The Governance Contradiction

"The entire governance model depends on the capacity it has already declared impractical."

In January 2026, Singapore's Infocomm Media Development Authority published the Model AI Governance Framework for Agentic AI. It is the most operationally detailed governance document any government has yet produced for agentic systems. It covers risk assessment, accountability structures, technical controls across the agent lifecycle, and end-user responsibility. It assigns clear roles to developers, deployers, and oversight personnel. It requires human approval at significant checkpoints, chains of accountability across the full value chain, and regular audits to confirm that oversight remains effective. As a governance document, it is thorough.

It is also the first major governance framework to state, plainly and early, that continuous human oversight of AI becomes impractical at scale.

Not in a footnote. Not buried in caveats. In the opening pages of the executive summary. The framework goes further. It states that the people supervising these systems will lose the skills required to do so, and that accountability chains break under their own weight.

But then the framework proceeds to build its entire governance model on exactly that oversight. Checkpoints, accountability chains, audits of effectiveness, all of it requiring the sustained human attention the executive summary has just said cannot be delivered.

I have read the framework several times now, and I cannot tell whether this was deliberate. One reading is that the authors understood the contradiction and included the admission as an honest signal, a way of saying: this is the best we can do, and we know it is not enough. The other reading is that the executive summary and the governance requirements were written by different hands, or at different stages, and nobody noticed that one invalidates the other. The first interpretation is more generous. The second is more likely. Either way, the result is the same: the most detailed governance framework for AI yet published contains, in its opening pages, the evidence that its own model cannot hold.

The framework made that admission about agents. It does not stop there.

A person approving an agent's actions from a contextual summary is doing the same thing as a person signing off on an AI-drafted report, an AI-generated analysis, or an AI-produced recommendation. A risk committee reviewing an AI-generated summary of exposures without reconstructing the underlying data is performing the same nominal oversight. A board member reading an AI-drafted strategy paper and checking whether it reads well, rather than whether the reasoning survives scrutiny, is occupying the same position. The oversight challenge the IMDA framework identified for agentic AI is present wherever AI generates the substance and a human provides the approval. It just described it too narrowly.

The IMDA framework deserves recognition for its candour. Governments do not usually publish documents that contain their own limitations so plainly. But the limitation it identified is not specific to agents. It is the condition of every AI approval process now operating at any meaningful scale. The framework is the first governance document honest enough to say so, even if it said it about the wrong thing.

If the governance model assumes oversight that cannot be delivered, the response is not better oversight. It is a different understanding of what AI does to the organisation it enters.

That is the work I am doing. Naming the structural dynamics that AI introduces to organisations, the patterns that appear when a reasoning technology meets an operating environment built for people. Arguing that AI has to be constrained, not governed, because governance assumes a reviewable output and AI produces volume that exceeds review. Examining what happens when a form of reasoning enters your organisation that does not originate from a mind and does not derive from intelligence. And building the archive of research to show how pervasive LLM-generated content already is in professional settings, and how few senior leaders can recognise its fingerprints or their implications.

AI is not a tool to implement. It is a force to wield and to control. The governance frameworks have not caught up with that, and the IMDA framework, to its credit, is the first to admit it.

References

IMDA, Model AI Governance Framework for Agentic AI, Version 1.0, 22 January 2026. https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf

Foster-Fletcher, R., The Structural Dynamics of AI Adoption. https://fosterfletcher.com/the-structural-dynamics-of-ai-adoption/

Foster-Fletcher, R., You Cannot Negotiate with Code: Why Physics, Not Policies, Will Govern AI, LinkedIn, 2025. https://www.linkedin.com/pulse/you-cannot-negotiate-code-why-physics-policies-govern-richard-d5pxc/

Foster-Fletcher, R., The Undesigned Company: The Artificial Evolution of the Modern Firm, What Still Matters, 2026. https://whatstillmatters.substack.com/p/the-undesigned-company?r=8r6ek

MKAI, Corporate Disclosure Prose Drift Analysis, Inquiry Brief. https://mkai.org/inquirybrief/corporate-disclosure-prose-drift-analysis/