Confidential

Success should not be measured by the breadth of principles agreed, but by how far the Dialogue closes the distance between principle and practitioner. The global conversation is not short of principles. The OECD AI Principles, UNESCO Recommendation, EU AI Act, NIST AI RMF, and ISO/IEC 42001 already converge on a familiar vocabulary. What remains unresolved is the operational layer beneath it. Three outcomes would mark genuine progress. Implementation coherence, not more principles. Institutions operating across jurisdictions increasingly face overlapping, non-reciprocal AI obligations while the technology outpaces any single regulatory cycle. The Dialogue should commit to interoperability mechanisms — shared conformity assessment methodologies, mutual recognition of AI audits, and a common taxonomy for high-risk system classification. Structural inclusion of Global South practitioner voices. Much of the current conversation reflects the regulatory imaginations of jurisdictions that also host the largest AI developers. The communities most exposed to AI-driven decisions in financial access, social benefits, and law enforcement are rarely in the rooms where standards are set. Practitioners carrying statutory accountability for AI system adequacy in regulated sectors — particularly in developing economies — should be structurally represented, not consulted as an afterthought. A living mechanism, not a moment. AI capability cycles run in 12-to-18 month generations. Governance drafted on three-to-five year review cycles is structurally behind the technology before it is enacted. A successful first Dialogue should establish how its principles remain responsive — with defined update triggers and practitioner feedback pathways between plenaries. I carry personal statutory accountability for the adequacy of compliance systems that increasingly rely on AI, in a jurisdiction where no regulatory standard yet defines what "adequate" means for such systems. That gap is not unusual — it is the quiet reality of AI deployment inside regulated sectors globally. The Dialogue's first test is whether it names that gap honestly.

• Interoperability of governance approaches
• Transparency, accountability, and human oversight
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Safe, secure and trustworthy AI

These four priorities reflect the operational reality of AI deployment inside a regulated sector carrying statutory accountability, viewed from a Global South jurisdiction. Interoperability of governance approaches is the most urgent because the fragmentation risk is already materialising. Institutions operating across jurisdictions are navigating overlapping AI obligations with no shared taxonomy, no mutual recognition of audits, and no coherent mapping between frameworks. For regulated entities in developing economies, this fragmentation compounds existing regulatory burden without proportionate risk reduction. A global dialogue that does not address interoperability leaves this burden to accumulate unchecked. Transparency, accountability, and human oversight sits at the core of practitioner reality. In my jurisdiction, I carry personal statutory accountability for the adequacy of AML/CFT systems increasingly dependent on AI - yet no regulatory standard defines what adequacy means for such systems. This accountability gap is not unique to South Africa. Globally, sectoral officers are being held responsible for AI outcomes without the governance infrastructure to discharge that responsibility defensibly. Social, economic, ethical, cultural, linguistic and technical implications of AI is the only theme broad enough to hold the distributional question honestly. AI systems that deny financial access, misclassify customers, or produce false positives against specific demographic groups generate harms that are not technical failures but equity failures. Frameworks like Ubuntu and Data Justice - invoked in South Africa's Draft National AI Policy - offer non-Western philosophical anchors for assessing these harms collectively rather than individually. The Dialogue should engage these framings substantively. Safe, secure and trustworthy AI is the umbrella. In sectors where AI mediates access to the financial system, safety is not abstract - it has immediate consequences for individuals wrongly excluded and for illicit flows undetected. This theme keeps the conversation grounded in high-stakes deployment realities, not governance theory.

The sectoral accountability gap. The listed themes address AI governance at the framework level but not the question of how statutory accountability for AI outcomes is discharged inside existing regulated sectors. Compliance officers, data protection officers, clinical governance leads, and financial supervisors are increasingly personally accountable for AI-driven outcomes under pre-AI legislation that was never designed with AI in mind. This is not the same as "accountability" in the abstract - it is the live question of what a named officer is supposed to do when an AI system produces a consequential output and no regulatory standard yet defines whether that output was adequate, explainable, or compliant. Know Your Agent (KYA) and agentic AI in regulated transactions. As AI agents begin to initiate financial transactions, execute contracts, and interact with regulated services autonomously, existing identification frameworks - KYC, beneficial ownership, authorised representative regimes - face a structural question they were not built to answer: who is the accountable human principal behind an autonomous agent, and how is that verified? This is emerging fastest in crypto, but the pattern will generalise. It sits awkwardly across safety, accountability, and interoperability without being captured fully by any. The pace-of-change asymmetry. AI capability cycles run in 12-to-18 month generations. Regulatory cycles run in three-to-five year generations. This asymmetry is not a theme - it is the structural condition under which every theme operates. Without an explicit mechanism for responsive governance, any framework produced by this Dialogue will be behind the technology before it is enacted. The Dialogue should name this asymmetry and commit to mechanisms that address it directly.

South Africa's financial services sector — banks, insurers, asset managers, CASPs, and intermediaries — operates under FSCA conduct supervision, Prudential Authority oversight, FIC Act accountability, SARB exchange control, and POPIA. AI systems are already embedded across the sector in credit scoring, insurance underwriting, transaction monitoring, sanctions screening, fraud detection, and customer risk rating. None of these regulatory instruments yet define adequacy standards for AI-driven decisioning. The Draft National AI Policy (Government Gazette No. 54477, April 2026) is in public comment but will not produce sectoral regulations before 2027/2028. Interoperability. South African FSPs operate across jurisdictions subject to overlapping AI governance overlays — FATF Recommendation 16 Travel Rule obligations for CASPs, EU GDPR-adjacent automated decision-making standards, US model risk management expectations, and emerging EU AI Act extraterritorial reach. Each jurisdiction is developing its own AI regime. Without interoperability, compliance cost compounds without proportionate risk reduction — a particular burden on developing-economy institutions already navigating FATF grey-list remediation. Transparency, accountability, human oversight. Sectoral accountable officers — MLCOs under FIC Act section 42, Key Individuals under FAIS, Information Officers under POPIA — carry personal statutory accountability for the adequacy of systems increasingly dependent on AI. POPIA section 71 restricts solely automated decision-making. Yet no South African regulatory instrument defines what adequate AI explainability, bias testing, or human-in-the-loop design looks like. Practitioners are accountable for a standard that does not yet exist. Social, economic, cultural implications. AI-driven false positives in credit, insurance, and transaction monitoring can restrict financial access for specific demographic groups — a collective harm that Ubuntu and Data Justice framings in the Draft Policy gesture toward but do not operationalise. SA's eleven official languages are underrepresented in training data. Safe, secure, trustworthy AI. The opportunity is significant: SA's regulated sectors are sophisticated, its regulators engaged, and the Draft Policy creates a window for practitioner input.

The Dialogue's most important role is to occupy the coordination space that no existing forum currently holds. The OECD, G7 Hiroshima Process, G20, UNESCO, Council of Europe, and standards bodies like ISO and NIST each produce valuable work, but none convene the full global membership with the mandate to reconcile their outputs. The Dialogue can play that reconciling role without duplicating what already exists. A clearinghouse for interoperability. The Dialogue can map where existing frameworks converge, where they diverge, and where the divergence is substantive versus merely terminological. This mapping is foundational infrastructure. Without it, mutual recognition, shared conformity assessment, and audit portability remain aspirational. A legitimate forum for Global South agenda-setting. Most current AI governance venues were designed around developed-economy membership and reflect developed-economy priorities. The Dialogue sits within the UN system, which gives it the institutional legitimacy to elevate concerns — data sovereignty, language representation, AI-enabled financial exclusion, capacity asymmetry — that other forums treat as secondary. Using that legitimacy deliberately would be a distinguishing contribution. A feedback channel between practitioners and frameworks. International cooperation often fails at the implementation layer because the practitioners operating inside regulated sectors have no formal route to surface operational realities back to framework-setters. The Dialogue could establish a structured feedback mechanism — sectoral practitioner panels, implementation case studies, empirical data on governance friction — that gives frameworks a feedback loop they currently lack. A commitment device for responsive governance. Finally, the Dialogue can do something no national regulator can do alone: hold member states accountable to the pace of AI development. A published annual assessment of implementation progress — against agreed benchmarks — would create quiet but durable pressure toward coherence.

The Dialogue enters a landscape already populated by substantive initiatives. Its added value lies not in duplication, but in connection. Within the UN system. The Dialogue's most natural partner is its sibling mechanism, the Independent International Scientific Panel on AI, which provides the evidence base that the Dialogue can translate into shared understanding and cooperation . UNESCO's Recommendation on the Ethics of AI and its AI Readiness Assessment Methodology — already implemented in over seventy countries — offer a tested diagnostic framework the Dialogue can build upon. The Global Digital Compact provides the foundational mandate. ITU's technical standards work and the AI for Good platform provide the convening infrastructure. Beyond the UN system. The OECD AI Principles, the G7 Hiroshima AI Process, the Council of Europe Framework Convention on AI, and the Global Partnership on AI each hold valuable normative and technical work but reflect developed-economy membership. The African Union Continental AI Strategy, Smart Africa initiatives, and regional frameworks from ASEAN and the Organization of American States represent Global South agenda-setting that deserves equal weight at the Dialogue table. In the standards ecosystem. ISO/IEC 42001, the NIST AI Risk Management Framework, and sectoral standards bodies (FATF for financial crime, Basel Committee for banking) produce the implementation-layer detail that governance frameworks depend on but rarely reference directly. The Dialogue's distinctive added value is its capacity to act as the connective tissue across this landscape — mapping convergence and divergence, surfacing where existing initiatives reinforce or contradict one another, and creating a single legitimate forum where Global South and developed-economy priorities can be reconciled rather than parallel-tracked. Its unique institutional position gives it something no other forum has: universal membership paired with the authority to speak back to the fragmentation the Secretary-General has identified as a core risk.

Meaningful stakeholder contribution depends on structure. Without it, "multi-stakeholder" becomes a framing that concentrates influence among those with the resources to participate continuously. Stakeholder roles. Member States bring regulatory mandate and should surface implementation friction, not only policy positions. Academia contributes frameworks and evidence. The private sector brings operational visibility but must be structurally balanced against civil society to prevent capture. Civil society and human rights organisations bring the voice of affected communities and need dedicated, resourced participation. Technical communities bring implementation detail. Sectoral practitioners — compliance officers, data protection regulators, sectoral supervisors — carry statutory accountability for AI outcomes in regulated sectors and are currently the least represented group in global AI governance venues. Format recommendations. First, establish sectoral practitioner panels alongside the main plenary to surface implementation realities directly into the Dialogue's record. Second, commit to regional pre-consultations with dedicated resourcing for Global South civil society and practitioner participation. Virtual participation alone is insufficient where time zones, language, and institutional support are asymmetric. Third, publish structured implementation case studies between plenaries to keep the Dialogue grounded in operational reality. Fourth, use thematic working groups to sustain momentum between annual sessions.

Global AI governance discussions currently overrepresent state representatives, large AI developers, and academic institutions from developed economies. Several voices with direct stakes in AI outcomes remain structurally underrepresented. Sectoral practitioners carrying statutory accountability. Compliance officers, data protection regulators, clinical governance leads, and sectoral supervisors are personally accountable for AI outcomes under pre-AI legislation, yet have no formal route into global AI governance. Inclusion mechanism: dedicated sectoral practitioner panels with standing representation. Affected communities themselves. People subject to AI-driven decisions in financial access, credit, insurance, social benefits, immigration, and policing are rarely in the rooms where governance is set. Inclusion mechanism: resourced civil society representation and community impact testimony as a formal input category. Global South civil society and regulators. Developed-economy civil society organisations and regulators dominate current forums. Inclusion mechanism: regional pre-consultations with travel, translation, and institutional support funded by the Dialogue secretariat, not by participating organisations. Indigenous and non-Western philosophical traditions. Frameworks like Ubuntu, Buen Vivir, and Indigenous data sovereignty principles offer governance lenses that the dominant rights-based and utilitarian frames do not capture. Inclusion mechanism: philosophical pluralism working group to operationalise these frameworks into governance standards. Linguistic minorities. Languages underrepresented in training data produce systematic disadvantage in AI-mediated services. Inclusion mechanism: language equity as a standing Dialogue theme, with published metrics on AI system performance across major language families. Workers and trade unions. Workers affected by AI-driven workplace monitoring, hiring algorithms, and automation are largely absent from governance venues focused on developers and regulators. Inclusion mechanism: formal stakeholder category for organised labour. Inclusion is a structural question, not a rhetorical one. The Dialogue's credibility depends on building these mechanisms into its design from the outset.

Innovation in engagement format should serve a specific purpose: surfacing perspectives that standard plenary structures systematically miss. Living case study sessions. Rather than abstract panel discussions, dedicate plenary time to structured case studies where sectoral practitioners walk participants through real AI governance decisions — the data they had, the frameworks they applied, the decisions they made, and the outcomes. This turns implementation reality into shared learning. Scenario simulations. Multi-stakeholder groups work through concrete governance scenarios — for instance, a hypothetical AI-driven financial exclusion incident spanning three jurisdictions — and negotiate responses in real time. This exposes framework friction that abstract discussions do not. Reverse briefings. Instead of developers and states briefing civil society, reverse the format: affected communities brief policymakers and developers on lived experience of AI systems. Dedicated time, structured format, resourced participation. Open challenge problems. Publish concrete, unresolved governance questions ahead of each plenary — for instance, "What does 'adequate human oversight' mean for a credit scoring AI?" — and invite structured submissions from all stakeholder groups. The best submissions are debated in plenary. Asynchronous deliberation platforms. Between plenaries, host moderated digital spaces where the substance of plenary discussion can continue. This extends the Dialogue beyond the privileged few who can attend in person. Youth and next-generation panels. People under thirty will live with the governance choices made now. Formal standing representation, not tokenistic inclusion. Published friction logs. Each session should produce a public record of disagreements — not only consensus. Disagreement is data. Format innovation that does not change who is heard is decoration, not innovation.

ISO/IEC 42001 - the first certifiable AI management system standard. Offers a pathway for organisational accountability ahead of sectoral regulation. NIST AI Risk Management Framework - voluntary but widely adopted. Its living-document model keeps pace with technology in a way statutory frameworks cannot. Singapore's Model AI Governance Framework and AI Verify - pairs principles with an operational testing toolkit, closing the gap between governance intent and implementation. UNESCO's AI Readiness Assessment Methodology - implemented in over seventy countries, offering a tested diagnostic for national AI capacity without ranking jurisdictions. EU AI Act - the most comprehensive statutory framework, valuable as a risk-classification reference even for non-EU jurisdictions. Council of Europe Framework Convention on AI - the first legally binding international AI treaty, grounding governance in human rights obligations.