Felaris Global LLC

Artificial intelligence governance frameworks currently under development at the United Nations, at the Organisation for Economic Co-operation and Development, at the European Union, and at national sector-regulator level implicitly assume enterprise-scale organisational infrastructure. A Chief Information Security Officer, a Chief AI Officer, an in-house legal and compliance function, a dedicated governance committee, and the budget to operate them are treated as prerequisites rather than as options. That assumption does not hold in the tier where a substantial share of regulated economic activity actually takes place. Mid-market regulated companies, with annual revenues between fifty million and five hundred million United States dollars, carry material compliance obligations under sector law. They also carry limited governance infrastructure. Operations in smaller jurisdictions compound the implementation gap. This submission offers three observations and four recommendations. The first observation is that framework design today treats the Fortune 500 reference organisation as the implementer of record. That choice produces instruments which mid-market regulated companies cannot adopt at the depth the frameworks intend, regardless of board-level commitment. The second observation is that organisational capacity in smaller jurisdictions is a distinct problem from sovereign capacity. Both matter. Only one is currently addressed in the Dialogue's structure. The third observation is that artificial intelligence governance instruments which do not interoperate with existing sector frameworks produce implementation-level resistance, duplication, and contradictory guidance. Validated regulated industries have spent two decades building computerised systems validation, electronic records discipline, and audit trail practice. Artificial intelligence oversight should be built on that foundation, not parallel to it. The four recommendations that follow are practical and address the Dialogue's named clusters. Each sits within the scope of what the Co-Chairs' summary can reasonably reflect.

• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

Felaris Global's priorities for the Dialogue reflect where the firm operates and where the gaps are largest for the clients it serves: mid-market regulated companies in medical device, pharmaceutical, manufacturing, and PE-backed financial services, across the United States, Europe, and the English-speaking Caribbean. AI capacity-building is the most pressing priority at the organisational level. Frameworks being built today assume enterprise-scale infrastructure. Most of the regulated economy does not look like that. Mid-market companies with material compliance obligations under FDA, EMA, EU MDR, SOX, and the newer AI instruments carry limited governance infrastructure. Organisational capacity-building, calibrated to this tier and to smaller jurisdictions, is the mechanism that makes nominal framework adoption substantive. The social, economic, and technical implications area matters because AI governance cannot be decoupled from the economic reality of the organisations that implement it. Frameworks that ignore mid-market economic constraints and the technical realities of validated systems do not translate into operational discipline. Felaris brings operator perspective on that translation. Interoperability of governance approaches is the defining design question for the next generation of AI governance instruments. Regulated industries have spent two decades building discipline under sector frameworks: 21 CFR Part 11, EU Annex 11, IEC 62304, SOX, and ISO/IEC 27001 and 42001. Instruments that interoperate with these frameworks produce additive governance. Instruments that run parallel produce duplicative documentation and contradictory guidance, and meet line-level resistance from staff who have internalised the existing discipline. Felaris Global's contribution is grounded in engagements with mid-market regulated operators and government entities in the Caribbean. The firm is actively building implementation tooling and training programmes, alongside the fractional executive advisory capacity needed to deploy them at mid-market scale.

Two issues sit across the listed themes and warrant explicit recognition in the Dialogue's work. The first is the organisational implementation gap at mid-market scale. Every theme above describes an intended outcome: safety, capacity, oversight, interoperability. None addresses the precondition, which is the organisational capacity to implement. Frameworks adopted nominally by mid-market regulated companies without the governance infrastructure of the Fortune 500 produce reporting artefacts rather than risk reduction. This gap is cross-cutting because it affects the realisation of every theme. A framework for safe, secure, and trustworthy AI that cannot be implemented at the tier where most regulated economic activity takes place does not produce safety outcomes in that tier. Mid-market regulated companies, particularly those operating subsidiaries in smaller jurisdictions, warrant treatment as a distinct implementation reality rather than a residual case. The second is shadow AI and embedded AI in third-party software. Artificial intelligence is increasingly deployed inside software-as-a-service products used by regulated operators, rather than procured as standalone AI systems. Organisations discover AI adoption after a vendor update, outside the procurement process. This reality cuts across safety, oversight, interoperability, and capacity. Current frameworks assume the organisation knows what AI it is operating, an assumption that is weak at mid-market scale. Model inventory and third-party AI risk management, addressed at the point of consumption rather than at procurement, would strengthen the Dialogue's output across themes. A third emerging issue is the sequencing between multilateral work and active sector regulation. Four dates sit on the same twelve-month horizon: FDA and EMA joint AI principles published January 2026, ISO/IEC 42001 in active adoption, the Dialogue's first session convening July 2026, and the EU AI Act's high-risk obligations taking effect August 2026. Interoperability guidance arriving after sector enforcement has begun is less load-bearing than guidance arriving earlier. Timing matters.

The impact is concrete across the mid-market regulated sector Felaris Global serves, and across the Caribbean region where the firm operates. On capacity-building, the most significant challenge is the widening gap between board-level AI governance expectations and mid-market organisational capacity. Boards in PE-backed medical device, pharmaceutical, and financial services companies receive the same questions the Fortune 500 receives from auditors, regulators, and customers. The operating scale to answer is not present. Hiring a Chief AI Officer and an in-house AI legal function is not financially rational for a $100M revenue company. The opportunity is delivery models calibrated to this tier: fractional executive capacity and tooling that reduces manual compliance overhead, complemented by training that builds internal competency. On social, economic, and technical implications, sector-regulator movement in the United States is fast: FDA and EMA published joint AI principles in January 2026, the SEC and PCAOB are examining AI in financial reporting, and SOX implications attach to AI-assisted controls. Mid-market absorptive capacity is limited, and non-adoption is not an option. In the English-speaking Caribbean, local AI frameworks are not yet in place, and regional operators build posture on multilateral reference material. A well-timed Dialogue output gives Caribbean regulators a reference architecture to adapt, saving the cost of building one from scratch. On interoperability, validated industries carry two decades of computerised systems validation discipline under 21 CFR Part 11, EU Annex 11, and IEC 62304. The challenge is that AI instruments risk being built parallel to, rather than additive to, this discipline. Duplicative documentation and contradictory guidance increase cost without reducing risk. An explicit interoperability mandate in the Dialogue's output would preserve existing inspection discipline while lowering mid-market implementation cost. It would also raise AI oversight from aspirational to operational.

International cooperation on AI governance is essential. The Dialogue can play four specific roles, each drawing on its position as a multilateral convening above the level of individual national and sector frameworks. First, it can coordinate interoperability between emerging AI governance instruments and established sector frameworks. The EU AI Act, the US NIST AI Risk Management Framework, UNESCO's Recommendation on the Ethics of Artificial Intelligence, ISO/IEC 42001, and the FDA/EMA joint AI principles are each moving on independent tracks. The Dialogue is uniquely positioned to require that new instruments interoperate with 21 CFR Part 11, EU Annex 11, IEC 62304, SOX, EBA model risk guidance, and ISO/IEC 27001 as a design constraint rather than as an afterthought. No other venue sits above that full set of instruments. Second, the Dialogue can provide a reference architecture for smaller jurisdictions. Regional governments in the Caribbean, in parts of Africa, Latin America, and the Pacific depend on multilateral reference material because local AI frameworks do not yet exist. A coherent Dialogue output lowers the cost of local framework development and accelerates credible regional posture. Third, the Dialogue can create structured space for voices that enterprise-tier and national-delegation participation does not surface. Mid-market regulated operators and sector-specific practitioners see implementation realities that do not register in Big Tech or G7 conversations. The same is true of operators in smaller jurisdictions. The Dialogue's private-sector track should include these voices deliberately. Fourth, the Dialogue can set the sequencing. Sector enforcement is accelerating on a visible twelve-month horizon. Dialogue guidance that arrives ahead of implementation deadlines shapes practice. Guidance that arrives after enforcement has begun describes it. The July 2026 session is early enough to shape; subsequent cycles should preserve that cadence.

The Dialogue should build on existing work across five layers. Its added value is a single multilateral convening that threads them together under UN authority. On multilateral principles and ethics, UNESCO's Recommendation on the Ethics of Artificial Intelligence, the OECD AI Principles and the OECD.AI observatory, and the Council of Europe Framework Convention on AI provide the values-level reference architecture. The Dialogue should treat these as the baseline and avoid restating them. On standards and management systems, ISO/IEC 42001 and the broader ISO/IEC JTC 1/SC 42 portfolio provide the operational layer regulated operators actually implement against. The Dialogue's interoperability work should treat ISO/IEC 42001 as a core reference point. On sector-specific mechanisms, FDA and EMA joint AI principles, the International Medical Device Regulators Forum, IOSCO, the European Banking Authority, and the Public Company Accounting Oversight Board are each advancing AI oversight within their respective sectors. The Dialogue should connect with these directly rather than build parallel guidance. On regional bodies, CARICOM, the African Union Continental AI Strategy, ASEAN's AI Guide, and the Inter-American Development Bank are the channels for translating Dialogue outputs into regional adoption. They should be engaged early. On government-led processes, the Global Partnership on AI, the G7 Hiroshima AI Process, and the Bletchley-Seoul-Paris AI Safety Summit series have produced technical and policy material the Dialogue can reference. The Dialogue's added value is convening authority across these tracks and explicit attention to implementation reality at mid-market scale and in smaller jurisdictions, which no existing mechanism addresses.

The Dialogue's structure should recognise that "stakeholders" is not a single category and that different stakeholder types produce different kinds of evidence. National delegations carry the enforcement context and political mandate. Intergovernmental bodies, including UNESCO, ITU, OECD, and ISO/IEC, provide normative and technical grounding. Civil society contributes rights-based and human-impact evidence. Academic research adds empirical and normative analysis. The technical community supplies standards, reference implementations, and open-source capacity. The private sector, by contrast, is usually treated as a single category. It should not be. Big Tech brings model-level visibility. Enterprise regulated operators carry governance-at-scale experience, while mid-market regulated operators see the implementation reality at the tier where most regulated economic activity takes place. Startups add deployment and product-level evidence. Treating these as distinct voices within the private-sector track would sharpen the evidence the Dialogue receives and prevent enterprise perspectives from substituting for mid-market ones. Structural recommendations: The Dialogue should establish thematic working groups mapped to the seven thematic areas, each with a designated rapporteur and a standing mandate between sessions. Sessions themselves should be cyclical, annual or biannual, with intersessional output rather than silence between convenings. The open written-submission portal should operate as a standing mechanism. Online consultations would broaden participation beyond travel-capable delegations. Implementation-focused segments should sit alongside principle-level discussions. Sector-regulator liaison seats should be formally integrated. Rotating regional co-chairing surfaces voices from smaller jurisdictions structurally. Published submission records, with summary analyses by the Co-Chairs' rapporteurs, maintain transparency.

Current global discussions on AI governance over-represent large jurisdictions and large organisations. They under-represent three categories of voice. Small and developing states, particularly those outside the G20 and its immediate partners, are structurally under-represented. The English-speaking Caribbean is a clear example. CARICOM member states, including Trinidad and Tobago and Jamaica, carry material operating exposure to AI adoption through local industry and through multinational subsidiaries. They also carry serious talent and growth potential in technology and services sectors that is not currently reflected in the international record. Jamaica's digital transformation programme and Trinidad and Tobago's technology and financial services base should be at the Dialogue's table as participants with speaking roles, rather than as subjects of capacity-building discussions. Mid-market regulated operators across all jurisdictions are the second under-represented voice. The Dialogue's evidence base skews to enterprise and Big Tech perspectives, which produces framework design calibrated to organisational realities that most regulated companies do not have. Practical implementation voices are the third. Workers deploying AI daily, and compliance officers translating framework text into operational practice, see evidence that principle-level consultation does not capture. End users affected by automated decisions carry equally valuable experience. These voices could be included through several mechanisms. Funded participation for small-jurisdiction delegations should cover travel and preparation costs. Online and written-submission tracks should carry equal weight with in-person participation. Structured solicitation through CARICOM, the African Union, ASEAN, and Pacific Islands Forum secretariats would bring regional bodies into the evidence base early. Dedicated seats for mid-market regulated operators, distinct from enterprise seats, would surface implementation reality. Targeted outreach in Trinidad and Tobago, Jamaica, and other Caribbean jurisdictions, where operational AI adoption is accelerating, should be part of the secretariat's standing engagement plan.

Traditional multilateral formats produce statements. The Dialogue needs formats that produce evidence the Co-Chairs' summary can build on. Five formats would shift engagement from performative to substantive. The first is implementation case clinics. Each thematic track opens with an anonymised operational case drawn from written submissions, for example a mid-market regulated company implementing framework X or a smaller-jurisdiction ministry adapting reference material for local adoption. The panel works the case in front of the room. Operators, regulators, academics, and civil society dissect the same evidence. Principles are tested against practice rather than declared. The second is distributed regional hubs. The Geneva plenary connects live to regional hubs: Port of Spain or Kingston for the Caribbean, Nairobi or Addis Ababa for Africa, Singapore or Jakarta for ASEAN, Santiago or Panama for Latin America. Participants contribute from their region at parity with in-room participants. Travel cost ceases to be the filter on whose voice is heard. The third is cross-stakeholder panels with mandated composition. Every panel includes a regulator, a mid-market operator, an enterprise operator, a civil society representative, and an academic. Single-category panels should not be permitted. This prevents enterprise substitution for mid-market perspective. The fourth is co-drafting sessions for interoperability matrices. Working groups co-author mapping tables between emerging AI instruments and existing sector frameworks in session. Outputs are published as working drafts for comment. This shifts the Dialogue from declaration to artefact. The fifth is a structured asynchronous pre-Dialogue phase. Four to eight weeks of moderated online thematic discussion before each in-person session, with a rapporteur-produced digest presented at plenary opening. Asynchronous participation broadens the evidence base before anyone steps into the room.

Four categories of approach have demonstrated that AI governance can move from principle to operational practice. The first is the validated-industry precedent in pharmaceutical and medical device sectors. FDA 21 CFR Part 11, EU Annex 11, and the Good Automated Manufacturing Practice framework have produced two decades of audited discipline around computerised systems, covering system description, validation lifecycle, change control, and electronic records integrity. Applied to AI systems, this precedent gives an immediate foundation for human oversight and auditability. The Dialogue should reference it rather than re-derive it. The second is ISO/IEC 42001:2023, the AI management system standard. It provides an auditable, scale-appropriate baseline that mid-market operators can implement without enterprise governance infrastructure. The Dialogue should treat it as the management-system reference for interoperability work. The third is regulatory sandboxes. The EU AI Act mandates sandboxes in every Member State from August 2026. Singapore's IMDA AI Verify provides an open-source testing toolkit that operationalises governance principles. The UK AI Safety Institute and Brazil's sandbox pilots offer working examples. Sandboxes give operators a controlled environment to test AI systems before full deployment, which closes the gap between principle and practice. The fourth is cross-regulator coordination mechanisms. The UK Digital Regulation Cooperation Forum brings the CMA, ICO, Ofcom, and FCA together to coordinate on cross-cutting AI questions. This model is directly transferable to multilateral coordination between sector regulators on AI oversight. At operator level, three practices are proving effective at mid-market scale: fractional executive accountability for AI, shared compliance platforms across PE portfolios, and AI registers modelled on the Amsterdam and Helsinki municipal examples. These approaches make AI governance tractable at tiers where dedicated infrastructure is not economically feasible.