Flux AI

A successful Gobal Dialogue on AI Governance must separate policy documentation from infrastructure enforcement. Agentic systems require runtime controls: pre-execution authorization, irreversibility boundary validation, and independent audit trails outside the system boundary. Currently, governance is treated as static paperwork.

• Interoperability of governance approaches
• Protection and promotion of human rights
• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight

These priorities reflect where governance gaps cause operational failure.

The themes miss the operational reality of agentic AI deployments.

In the EU, the asymmetry between AI adoption speed and governance readiness is creating a surface of paper compliance. 66% of Danish SMEs use AI, but under 10% have integrated it into governed workflows. 35% of employees have entered proprietary data into public AI tools without organisational awareness. The EU AI Act's high-risk enforcement begins August 2026, and most organisations in regulated sectors across Western Europe are preparing compliance documentation rather than building the runtime infrastructure required to enforce it. This gap produces specific operational failures in European deployments. LLM providers push silent model updates that alter system behaviour without notification to deployers. An organisation that validated its AI system against Article 15 in Q1 may be running materially different inference by Q2, with no governance mechanism to detect the shift. In Danish and European public sector organisations beginning to deploy agentic AI across welfare, healthcare, and financial administration, multiple agents now coordinate across sensitive data boundaries. Existing governance frameworks govern individual systems but do not address what happens when two individually compliant agents produce an unauthorised interaction together. Compounding this, deployed AI systems across the sector generate their own compliance logs. In regulated European environments handling citizen data, a system that produces its own audit evidence is structurally inadequate. Independent audit infrastructure is absent from current frameworks. Europe has the regulatory foundation through the AI Act, GDPR, and sector-specific directives. What it lacks is the technical enforcement architecture to operationalise these instruments at runtime, where governance decisions are made in milliseconds, not quarterly review cycles.

The Dialogue can advance international cooperation by addressing the structural cause of duplicated governance engineering across jurisdictions. When I mapped a deployed European public sector system to IMDA Singapore's Model AI Governance Framework in March 2026, the exercise revealed that concepts like "meaningful human oversight" under EU AI Act Article 14 and "human accountability" under IMDA MGF Dimension 2 prescribe functionally similar controls. Organisations operating across both jurisdictions currently build and maintain parallel compliance architectures for what is essentially the same requirement. A vocabulary mapping does not solve this. Standardised technical protocols do. If a tamper-evident audit log from an EU-regulated system cannot be ingested by an IMDA-compliant verification process without reformatting, the systems are not interoperable regardless of how well-aligned the policy language is. The Dialogue should therefore use its political weight to mandate the development of common audit log formats and independent audit infrastructure requirements by technical standards bodies. Concrete outcomes include mandating that standards organisations deliver interoperable governance infrastructure protocols, requiring minimum technical requirements for audit trail independence in agentic AI systems, and establishing how model version integrity should be documented so that a compliance validation performed in one region remains verifiable in another. The Dialogue should also establish a continuous feedback mechanism from practitioners deploying AI in regulated sectors. Current governance standards are shaped primarily by regulators, academics, and large technology companies. Organisations deploying agentic AI in healthcare, public administration, and financial services encounter governance failures that emerge at runtime, not at design time. A structured channel from deployers to standard-setters, maintained beyond the July meeting, would ground future governance instruments in deployment evidence rather than theoretical risk taxonomies.

The Dialogue should connect with governance instruments that have moved beyond principles into operational detail, and with the technical standards bodies that must build the infrastructure. The UK ICO's Tech Futures report on Agentic AI (January 2026) is the most detailed regulatory analysis of autonomous AI agent risks published by a data protection authority. It identifies accountability gaps specific to agentic architectures that current governance frameworks do not address. The EU AI Act represents the most advanced binding legislation, though the Digital Omnibus Regulation has already weakened key provisions including mandatory AI literacy requirements and shifted compliance timelines to 2027-2028. The Dialogue should treat the EU's implementation experience as a live case study in the gap between legislative ambition and enforcement reality. IMDA Singapore's Model AI Governance Framework for Agentic AI provides the most operationally detailed governance architecture for autonomous systems. My submission to IMDA's public consultation demonstrated that its four dimensions can be mapped to EU AI Act articles and deployed against real European public sector systems. The Dialogue should use this cross-jurisdictional mapping experience to inform interoperability requirements. MIT's AI Risk Initiative has mapped over 1,000 governance documents through the AGORA dataset and found that governance coverage concentrates on model safety while emerging concerns like multi-agent risks receive comparatively little attention. Consumer-facing and labour-intensive sectors have lower representation. The Dialogue should build on this empirical evidence of governance gaps rather than constructing its own taxonomy. The Dialogue's added value is in creating political mandates that direct technical standards bodies toward governance infrastructure interoperability. ISO/IEC JTC 1/SC 42 is developing AI management system standards. IETF and W3C have established processes for building interoperable technical protocols. These are the organisations that must deliver common audit log formats, model version integrity documentation, and independent audit trail requirements. The Dialogue should mandate this work, not duplicate it.

The Dialogue's value depends on whether it can incorporate evidence from organisations that deploy AI in regulated environments, not only from those that develop models or write policy. The UK ICO's Tech Futures report on agentic AI identified that organisations may need to rely on supervisor agents to monitor increasing numbers of deployed agents, but acknowledged it remains unclear whether this approach is effective in practice. The Dialogue should require stakeholders to declare where governance frameworks have been tested in production and where they remain theoretical. MIT's AI Risk Initiative found through their AGORA dataset analysis that AI Deployers and AI Developers appear as targets in over 500 governance documents but have minimal involvement in enforcement and monitoring roles. The governance conversation is dominated by governance actors writing rules for others to follow. The Dialogue should actively recruit deployers from regulated sectors, healthcare, public administration, financial services, who encounter governance failures at runtime rather than at policy review. Technical standards bodies should participate not as observers but as recipients of specific mandates. If the Dialogue concludes that independent audit infrastructure is necessary for agentic AI systems, ISO/IEC JTC 1/SC 42 and IETF need to leave with a defined deliverable and timeline. On format: the consultation should require stakeholders to declare their deployment evidence alongside their recommendations. The EU Digital Omnibus demonstrated what happens when governance processes are shaped primarily by cost-reduction arguments from industry without counterweight from practitioners and civil society. Mandatory disclosure of whether a submitter has deployed, audited, or enforced the governance controls they recommend would allow the Dialogue to distinguish between theoretical positions and recommendations grounded in operational experience

Deployers in regulated public sector environments handle the consequences of governance gaps daily but rarely participate in the processes that produce governance standards. MIT's AGORA dataset analysis confirms this empirically: deployers are targeted by over 500 governance documents but have minimal involvement in enforcement and monitoring roles. The organisations implementing AI in municipal welfare systems, hospital administration, and public finance operate under regulatory constraints that technology companies and academic institutions do not face. Their implementation evidence is the most direct test of whether governance frameworks function in practice. Small and mid-sized governance infrastructure companies are absent from consultations dominated by large technology platforms. The companies building runtime enforcement tools, audit infrastructure, and compliance automation for regulated sectors operate at the intersection of policy and engineering. Their perspective differs fundamentally from both the technology providers whose products they govern and the regulators whose requirements they operationalise. The Dialogue should establish participation mechanisms that do not require the resources of a large corporate affairs department to engage with. Workers and citizens affected by AI-driven decisions in public services are the ultimate stakeholders of governance quality. When an AI system in a municipal welfare office processes a benefits decision incorrectly, the affected citizen encounters the governance failure directly. These experiences are not captured by stakeholder consultations framed around technical or policy expertise. Civil society organisations document this harm, but it rarely translates into infrastructure requirements. The Dialogue should establish a mechanism where documented harm from AI-driven decisions systematically triggers a technical audit requirement, so that a citizen's experience becomes a structured input for identifying infrastructure failures rather than remaining an isolated ethical concern. Including these groups requires deployment-evidence requirements that value operational experience alongside academic credentials, and feedback channels that persist beyond single consultation windows.

Replace position statements with structured case submissions. Stakeholders proposing governance requirements should demonstrate where those requirements have been implemented, what failed, and what the measurable outcome was. IMDA Singapore's public consultation used this model: submissions required a case study section alongside policy feedback. The result was recommendations grounded in deployment evidence rather than theoretical preferences. The Dialogue should adopt the same structure for its July meeting. Live interoperability mapping would be more productive than thematic panels. Select a specific governance scenario: an agentic AI system processing citizen welfare data across two jurisdictions, and require participating stakeholders to map the applicable governance requirements from their respective frameworks in real time. This exercise would surface interoperability failures concretely rather than discussing them abstractly. The ICO's Tech Futures methodology, which used four scenario-based futures to explore agentic AI risks, demonstrates that structured scenario work produces sharper analysis than open consultation. Mandated protocol drafting should run parallel to the political dialogue. Assign ISO/IEC JTC 1/SC 42 and IETF representatives a defined deliverable during the July meeting itself: draft a common audit log schema for agentic AI systems within the session timeframe. A working prototype, even incomplete, would demonstrate more political will than a communiqué promising future work. Written submissions should remain open beyond the meeting, with a requirement that submitters disclose their operational relationship to AI governance: developer, deployer, regulator, or affected party. This disclosure transforms the submission archive from a collection of opinions into a structured evidence base weighted by proximity to deployment

The UK ICO's Tech Futures report on agentic AI requires organisations deploying multiple agents to map accountability before processing begins. This represents a functional shift from retrospective compliance to architectural governance.