Skip to content

Latest commit

 

History

History
60 lines (43 loc) · 3.04 KB

File metadata and controls

60 lines (43 loc) · 3.04 KB
title IMDA — Four Dimensions of Agentic AI Governance
source-id imda-agentic-ai-2026
wiki-page (02_References/enterprise-ai/wiki/frameworks/four-dimensions-of-agentic-ai-governance.md)
last-synced 2026-06-07

IMDA — Four Dimensions of Agentic AI Governance

What this reference contains

IMDA's Model AI Governance Framework (Agentic AI extension), 2026. The central governance model for responsible agentic AI deployment: four dimensions applied iteratively across the agent lifecycle.

The four dimensions

# Dimension Core question
1 Assess and bound the risks upfront Is this use case suitable for an agent, and how do we limit its blast radius by design?
2 Make humans meaningfully accountable Who is responsible for what, and how do we ensure human oversight remains effective at scale?
3 Implement technical controls and processes How do we operationalize safety across the lifecycle?
4 Enable end-user responsibility How do we equip the people who use or oversee agents to do so safely?

Source: IMDA Agentic AI Framework 2026, p.13.

Dimension details

Dimension 1 — Assess and bound:

  • Score use case on risk factors (impact × likelihood)
  • Set structural limits on tools, data, autonomy, and area of impact
  • Issue agent identities and authorizations before deployment
  • Avoid single all-powerful agents; restrict device installation in mission-critical contexts

Dimension 2 — Human accountability:

  • Allocate responsibility across the agentic value chain (model devs, platform, deployer, users)
  • Design human-approval checkpoints
  • Audit override rates and response times (automation bias is a measurable risk)

Dimension 3 — Technical controls:

  • Pre-deployment: structural controls (not prompt-layer only); agent-specific testing (task execution, policy adherence, tool calling, robustness)
  • Post-deployment: gradual rollout, continuous monitoring, change management
  • Tighten permissive defaults; dedicated agent credentials; log all actions

Dimension 4 — End-user responsibility:

  • Transparency to users (capabilities, actions, escalation paths, data handling)
  • Training on autonomous-agent risks
  • Protect foundational skills and tradecraft from agent erosion

Iterative principle

"These four dimensions should be viewed as an iterative process. If an anomaly is detected during monitoring, organisations should re-evaluate the earlier dimensions." — IMDA 2026, p.13

When to consult

  • For skill tech-agent-guardrail: this IS the core content — the skill operationalizes all four dimensions as a structured assessment.
  • For skill tech-stack-diagnostic: Dimension 3 (technical controls) maps to the oversight and ops layers of the stack.
  • For skill people-readiness-conversation: Dimension 2 (human accountability) surfaces the governance-ownership gap in leadership teams.

Source

IMDA Model AI Governance Framework (Agentic AI), 2026, pp.13–47. The four-dimension model, iterative principle, and worked examples including the OpenClaw case.