland, the United States and Ghana differ on nearly every institutional variable of interest, legal family, supervisory architecture, market depth, litigation intensity, actuarial capacity, while facing a common technological stimulus. Under this design, commonalities in supervisory response would suggest technology-driven convergence, while systematic differences tracking institutional variables support the institutional-contingency thesis.

3.2. Sources and Factual Grounding

The paper is conceptual and practical in orientation: its purpose is to develop an analytical framework and derive design guidance, not to report a coded documentary dataset. Its factual grounding is nonetheless direct. The characterisations in Section 4, Section 5 and Section 6 rest on the primary instruments and statutes themselves, all publicly available and cited in the reference list (FINMA, 2023, 2024; NAIC, 2023, 2026; Colorado Division of Insurance, 2023; Republic of Ghana, 2012, 2021, 2025; European Parliament and Council, 2024; Council of Europe, 2024), on the issuing bodies' own communications, and on the international supervisory layer (Financial Stability Board, 2024; International Association of Insurance Supervisors, 2025), with practitioner legal commentary consulted for corroboration in each jurisdiction.

Several factual anchors used in the analysis deserve explicit statement because they are recent and verifiable against the sources cited. FINMA Guidance 08/2024 was published on 18 December 2024 and articulates supervisory observations across seven areas: governance, inventory and risk classification, data quality, tests and ongoing monitoring, documentation, explainability and independent review (FINMA, 2024). The NAIC model bulletin was adopted by the association's Executive Committee and Plenary on 4 December 2023, and twenty-four states and the District of Columbia had adopted it or substantially similar guidance as of the NAIC's April 2026 implementation record; the AI Systems Evaluation Tool is being piloted by twelve states from March to September 2026, with adoption of an updated version anticipated at the 2026 Fall National Meeting (NAIC, 2023, 2026). Ghana's national artificial intelligence strategy was released in December 2025 by the Ministry of Communication, Digital Technology and Innovations and formally launched in April 2026, and organises implementation around eight pillars, among them a sectoral-adoption pillar that names financial services among its key sectors (Republic of Ghana, 2025). Two limitations of instrument-based analysis are flagged now and revisited in Section 10: instruments reveal supervisory intent more reliably than supervisory practice, and the NAIC tool was analysed in its pilot form, so its adopted form may differ.

4. Case One: Switzerland, Principle-Based Integration

4.1. The Instrument and Its Context

Switzerland has, deliberately, no artificial intelligence statute and no artificial-intelligence-specific binding regulation for the financial sector. The Federal Council's approach, set out in the Federal Council's February 2025 decision on the regulation of artificial intelligence (Swiss Federal Council, 2025) and Switzerland's signature of the Council of Europe's framework convention on artificial intelligence (Council of Europe, 2024), favours sector-specific application of existing law, with a legislative consultation draft expected, at the time of writing, toward the end of 2026. Within this posture, FINMA published Guidance 08/2024 on governance and risk management when using artificial intelligence in December 2024 (FINMA, 2024). The guidance is not a circular: it creates no new obligations, but communicates supervisory observations and expectations about how existing obligations, principally the governance and operational-risk requirements that already bind supervised institutions, apply when institutions use artificial intelligence. The guidance itself is addressed to supervised institutions generally; the most developed codification of the underlying operational-risk requirements, Circular 2023/1 on operational risks and resilience (FINMA, 2023), is formally addressed to banks, with insurers subject to equivalent governance and risk management expectations under insurance supervisory law.

Substantively, the guidance concentrates supervisory attention on identified risk areas, model robustness and correctness, explainability, bias, data quality and security, third-party dependency, and articulates supervisory observations and assessments across seven areas: governance, inventory and risk classification, data quality, tests and ongoing monitoring, documentation, explainability and independent review (FINMA, 2024). FINMA has complemented the instrument with organisational capacity, including a dedicated artificial intelligence desk, and with signalling through its annual risk monitor. As part of the same federal programme, the financial administration is reviewing financial market regulation for artificial intelligence gaps (Swiss Federal Council, 2025).

4.2. Reading the Case Through the Six Dimensions

Legal form: non-binding guidance interpreting binding but technology-neutral norms; non-compliance is not sanctionable as such, but failure to meet the underlying governance and organisational requirements is.

Philosophy: strongly principle-based, technology-neutral and proportionate; artificial intelligence is treated as a new source of familiar risk categories rather than a new regulatory object. Institutional carrier: a single integrated supervisor covering banking and insurance, with high internal technical capacity and a small number of sophisticated supervised institutions.

Accountability locus: the institution and its governing bodies; responsibilities must be clearly allocated internally, but the instrument names no natural person, in contrast to the responsible-actuary institution that Swiss insurance supervision applies to technical provisions.

Enforcement mechanism: ongoing supervision, supervisory dialogue and the credibility of an supervisor able to escalate to enforcement under the general framework; litigation plays essentially no role.

Market context: a concentrated, internationally integrated market with deep actuarial and model-risk capacity, where the principal supervisory concern is that governance keeps pace with sophisticated adoption rather than that adoption is reckless.

The internal coherence of the Swiss equilibrium deserves emphasis. Principle-based guidance without new binding rules is a rational instrument where the supervised population is small and sophisticated, where the supervisor's expectations are credible without formal sanction because supervisory dialogue is continuous, and where a broader constitutional preference for technology-neutral, framework legislation constrains the instrument menu. The same instrument issued by a supervisor without continuous-dialogue credibility would be a suggestion, not supervision. Practitioner readings of the guidance corroborate this characterisation: legal and advisory commentaries in the Swiss market present Guidance 08/2024 as an extension of existing governance and operational-risk obligations to a new risk source, emphasising central allocation of accountability, data-quality testing, robustness and independent review rather than any new regulatory object (MLL Legal, 2025; PwC Switzerland, 2025).

5. Case Two: The United States, Coordinated Examination in a Fragmented System

5.1. The Instrument and Its Context

United States insurance regulation is constitutionally fragmented across the states, and the National Association of Insurance Commissioners exists to manufacture coordination without central authority. Its instrument of December 2023, the model bulletin on the use of artificial intelligence systems by insurers (NAIC, 2023), reflects that machinery: it is a template that individual state insurance departments adopt, requiring insurers to maintain a written artificial intelligence systems program with governance, risk management and internal controls proportionate to the insurer's use of artificial intelligence, and clarifying that decisions made or supported by artificial intelligence remain subject to existing law on unfair trade practices, unfair discrimination and market conduct. The bulletin operationalises the association's 2020 principles on artificial intelligence, which commit insurers to systems that are fair and ethical, accountable, compliant, transparent and secure (NAIC, 2020), and practitioner analyses read it as prescriptive in operation, centred on the written program, board-level governance, vendor oversight and documentation (Kennedys, 2025). Twenty-four states and the District of Columbia had adopted the bulletin or substantially similar guidance as of the NAIC's April 2026 implementation record, with several further states maintaining their own pre-existing frameworks. Individual states have gone further on specific fronts, most prominently Colorado's statute and regulations on insurers' use of external consumer data and information sources and algorithms, which impose quantitative bias-testing obligations in life insurance (Colorado Division of Insurance, 2023).

The distinctive second step is the AI Systems Evaluation Tool developed by the NAIC's Big Data and Artificial Intelligence Working Group (NAIC, 2026): a structured instrument for use by examiners in market conduct and financial examinations to gather information on the extent of an insurer's artificial intelligence use, its governance and risk mitigation, potentially high-risk models and input data. The tool is being piloted by twelve participating states from March to September 2026, with adoption of an updated version anticipated at the NAIC's 2026 Fall National Meeting. Where FINMA writes expectations, the NAIC operationalises questions: the American instrument's centre of gravity is the examination, the setting in which state supervision has always exercised its most granular authority.

The backdrop, finally, is litigation: private actions alleging algorithmic discrimination in underwriting and improper automated claims-handling practices have survived motions to dismiss and entered discovery, meaning that in the United States, unlike Switzerland or Ghana, courts are a live, parallel channel of artificial intelligence accountability that shapes insurer behaviour independently of supervisors.

5.2. Reading the Case Through the Six Dimensions

Legal form: a model bulletin acquiring force state by state through adoption, interpreting existing statutory prohibitions; plus an examination tool, which is not a norm at all but a supervisory instrument; plus, in some states, binding technology-specific regulation.

Philosophy: nominally principle-based at the bulletin level but operationally rule-tending, because examination questionnaires and litigation risk convert principles into de facto checklists.

Institutional carrier: a coordination body without direct authority, achieving harmonisation through voluntary adoption, peer pressure and accreditation dynamics; fragmentation is managed, never abolished.

Accountability locus: the insurer as entity, with board-level accountability for the artificial intelligence program; as in Switzerland, no named natural person, though the appointed actuary and market-conduct liability create adjacent personal exposure.

Enforcement mechanism: a triple channel of examination, state enforcement and, distinctively, private litigation, with the litigation channel arguably the strongest behavioural driver.

Market context: the world's largest and most heterogeneous insurance market, with thousands of supervised entities ranging from global groups to single-state carriers, which is precisely why the system's instrument of choice is a scalable, delegable questionnaire rather than bespoke supervisory dialogue.

The American equilibrium is thus coherent on its own terms: in a system of many supervisors, many insurers and strong courts, coordination templates plus examination tooling plus litigation deterrence produce coverage that no single principle-based dialogue could. The cost is the familiar one of checklist supervision, a compliance industry optimising to the questionnaire, and fifty-one opportunities for divergence every time the template evolves.

6. Case Three: Ghana, Strategy Without (Yet) an Instrument

6.1. The Supervisory and Policy Context

Ghana's National Insurance Commission supervises under the Insurance Act 2021 (Act 1061), a modernised statute that consolidated risk-based supervision, strengthened governance and actuarial requirements, including the appointment of approved actuaries and the submission of actuarial valuations, and expanded the Commission's rule-making powers through directives (Republic of Ghana, 2021). The market the NIC supervises is structurally unlike either comparator: several dozen licensed insurers of modest scale, low but growing penetration, an accelerating digital distribution layer, mobile-money-mediated microinsurance, and a thin domestic actuarial profession concentrated in a small number of qualified professionals serving many institutions. Data protection is governed cross-sectorally by the Data Protection Act 2012 (Act 843) (Republic of Ghana, 2012).

On artificial intelligence, the operative national instrument is Ghana's national artificial intelligence strategy covering 2025 to 2035, released in December 2025 by the Ministry of Communication, Digital Technology and Innovations and formally launched in April 2026. The strategy is the product of a consultation process begun in 2023 under the then Ministry of Communications and Digitalisation (Republic of Ghana, Ministry of Communications and Digitalisation, 2023), giving its sectoral direction continuity across a change of government, which strengthens its value as political cover for regulatory action. The strategy organises implementation around eight pillars spanning enablers and accelerators, among them a sectoral-adoption pillar that names financial services among the key sectors in which artificial intelligence uptake is to be driven, the pillar under which insurance applications fall (Republic of Ghana, 2025; on its operationalisation in insurance, see Botchey, 2026a). The strategy is a policy document, not a supervisory instrument: it creates direction and legitimacy for sectoral regulators to act, but it does not itself impose obligations on insurers. As of mid-2026, and to the author's knowledge, the NIC has issued no insurance-specific artificial intelligence directive or guideline, while Ghanaian insurers are already deploying the technology, in distribution and chatbots, in claims triage, and, through vendor platforms, increasingly in pricing and reserving support. Ghana therefore presents the analytically crucial third configuration: a supervised market adopting the technology, a national strategy legitimising regulatory action, statutory powers adequate to act, and an instrument gap.

6.2. Reading the Case Through the Six Dimensions

Legal form: no artificial-intelligence-specific instrument; general statutory duties, directives powers and cross-sectoral data protection law constitute the current legal surface.

Philosophy: not yet chosen, which is the point; the choice is live.

Institutional carrier: a single sectoral supervisor with concentrated authority, short internal distances and directive powers that can bind quickly, but a small technical staff for whom every new supervisory instrument competes with existing risk-based supervision workload.

Accountability locus: under Act 1061, unusually promising raw material, because Ghanaian insurance law already names natural persons, approved actuaries, principal officers, board members subject to fit-and-proper requirements, giving the NIC a personal-accountability infrastructure that the Swiss and American artificial intelligence instruments conspicuously lack.

Enforcement mechanism: licensing, approvals and directives are the credible channels; examination capacity is scarce and litigation is not a practical accountability channel for policyholders.

Market context: institutional voids in the classic sense, thin professional infrastructure, data scarcity, vendor dependence on foreign platforms, but also the absence of legacy artificial intelligence deployments at scale, and hence the absence of an installed base with an interest in weak rules.

The temptation the Ghanaian configuration invites is transplantation: adopt a translated FINMA guidance or a localised NAIC bulletin and declare the gap closed. The six-dimension reading shows why both transplants would likely underperform or mislead, because their institutional presuppositions are absent. Principle-based guidance presumes supervised institutions that can convert principles into control frameworks and a continuous supervisory dialogue to calibrate them; most NIC-supervised insurers have neither model-risk functions nor the staff to build them unaided. Examination tooling presumes examiners; deploying a fifty-page questionnaire into a supervisor with a handful of technical staff generates unread paper. The design question for Ghana is therefore not which template to copy but which supervisory channels actually possess force in the Ghanaian institutional endowment, and the answer, developed in Section 9, runs through the channels the NIC already controls: licensing, approvals, the approved-actuary institution and directive powers.

7. Structured Comparison

Table 1 assembles the three cases across the six dimensions. Three observations organise the reading that follows.

First, the instruments differ most where the institutions differ most, and least where the technology dictates. All three configurations converge on the same underlying risk vocabulary, governance, data quality, model robustness, explainability, third-party risk, because the technology supplies it. They diverge on everything institutional: what kind of legal object to issue, through which body, aimed at whom, enforced how. Convergence at the level of risk taxonomy combined with divergence at the level of instrument design is precisely the pattern the institutional-contingency thesis predicts and the functionalist thesis does not.

Second, none of the three instruments resolves the accountability question at the level of the natural person. The Swiss and American instruments both stop at the institution and its governing bodies, an omission the author has examined in depth elsewhere in the context of statutory actuarial sign-off over agent-produced work (Botchey, 2026c). The comparative irony of Table 1 is that the jurisdiction with no artificial intelligence instrument at all, Ghana, possesses the strongest statutory named-person infrastructure on which such an instrument could be built, because Act 1061's approved-actuary and fit-and-proper architecture already runs accountability through identified individuals whom the supervisor licenses and can therefore reach directly. Named-person accountability is not exotic in financial regulation: the United Kingdom's senior managers and certification regime, created by the Financial Services (Banking Reform) Act 2013, rests on precisely the premise that supervisory reach improves when responsibility attaches to identified individuals (United Kingdom, 2013). The observation here is that Ghanaian insurance law already embeds that premise, while the artificial intelligence instruments of both comparators stop short of it.

Third, the enforcement columns explain the instrument columns. FINMA can afford expectations because its dialogue is continuous and its escalation credible. The NAIC system needs questionnaires because its enforcement is episodic, delegated and scaled across thousands of entities, and it can rely on courts to do deterrence work that supervisors elsewhere must do themselves. The NIC's credible channels are licensing and approvals, which is exactly why Section 9 routes the proposed Ghanaian instrument through them. Instruments are downstream of enforcement endowments, not the reverse.

8. Three Equilibria, Not One Maturity Curve

It is tempting, and common in policy commentary, to arrange supervisory responses on a maturity scale: binding rules as advanced, guidance as intermediate, silence as lagging. The comparison resists that arrangement at every point. Switzerland's non-binding instrument is not a way-station toward binding rules; it expresses a settled constitutional preference for technology-neutral framework law, and the pending federal legislative process may well confirm rather than replace it.

The American questionnaire is not a more evolved form of the Swiss dialogue; it is the form supervision takes when it must be performed by many hands across many jurisdictions at scale. And Ghanaian silence to date is not simply absence: issuing an unenforceable instrument would consume the NIC's scarcest resource, credibility, and a supervisor that waits until it can enforce is behaving rationally, provided the waiting is used to design. The proviso carries real weight, and the risk on its other side should be acknowledged plainly: vendor-driven deployment does not pause for supervisory reflection, and waiting that is not used to design is simply falling behind.

This is the supervisor-level analogue of institutionally contingent adoption hierarchies. A firm in an institutional void rationally adopts technologies in a different sequence from a firm in an institutionally dense environment. In the same way, a supervisor in a capacity-constrained setting rationally adopts supervisory technologies, instruments, tools, examination practices, in a different sequence from FINMA or the NAIC, and judging that sequence against the metropolitan template mistakes difference for delay. Two strands of literature converge on this reading. Methodologically, the pattern is exactly what a most-different-systems design is built to detect: under George and Bennett's (2005) logic, divergent outcomes tracking the variables on which the cases differ, here the institutional endowments, while the common stimulus is held constant, support institutional rather than technological causation. Substantively, the finding extends Khanna and Palepu's (1997) institutional-voids argument from firm strategy to regulatory strategy: just as focused metropolitan strategies misfire in void-laden markets because the intermediary infrastructure they presuppose is absent, metropolitan supervisory instruments misfire where their enforcement infrastructure is absent, and the rational response in both settings is design for the endowment, not imitation of the template. The practical corollary, however, cuts against complacency: equilibria are only defensible while their presuppositions hold. Ghanaian insurers' artificial intelligence adoption is accelerating on vendor rails regardless of supervisory posture, and the equilibrium of designed waiting has a shelf life. The question is what to design.

The comparison also surfaces the specific advantage of moving late, which this paper has elsewhere called the sequencing opportunity. The Swiss and American instruments were both written for a world of predictive machine learning and have been overtaken, visibly, by the agentic turn: neither instrument distinguishes degrees of delegation, and both stop accountability at the institutional layer. A late-moving supervisor is not obliged to inherit those omissions. It can write delegation-aware expectations, distinguishing systems that assist, systems that propose and systems that decide, and condition-based reliance rules for the named professionals its statute already identifies, from the first instrument onward, before any installed base exists to resist them.

Regulatory leapfrogging of this kind has a well-known precedent in the region: mobile-money regulation in East and West Africa did not recapitulate the branch-banking rulebook, and the welfare gains of that endowment-fitted approach are among the best-documented results in development economics (Jack & Suri, 2014; Suri & Jack, 2016). Its designers were not behind; they were unencumbered.

9. Design Guidance: An Architecture for an NIC Instrument

9.1. Design Principles From the Comparison

Five principles follow from the analysis for any capacity-constrained insurance supervisor, stated here in general form. Route obligations through channels the supervisor already controls and credibly operates, licensing, approvals, named-person accreditation, rather than channels that presuppose absent capacity, mass examination, continuous dialogue. Attach accountability to the natural persons the statute already names, because personal accountability economises on supervisory attention: one approved actuary or principal officer with personal exposure substitutes for many examiner-hours. Regulate by delegation level rather than by technology, so the instrument survives technological turnover and scales obligations to autonomy rather than to buzzwords. Impose proportionate, verifiable artifacts, an inventory, a materiality classification, a named accountable person per material system, rather than comprehensive programs that only large insurers can staff. And harvest, do not replicate, foreign instruments: the risk taxonomy converges internationally and can be adopted wholesale; the machinery must be local.

9.2. The Proposed Architecture

Applied to Ghana, these principles yield a guideline architecture in four parts, offered as a public design contribution and deliberately drafted to be operable with the NIC's existing powers and realistic staffing.

• Part one: Register and classify. Each licensed insurer files, annually and on material change, an artificial intelligence system register: every system in use touching pricing, underwriting, reserving, claims or customer decisions, its vendor or internal origin, its delegation level on a defined scale from assistance through proposal to decision, and a materiality classification. The filing is short, structured and reviewable in minutes per insurer; its function is supervisory visibility, which is the binding constraint the NIC currently faces.
• Part two: Name the person. Every system classified material must have a named accountable individual who is either the approved actuary, for actuarial systems, or the principal officer or a designated senior officer approved by the Commission for others; the accountability attaches through the existing fit-and-proper and approval machinery, requiring no new legal category. For actuarial systems operating at the proposal or decision level, the approved actuary's statutory reporting must confirm the conditions under which reliance was placed, aligning the guideline with emerging professional standards on reliance over system-produced work.
• Part three: Condition the reliance. For material systems at higher delegation levels, the guideline specifies minimum reliance conditions, reproducibility of quantitative outputs from logged inputs, traceability of material judgements to identifiable loci, and demonstrated human ability to contest and override before outputs take effect (Botchey, 2026c), framed as conditions the named person must be able to evidence on request, not as documents to be filed. This imports the substance of international expectations, including the risk areas common to the Swiss and American instruments (FINMA, 2024; NAIC, 2023), while binding it to persons and evidence rather than to programs and paper.
• Part four: Gate the perimeter. Vendor artificial intelligence platforms serving multiple licensed insurers in pricing or reserving may be designated by the Commission for enhanced disclosure through the insurers they serve, using existing outsourcing and directive powers; and artificial intelligence representations in product filings, a channel the NIC already controls transaction by transaction, trigger the register and named-person requirements automatically. The perimeter gates convert the NIC's strongest existing chokepoints, approvals and filings, into the instrument's enforcement surface.

The architecture's supervisory economics are its argument: it generates one short filing per insurer per year, concentrates supervisory attention on the material and the autonomous, reuses every enforcement channel the Commission already operates, and reaches the individuals the statute already licenses.

It also positions Ghana, deliberately, ahead of rather than behind the comparators on the dimension that the agentic turn is making decisive, personal accountability for delegated machine judgement, which neither FINMA Guidance 08/2024 nor the NAIC bulletin addresses (FINMA, 2024; NAIC, 2023). The sectoral pillar of the national artificial intelligence strategy supplies the policy mandate (Republic of Ghana, 2025); Act 1061 supplies the powers (Republic of Ghana, 2021); the design supplies the economy.

9.3. Implications Beyond Ghana

Although drafted for the NIC, the architecture generalises to the many supervisors sharing Ghana's endowment profile, concentrated authority, thin examination capacity, named-person statutes, growing vendor-mediated adoption, across Sub-Saharan Africa and beyond. It also feeds back to the mature comparators. The Swiss legislative process now underway and the NAIC tool's post-pilot evolution will both confront the agentic delegation question sooner rather than later, and the register-person-conditions-perimeter structure is portable in the reverse direction precisely because it is built on institutional primitives, filings, named persons, evidence on request, that every insurance supervisor possesses.

10. Limitations

Four limitations bound the analysis. The comparison is of instruments and institutional design, not of supervisory outcomes; whether any of the three configurations better contains artificial intelligence risk in practice is an empirical question that only time and incident data can answer. The NAIC evaluation tool was analysed in its pilot form and its adopted form may differ. The Ghanaian case characterises the supervisory and policy position from the cited instruments and official communications; a systematic survey of actual artificial intelligence deployment across NIC-licensed insurers would materially strengthen the adoption claims and is identified as follow-on work. And the three-case design, chosen for contrast, cannot establish how the framework performs across the middle of the institutional distribution; the European Union's supranational configuration under the Artificial Intelligence Act (European Parliament and Council, 2024), and large emerging markets such as Nigeria, Kenya and South Africa where supervisory artificial intelligence activity is also underway, are natural extensions.

11. Conclusions

One technology, three regulators, three instruments that could hardly be more different: a Swiss guidance that binds nothing and moves everything, an American questionnaire that scales supervision across a continent of jurisdictions, and a Ghanaian silence that is better read as an open design brief than as a gap. The comparison supports a simple but consequential conclusion: supervisory responses to artificial intelligence are institutionally contingent, and the search for a single best-practice template is not merely futile but harmful, because instruments transplanted without their presuppositions produce paper in place of protection. For the late-moving supervisor, the practical message is emancipating. The task is not to imitate Bern or Kansas City but to route new obligations through old channels that already carry force, and lateness, used deliberately, is an asset: the next generation of supervisory instruments will have to reckon with delegation and personal accountability, and nothing prevents a supervisor in Accra from writing that generation first.