UnRisked (Pty) Ltd

The success of the Dialogue hinges on its ability to move beyond abstract principles and develop actionable architecture. This entails creating a framework taxonomy for pre-deployment governance instruments tailored for agentic AI systems. Additionally, it is crucial to establish international standards for AI-generated accountability documentation in regulated sectors. Equally important is the formal recognition of practitioners from the Global South as a structured stakeholder group, ensuring they have ongoing channels for input rather than a single consultation. True success cannot be equated with merely issuing another declaration. The aim is for the Dialogue to leave Geneva with three definitive outcomes: first, the establishment of mutually agreed reference architectures that Member States can adapt to their own contexts; second, a mandate for the Scientific Panel to prioritize addressing accountability gaps in both agentic and multi-agent systems; and third, a commitment that future sessions will evaluate progress based on concrete milestones in governance infrastructure rather than merely on intentions.

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

These four priorities are interconnected in practice. Trustworthiness cannot be verified without a robust accountability infrastructure. Similarly, accountability cannot be enforced without transparency. Moreover, neither concept holds value if governance frameworks are designed in high-resource jurisdictions and then imposed on the Global South without considering the local social, economic, and regulatory realities. My work on developing the Agentic Risk Schedule, Record of Advice, and Shared Cognition Liability frameworks within South Africa's FAIS-regulated financial services sector illustrates that interoperability and accountability are not merely abstract ideals; they are concrete engineering challenges that require practical solutions. The four priorities identified by the Dialogue represent precisely where these solutions need to be developed.

The most critical cross-cutting issue not explicitly captured is the accountability gap specific to agentic and multi-agent AI systems - systems that perceive, decide, and act autonomously without per-step human instruction. Existing governance frameworks assume a human remains meaningfully in the loop at the point of consequence. For an expanding category of deployments, this assumption is false. When AI agents interact with other AI agents in complex pipelines, responsibility becomes distributed across developers, deployers, operators, and users in ways no existing legal or regulatory structure can cleanly assign. This is what I term the Shared Cognition Liabilityâ"¢ problem - a framework I have developed and published, and it is live in financial services, logistics, healthcare, and public administration right now. A second gap is audit independence. Current practice allows AI systems to generate their own audit trails - a structural conflict of interest that undermines every accountability mechanism built on top of it. International governance standards must require that audit infrastructure operate independently of the systems being audited. Both gaps are most acutely visible in regulated sectors in the Global South, where accountability failures have direct consequences for real people, and where practitioners have already begun developing deployable solutions in the absence of global consensus. Those solutions deserve formal recognition in this Dialogue.

South Africa operates a sophisticated financial services regulatory framework - the Financial Advisory and Intermediary Services (FAIS) Act, the Financial Sector Conduct Authority (FSCA), and the Twin Peaks model established under the Financial Sector Regulation Act - that imposes fiduciary accountability on human advisers but has no equivalent mechanism for AI agents acting in advisory roles. The gap is not theoretical. AI-assisted underwriting, claims assessment, and financial advice are already deployed in South Africa's formal insurance and banking sectors, with no governance infrastructure to establish, dispute, or enforce accountability when those systems cause harm. The challenge is compounded by a dual economy: AI adoption in the formal sector is outpacing governance capacity at every level, while the majority of South Africans interact with AI systems through informal and semi-formal channels where oversight is effectively absent. The opportunity is equally significant. South Africa's regulatory environment is demanding enough to stress-test governance frameworks that would fail in more permissive jurisdictions. Frameworks developed here - like the Agentic Risk Schedule, Record of Advice, and Shared Cognition Liabilityâ"¢ frameworks developed by UnRisked (Pty) Ltd - are built to function under genuine constraint. They represent governance infrastructure that works in the real world, not only in well-resourced innovation labs. Africa, more broadly, is not a governance laggard. It is a governance laboratory. The continent's cross-border trade complexity, jurisdictional fragmentation, and resource constraints mean that AI governance solutions developed here are, by necessity, more robust and more portable than those developed elsewhere. The Dialogue should treat African practitioner experience not as a capacity gap to be addressed, but as a knowledge asset to be integrated.

The Dialogue plays a vital role that no bilateral or regional process can match: it creates a legitimate, inclusive space for testing the compatibility of governance frameworks from different jurisdictions without forcing them into a single mold. International cooperation on AI governance struggles when it confuses harmonization with homogeneity. A framework that works well in the EU's resource-rich regulatory environment may not be suitable for countries like South Africa, Nigeria, or Indonesia - not due to a lack of capacity, but because their regulatory contexts, economic structures, and deployment environments differ fundamentally. To foster meaningful cooperation, the Dialogue should focus on three key actions: First, it can establish a mutual recognition framework for AI governance instruments that sets minimum accountability standards while allowing for jurisdictional flexibility. Second, it should facilitate two-way knowledge transfer that values governance innovations from the Global South, treating them as valuable contributions rather than solely importing frameworks from the Global North. Third, it can anchor international collaboration in practical, tested solutions instead of relying on theoretical consensus documents. The AI governance frameworks created by UnRisked (Pty) Ltd in South Africa - the Agentic Risk Schedule, Record of Advice, and Shared Cognition Liabilityâ"¢- demonstrate that effective governance structures can be built even in challenging conditions. This represents the kind of solution that international cooperation should strive to scale.

The Dialogue should focus on enhancing frameworks such as the ITU's AI for Good ecosystem, UNESCO's Recommendation on the Ethics of AI, and the African Union's Continental AI Strategy. Its unique value lies in connecting these normative frameworks to practical governance infrastructure. While previous initiatives have established principles, a significant gap remains in terms of architectural frameworks. The Dialogue's key contribution should be to bridge the divide between normative consensus and practical implementation. This can be achieved by creating adaptable reference architectures for Member States rather than producing declarations that require interpretation on their own. Moreover, the Dialogue should engage with the FSCA and similar regulators in the Global South, who are already addressing AI accountability in regulated sectors. It is also essential to involve InsurTech and FinTech practitioners who have developed governance solutions in the absence of global standards. These practitioners offer a valuable yet underutilized knowledge base that current multilateral initiatives have overlooked. The Dialogue's added value comes from its ability to bring legitimacy and reach. It can facilitate the integration of governance innovations that emerge at the edges of the system into globally recognized frameworks with established adoption pathways.

The Dialogue should structurally distinguish between stakeholders who have governance frameworks to contribute and those who have governance needs to articulate, and create a dedicated space for both. Practitioners from regulated sectors in the Global South, solo founders, and domain-specific technical experts are consistently excluded from multilateral formats designed for large organisations with dedicated policy teams. Format recommendations: replace generic open-floor sessions with structured practitioner panels where applied governance experience is presented alongside theoretical frameworks. Require that each thematic session include at least one voice from a developing economy who is actively deploying - not studying - AI systems in regulated contexts. Create a written input track that is genuinely integrated into session design, not processed separately and summarised into invisibility.

The voices that are most often underrepresented in AI governance discussions are those of regulated-sector practitioners from the Global South, including underwriters, compliance officers, and founders of FinTech and InsurTech companies. These individuals are making real-time decisions regarding AI governance that directly affect people's lives, yet they have no established pathway to participate in multilateral processes. Additionally, solo founders and small enterprises working to build governance infrastructure without institutional support are also underrepresented. The assumption that meaningful contributions to AI governance require large organizations excludesthose whose constraint-driven innovation produces the most robust and portable solutions. A formally recognised practitioner constituency within the Dialogue structure, with dedicated input pathways and session representation, would begin to address this gap.

The most effective change in format would involve replacing broad declaratory statements with structured problem-solving sessions. These sessions would present a live governance gap, invite practitioners to propose competing solutions, and engage participants in identifying which elements could form the foundation for international standards. Additionally, implementing a pre-Dialogue open submission review process would be beneficial. Written inputs would be synthesised by thematic working groups rather than solely by secretariat staff, ensuring that practitioner knowledge actively shapes session design instead of being merely catalogued afterward. Finally, the Dialogue should pilot a "governance stress-test" format: present a real-world AI deployment scenario from a Global South-regulated sector and invite multistakeholder teams to identify the governance failures and propose solutions in real time. This format would surface the gaps between existing frameworks and deployment realities more effectively than any panel discussion.

UnRisked (Pty) Ltd has created a suite of AI governance frameworks that provide concrete and deployable solutions to the accountability gaps identified in this submission. The Agentic Risk Schedule (ARS) is a pre-deployment governance tool that outlines the operational boundaries, autonomy limits, escalation conditions, and human oversight requirements for an AI agent before it is authorized to act. It serves as the agentic equivalent of a regulatory license condition. (CIPC Provisional Patent Record ID 1537459, March 2026.) The Record of Advice (RoA) is an AI-native accountability document tailored for AI agents in advisory roles. It captures reasoning chains, data inputs, and accountability attribution, thereby creating the necessary evidentiary infrastructure to establish or enforce professional accountability. (CIPC Provisional Patent, March 2026.) The Shared Cognition Liabilityâ"¢ (SCLâ"¢) framework addresses liability attribution in multi-agent environments where no single actor bears full responsibility for an AI system's outputs. It proposes proportional allocation mechanisms adapted from insurance and tort law principles. (Published: SSRN Abstract ID 6374198.) The Cognitive Autonomy Ratingâ"¢ (CARâ"¢) provides a standardized classification schema for AI agent autonomy levels, allowing for consistent regulatory communication across jurisdictions and facilitating governance interoperability. These frameworks were developed within South Africa's FAIS-regulated financial services environment, under genuine regulatory constraints and without institutional support. They are designed to be applicable across different jurisdictions and are available as reference architectures for consideration by the Dialogue and the Scientific Panel.