LI Advisory Studio

The first Global Dialogue on AI Governance will only be a success if it resists becoming another well-intentioned forum that produces alignment without consequence. Success is not participation. It is traction. First, the Dialogue must confront fragmentation head-on. Today, AI governance is accelerating, but in parallel not together. What is needed is not another set of high-level principles, but a clear path to interoperability across jurisdictions: how rules translate, how standards connect, and how organizations operating globally can realistically comply without navigating contradictions. Second, it must close the gap between science, policy, and execution. AI is evolving faster than institutional cycles can absorb. If the Dialogue does not produce mechanisms to continuously translate technical insight into policy that is implementable in real environments, it will quickly become obsolete. Third, inclusion must move beyond representation to capability. Many countries are present in the conversation but absent from the infrastructure that shapes it: compute, data, talent. Without addressing this imbalance, governance risks reinforcing dependency rather than enabling sovereignty. Fourth, the outcome should be a focused, time-bound action agenda: a small number of concrete initiatives where global coordination is not only necessary, but feasible. Progress in AI governance will not come from trying to solve everything, but from proving that some things can be solved together. Ultimately, success will be measured by whether this Dialogue shifts the system from discussion to execution. If it does, it becomes a foundation. If it does not, it becomes noise.

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

These four priorities reflect a simple conviction: AI governance will fail if it remains conceptual. It must become operational. Safe, secure and trustworthy AI is the foundation, but trust is not declared, it is engineered. This requires moving beyond principles to verifiable mechanisms: testing, monitoring, and clear accountability across the lifecycle. Without this, "trustworthy AI" remains an aspiration rather than a system property. The social, economic, ethical, cultural, linguistic and technical implications matter because AI is not neutral: it scales bias, reshapes labor, and redefines access to knowledge and opportunity. Governance must therefore anticipate second-order effects, not just immediate risks, and ensure that AI does not deepen structural asymmetries under the guise of innovation. Interoperability of governance approaches is, in my view, one of the most urgent and under-addressed challenges. Organizations operate across borders; governance does not. Without alignment, we are creating a fragmented landscape that is costly to navigate, difficult to enforce, and ultimately ineffective. Interoperability is what will determine whether governance can scale. Finally, transparency, accountability, and human oversight are what make governance real. Not as abstract safeguards, but as embedded practices: who is responsible, who can intervene, and how decisions can be understood and challenged. As AI systems become more autonomous, the clarity of human accountability becomes more critical. Taken together, these priorities are not separate domains but interdependent levers. The goal is not to slow AI down, but to ensure it scales with intent, with control, and with consequence.

The listed themes are all necessary, but they remain largely descriptive of the problem space. What is still missing are the structural conditions that will determine whether AI governance is enforceable, not just aspirational. First, the question of infrastructure sovereignty cuts across all themes but is not explicitly addressed. AI governance cannot be meaningfully exercised without control, or at least access, to compute, data ecosystems, and critical digital infrastructure. Without this, capacity-building risks becoming symbolic, and many actors remain dependent on a small number of providers shaping both technology and standards. Second, we are underestimating the importance of operational governance models. It is one thing to define principles such as transparency or human oversight; it is another to embed them into decision-making processes, accountability structures, and day-to-day operations. The gap between policy intent and organizational execution is where most governance efforts will fail. Third, the Dialogue should more explicitly address economic concentration and value distribution. AI is not only a technological shift, it is rapidly concentrating power, capital, and influence. Questions around open-source, open models, and human rights intersect here, but the underlying issue is how value and control is distributed globally. Finally, there is an emerging need to govern autonomy at scale. As systems evolve from assistive tools to agentic actors, existing governance paradigms designed for static systems will become insufficient. This raises new questions around delegation, liability, and real-time oversight that are not yet fully captured in current frameworks. In short, the next frontier of AI governance is not defining what good looks like, but building the conditions to make it work in practice, at scale, and across asymmetries.

In Europe, and particularly in Switzerland where I am based, we are entering a phase where AI governance is advancing faster on paper than in practice. The gap is no longer regulatory absence, but operational readiness. The most significant challenge is fragmentation under acceleration. With frameworks such as the EU AI Act shaping the landscape, organizations are facing increasing pressure to comply, yet lack the internal structures, tooling, and clarity to do so effectively. Governance is often interpreted as a legal exercise, while in reality it requires deep integration into technology, operations, and decision-making processes. A second challenge is asymmetry of capability. Large organizations are beginning to build internal AI governance functions, while mid-sized companies that are critical to the European economy, struggle to translate requirements into actionable models. This creates a two-speed system where compliance, innovation, and risk exposure diverge. At the same time, there is a growing tension between sovereignty and dependency. Europe is defining standards, but remains reliant on external infrastructure and AI ecosystems. This weakens the ability to fully exercise governance and raises questions about enforceability, resilience, and long-term competitiveness. However, these gaps also create a unique opportunity. Europe has the potential to lead not by regulation alone, but by operationalizing trustworthy AI at scale, turning governance into a competitive advantage. This means building interoperable frameworks, investing in execution capabilities, and embedding accountability into real systems, not just policies. For Switzerland, with its strong position in innovation, neutrality, and multilateral engagement, there is a distinct opportunity to act as a bridge between regulatory ambition and practical implementation, helping translate global principles into workable models across industries. Ultimately, the question is not whether governance exists but whether it can function under real-world conditions.

The AI Dialogue can play a critical role but only if it moves from being a convening platform to becoming a coordination mechanism. International cooperation on AI governance does not fail because of lack of intent; it fails because of misalignment in execution. The Dialogue has the opportunity to address this by focusing on three shifts. First, from principles to interoperability. Many jurisdictions are converging on similar values: safety, transparency, accountability, but diverging in how these are implemented. The Dialogue can act as a neutral space to align not what countries believe, but how those beliefs translate into compatible standards, processes, and controls. Second, from representation to capability-sharing. True cooperation requires more than inclusion; it requires enabling actors, particularly in emerging economies, to participate meaningfully in shaping, deploying, and governing AI. This means facilitating access to knowledge, infrastructure, and implementation models, not just seats at the table. Third, from discussion to execution. The Dialogue should prioritize a limited number of high-impact areas where coordination is both necessary and achievable, and drive them forward through time-bound, multi-stakeholder initiatives. Progress in AI governance will be built through demonstrable outcomes, not comprehensive frameworks. Finally, the Dialogue can play a unique role in building trust across systems: between public and private sectors, between regions, and between technical and policy communities. But trust will not emerge from alignment alone: it will come from transparency in how decisions are made, and accountability in how they are implemented. If designed with this level of intent, the AI Dialogue can become more than a forum: it can become the infrastructure through which global AI governance actually works.

The AI governance landscape is currently lacking coherence. We get the foundations from the OECD AI Principles to the G7 Hiroshima Process, the Global Partnership on AI (GPAI), and emerging regulatory frameworks such as the EU AI Act. The challenge is that they operate in parallel, with limited interoperability and uneven translation into practice. First, the Dialogue can reduce this friction by driving convergence at the level of implementation, not just intent, by acting as a coordination layer across those foundations: aligning how existing principles, standards, and regulatory approaches map to each other. Today, organizations are forced to navigate overlapping frameworks with different interpretations of similar concepts. Second, it can serve as a bridge between policy and execution. Many initiatives are strong on normative guidance but weaker on operational models. The Dialogue can add value by curating and scaling practical approaches: reference architectures, governance models, and implementation patterns that organizations can adopt across sectors. Third, it can strengthen inclusion through capability, not only participation. By connecting existing capacity-building efforts with concrete governance use cases, the Dialogue can help ensure that a broader set of countries and organizations are not only represented, but able to act. The added value of the AI Dialogue lies in its ability to create continuity. Most initiatives are time-bound or regionally anchored. The Dialogue can provide a persistent, neutral platform that tracks progress, maintains alignment, and evolves governance alongside the technology. In essence, its impact will not come from adding another layer but from making the existing layers work together, at scale, and in practice.

The effectiveness of the AI Dialogue will depend less on who is present, and more on how participation translates into contribution. Different stakeholders should not simply be represented, they should be designed into the system with clear roles and outputs. Governments should focus on alignment and enforceability bringing clarity on regulatory intent, identifying areas for convergence, and committing to pilot interoperable approaches rather than debating abstract principles. Industry must contribute beyond advocacy, by sharing implementation realities: what works, what fails, and what creates friction at scale. This includes opening up governance models, risk management practices, and technical approaches that can inform more practical standards. The scientific and technical community should act as a continuous translation layer ensuring that fast-evolving technical developments are understood, stress-tested, and integrated into governance in near real time. Civil society's role is critical in grounding the Dialogue in societal impact and accountability, ensuring that governance reflects not only institutional perspectives, but lived realities. To enable this the Dialogue should move away from plenary-heavy formats toward a modular, execution-oriented structure: - Focused working tracks tied to specific outcomes (e.g., interoperability, safety validation, oversight models) - Time-bound deliverables with clear ownership across stakeholder groups - Pilot initiatives to test governance approaches in real-world settings before scaling - A continuous feedback loop between technical, policy, and operational communities Finally, the Dialogue should operate as an ongoing mechanism, not a periodic event with progress tracked, shared, and iterated over time. If structured this way, stakeholders move from participants to co-builders of governance that actually works.

The gap in AI governance is not only who is missing from the table but whose reality is not shaping the agenda. First, mid-sized enterprises and operational leaders are largely absent. Much of the discussion is dominated by large technology players and policymakers, while those responsible for actually implementing AI in real business environments are underrepresented. As a result, governance often fails at the point of execution. Including them requires moving beyond representation to structured input on real constraints, trade-offs, and failure modes. Second, emerging economies are present but not empowered. Participation does not equal influence. Without access to infrastructure, data, and technical capabilities, many actors remain rule-takers rather than rule-shapers. Inclusion here means coupling dialogue with capability-building and co-development, not just consultation. Third, domain experts outside of technology from healthcare, manufacturing, education, and public services, are under-leveraged. AI governance is often treated as a technical or regulatory issue, while its impact is deeply cross-functional and divisional. These voices are critical to understanding real-world consequences and should be embedded directly into thematic working tracks. Fourth, there is limited representation of those working on the frontlines of societal impact, including labor organizations, local communities, and practitioners dealing with the downstream effects of AI systems. Their insights are essential to move from abstract ethics to lived accountability. Finally, a critical missing perspective is that of implementation practitioners, like myself, those designing governance models inside organizations. Without them, we risk building frameworks that are theoretically sound but practically unworkable. Inclusion, therefore, is not about expanding the audience, but about redistributing influence, ensuring that those who build, deploy, and are affected by AI actively shape how it is governed.

Most global dialogues over-index on discussion and under-deliver on progress. If the goal is meaningful engagement, the format must be designed for output, not airtime. First, shift from panels to working sprints. Small, cross-functional groups: policy, technical, industry, tasked with solving a specific governance problem within a defined timeframe, producing concrete outputs (e.g., interoperability mappings, oversight models, risk frameworks). Engagement becomes contribution, not commentary. Second, introduce "live case" sessions. Organizations bring real AI use cases with unresolved governance challenges: compliance gaps, accountability questions, cross-border friction, and participants work through them in structured formats. This grounds the Dialogue in reality and surfaces where existing frameworks break. Third, create policy–technology translation labs. These are spaces where technical experts and policymakers jointly stress-test emerging developments (e.g., agentic systems, foundation models) and translate them into actionable governance implications in near real time. Fourth, implement pilot showcases with accountability loops. Instead of showcasing success stories, participants present ongoing initiatives with defined goals, metrics, and constraints and commit to reporting back on progress. This introduces continuity and shared accountability. Fifth, leverage rotating leadership models within working tracks, ensuring that different regions and stakeholder groups not only participate, but actively shape agendas and outputs. Finally, complement physical convenings with a persistent digital layer: a platform where work continues between sessions, outputs are tracked, and collaboration is sustained over time. The objective is simple: shift from engagement as exchange to engagement as co-creation under constraints. That is where real progress and real alignment emerges.

Effective AI governance is already emerging but not as a single model. It is taking shape through a combination of regulatory frameworks, technical standards, and operational practices that begin to close the gap between principle and execution. At the regulatory level, the EU AI Act is a critical step in translating risk-based governance into enforceable requirements. Its strength lies not only in defining obligations, but in forcing organizations to operationalize concepts such as risk classification, documentation, and oversight. Its challenge, however, remains implementation at scale. From a standards perspective, frameworks such as the NIST AI Risk Management Framework and ISO/IEC 42001 provide practical structures for embedding governance into organizational processes. They move the conversation from "what good looks like" to "how to build it," particularly in areas such as risk identification, controls, and lifecycle management. At the organizational level, leading practices are emerging around AI governance operating models, dedicated cross-functional structures that integrate legal, technical, and business accountability. These include model registries, audit trails, human-in-the-loop controls, and continuous monitoring systems. What distinguishes effective approaches is not sophistication, but integration into real decision-making processes. There is also growing momentum around open evaluation and benchmarking ecosystems, where models are stress-tested against shared safety and performance criteria. These initiatives increase transparency and create a baseline for trust, even across competing actors. Finally, regulatory sandboxes and pilot environments are proving essential to bridge innovation and compliance allowing governance approaches to be tested, iterated, and scaled before formal enforcement. The common thread across all these examples is clear: governance becomes effective when it is designed into systems, measured in practice, and continuously adapted not simply declared.