Model AI Governance Framework for Agentic AI

Strategic Center Evaluation

Adjacent Publisher Context

The following material is published by IMDA and addresses topics relevant to the Five Forces, but is not explicitly positioned as a component of the Model AI Governance Framework for Agentic AI. It is not included in the /25 coverage score above.

Narrative Strengths & Gaps

Internal Pressures & Capabilities. This is the framework’s centre of gravity, with deep, implementable treatment. Section 2.1 asks organisations to weigh each agentic use case by impact (domain, access to sensitive data and external systems, scope and reversibility of actions) and likelihood (autonomy, task complexity, external provisioning, system complexity), then to bound risks by design: “prefer deterministic rather than non-deterministic limits” and give agents “only the minimum tools and data access needed”. Section 2.2 allocates responsibilities across decision makers, product teams, cybersecurity teams and users, and designs human checkpoints for high-stakes or irreversible actions, then audits that oversight through override rates and response times to catch rubber-stamping. Section 2.3 covers technical controls, end-to-end testing, gradual rollout by user cohort, tools and connected systems, and continuous monitoring. Section 2.4.3 adds a distinctive workforce point: organisations should provide training and work exposure “so that users retain foundational skills”. The framework offers no self-assessment of where an organisation stands across its AI strategy; the MSD case study’s five levels of agency illustrate one organisation’s governance pathways rather than a diagnostic (Map: Not met). Its four iterative areas give each agent deployment a governed route from risk assessment to monitored operation, but not a route for the organisation’s strategy as a whole (Course: Partly met).

Global & Geopolitical Dynamics. The framework does not address this force. It recommends technical interoperability standards (MCP, A2A, OAuth 2.1, OpenTelemetry) and notes the difficulty of agents that “cross system or organisational boundaries”, but cross-border data, sovereignty, national strategy and geopolitical competition fall outside its scope, consistent with its focus on how an individual organisation governs the agents it deploys.

Societal & Regulatory Pressures. Regulation is built into the governance method rather than left as a statement of principles. Changes in compliance requirements are named as one of four triggers for a change review, alongside technical, environmental and performance triggers. Actions that may trigger liability, or downstream obligations such as “entering into a contract or sale”, sit behind human approval checkpoints. Audit trails are kept immutable “for analysis and compliance purposes”, vendor contracts are expected to cover data protection, and user data is to be handled under the organisation’s privacy policies with explicit consent where needed. Case studies show compliance functions in action: PwC’s Technology Risk Management team advises on “any regulatory requirements”, and compliance officers at Ant International review agent specifications. The existing trusted-AI principles of transparency, accountability and fairness are carried over and “translated in practice for agents”. The boundary is specific regulation and societal expectations beyond compliance: the framework names no laws or standards to align with, and does not analyse what the public expects of agents once legal requirements are met.

Industry & Competitive Pressures. The framework treats the agentic ecosystem as a source of risk to manage. It maps a six-actor value chain from model developers to end users and asks organisations to clarify “distribution of obligations in any terms and conditions or contracts” with external parties. It addresses third-party opacity directly, recommending disclosures, scoped credentials and logging from vendors, or alternative or in-house solutions where these are lacking. MSD’s containment of vendor-embedded agents within their own ecosystems shows the approach in practice. Its systemic-risk discussion recognises that “pricing agents used by different organisations could observe each other’s prices and converge on higher prices rather than competing”. Competitive positioning and industry adoption strategy fall outside its scope.

Human Needs & Aspirations. Section 2.4.2 addresses users who interact with agents acting on an organisation’s behalf, typically external-facing customer service or intake agents. The guidance is to “Declare upfront in the user interface that the users are interacting with agents”, explain the agent’s range of actions and data use, and provide human contact points users can alert “if they are dissatisfied with a decision”. It also suggests allowing users to “set their own approval thresholds and boundaries for agent actions based on their individual risk tolerance”. The Workday case study illustrates transparency towards end users of HR and finance agents. Broader public sentiment about displacement and citizens’ aspirations for AI lie outside the framework’s operational scope.

What this framework does not give you. A map of where the organisation stands across its AI strategy; its risk factors assess individual use cases. Without a clear picture of where it stands, an enterprise cannot tell which moves matter most. A route for the organisation as a whole; its four areas govern individual agent deployments. Without a route, even a well-chosen destination stays an aspiration. A way to choose your own destination; it supports confident, accountable adoption of agentic AI. An enterprise that has not stated its own ambition cannot check whether this destination is the right one for it. A view of geopolitical forces such as sovereignty, cross-border data and national AI competition. Sovereignty, cross-border data rules and export controls can change which AI an enterprise can use, and where.

Synthesis. The framework suits chief risk officers, AI governance and platform teams, and technology leaders in organisations that are moving from generative AI pilots to agents that take actions in production, particularly where agents touch customer-facing processes, payments or sensitive data. Its distinctive contributions are a risk-based way to decide how much autonomy each agent should have, through four levels of human involvement, and practical measures for keeping human oversight meaningful as agent volumes grow. Used alongside a strategy framework, it deepens the trust and governance discipline for agentic AI. The Five Forces lens adds what an agent-governance framework deliberately leaves to others: a map and route for the strategy as a whole, geopolitical context, societal expectations beyond compliance, a declared time horizon, and how falling execution costs change the value of ideas.

Sources

Tier 1 (scored), recency-gate pass unless noted:

1. Model AI Governance Framework for Agentic AI, Version 1.5 (PDF) – published 20 May 2026, updated 5 June 2026; 53 pages; IMDA’s current address for the framework; IMDA’s server blocks automated requests, so the link was manually verified by the Sapience team on 31 October 2026 and the text read from the downloaded PDF
2. AI Verify Foundation resources page – current; lists IMDA’s agentic AI governance resources; live when checked on 31 October 2026

Tier 2 (reported in Section 2.75, not scored):

1. Starter Kit for Testing Agentic Systems for Safety and Reliability (AI Verify Foundation resources page) – current; live when checked on 31 October 2026

Tier 3 (third-party corroboration, not scored): none used. Publication dates are taken from the framework PDF itself.

Excluded on recency or quality (not scored):

1. Model AI Governance Framework for Generative AI – 30 May 2024; outside 24-month window; earlier edition in the same series