Skip to main content

Map your agent policy to the world's first agentic framework

The short answer

Grade the policy you have; do not rebuild it from zero

Your Topic 7.7 policy is a strong internal artifact. This topic's job is comparing it to an external standard, evidence by evidence, and fixing only the real gaps that comparison finds.

What you will be able to do

  • Explain what the Model AI Governance Framework (MGF) for Agentic AI is, who published it, when, and what kind of instrument it is (voluntary guidance, not statute), stating the doctrine that organisations remain legally accountable for their agents' actions regardless of adoption.
  • Map each of the four Agentic MGF dimensions, assess and bound risks upfront, meaningful human accountability, technical controls and processes, and end-user responsibility, to the specific line or lines of your own one-page agent governance policy (see Topic 7.7) that address it.
  • Grade each dimension as fully addressed, partially addressed, or not addressed, using evidence from your own policy text rather than a general impression of how thorough the policy feels.
  • Diagnose the specific gap most one-page policies share against this framework (no line addresses the end-user), and explain why a policy that governs only the organization's side of an agent interaction is structurally incomplete against a framework built for agentic AI specifically.
  • Judge what an end-user responsibility rule adds that an internal control cannot supply on its own, disclosure, a way to recognize failure, and a path to a human, and write one enforceable rule for each.
  • Defend the claim that a voluntary framework is still worth full compliance effort, distinguishing "voluntary" (who wrote the rule) from "consequence-free" (whether you can be held to the outcome).
  • Compare the framework's own vocabulary ("meaningfully accountable") against your policy's vocabulary (the owner, trigger, consequence test), and show that they describe the same underlying requirement in different words.
  • Produce a dimension-by-dimension gap map of your own policy: for each of the four dimensions, the evidence that it is met, partially met, or missing, and the one fix you would ship first if you could ship only one.
  • Anticipate how this mapped policy strengthens your position in Module 8's investment defense (see Topic 8.3, 8.6) and Module 13's board inspection (see Topic 13.2), where "governed against which standard" is a harder question to answer than "governed."
  • Recognize that a full match against this one general framework is necessary but not sufficient for a regulated sector, and name the additional layer a financial institution, healthcare provider, or other regulated organization must still check.

The ideas, one by one

The Agentic MGF is the world's first published agentic-specific governance framework

IMDA launched Version 1.0 on 22 January 2026 at the World Economic Forum, and Version 1.5 on 20 May 2026, updated 5 June 2026. Four dimensions: assess and bound risks upfront, meaningful human accountability, technical controls and processes, end-user responsibility.

Voluntary describes authorship, not consequence

The framework is not statute, but organisations remain legally accountable for their agents' actions regardless of adoption. Skipping a voluntary framework does not skip the underlying liability it describes.

Your seven-part policy speaks fluently to the first three dimensions and is usually silent on the fourth

Risk classification, named ownership, and technical controls map cleanly onto your existing gate, owner line, and control requirements. End-user responsibility, disclosure, recognizability of failure, a path to a human, is a different audience your original policy was never built to address.

Grade with evidence, not impression

Every dimension gets a grade of Full, Partial, or Not Addressed, backed by a quoted line or an honest statement that no line exists. "We probably cover this" is not a grade.

A line existing is not the same as a line meeting the requirement

Check whether the line carries an owner, a trigger, and a consequence, and whether its scope actually matches what the dimension asks, before marking it Full.

"Meaningfully accountable" and "owner, trigger, consequence" describe the same requirement

Translating between a framework's vocabulary and your own enforceable form, and showing the translation explicitly, is itself evidence of understanding.

The end-user responsibility fix has three parts

Disclosure that the user is talking to an agent and what it may do; a plain-language way to recognize a failure; a real, reachable path to a human. Each becomes its own owner-trigger-consequence line.

No certification exists for this framework, and none should be claimed

The honest claim is "graded against the Agentic MGF's four dimensions, with evidence," never "certified" or "compliant" in a sense the framework does not offer.

Rank your fixes and commit to a date

A gap map with four ungraded gaps and no priority is an aspiration. A gap map with one committed, dated fix is a plan a board can hold you to.

The gap map is a snapshot against a version, and it decays like any other document

A new framework version, a new agent capability, or a real incident is a trigger to re-grade, the same living-document discipline from Topic 7.7.

Three independent lenses converging is stronger evidence than any one alone

Your own policy, this national framework, and a management-system standard like ISO/IEC 42001 (see Topic 6.3) agreeing on the same requirements is a more defensible position than any single source, and shows your original judgment was not a private opinion.

A general framework is necessary, not sufficient, for a regulated sector

A full gap map against the Agentic MGF answers one question well; a financial institution, healthcare provider, or other regulated organization still has a sector-specific layer to check on top of it (see Topic 6.8).

A policy's language and each agent's actual practice are two different questions

Grading the words your policy uses against a dimension is not the same as confirming every agent the policy covers actually follows those words; the second question is answered by your audit and your logs, not by the gap map alone.

Listen to it

The same lesson, talked through by two hosts as episode 59 of the podcast.

Real cases

The Model AI Governance Framework for Agentic AI, IMDA, 2026 (the anchor). Version 1.0 launched 22 January 2026 at the World Economic Forum in Davos; Version 1.5 published 20 May 2026, updated 5 June 2026 to add case studies and best practices. Four dimensions: assess and bound risks upfront; make humans meaningfully accountable; implement technical controls and processes; enable end-user responsibility. Voluntary guidance from a national authority, not statute, and organisations remain legally accountable for their agents' actions regardless of adoption. This is the world's first published governance framework built specifically for agentic AI, and the reference point this entire topic uses.

The lineage the Agentic MGF completes: the Model AI Governance Framework family, IMDA and PDPC, 2020 to 2026 (pointer, owned by Topic 6.7). The Model AI Governance Framework, Second Edition (PDPC with IMDA, 21 January 2020) set general AI governance principles; the Model AI Governance Framework for Generative AI (IMDA and the AI Verify Foundation, 30 May 2024, nine dimensions) extended the family to generative systems. The Agentic AI edition is the third member of this arc (see Topic 6.7), and this topic uses only its four dimensions, not the earlier editions' content.

The maturity gap this topic exists to close: Deloitte AI Institute, Singapore, 2026 (pointer, statistical evidence). Deloitte's "The State of AI in the Enterprise: The Untapped Edge," published 3 February 2026, surveyed 3,235 business leaders across 24 countries including 75 in Singapore (fieldwork August to September 2025): 72% of Singapore businesses plan to deploy agentic AI in several operational areas within two years (15% do so today); only 14% of Singapore leaders report a mature model for agentic AI governance, against a global average of 21%. The gap between the two numbers, adoption racing ahead of governance maturity, is the exact problem this topic's grading exercise is built to narrow inside one organization.

The named-owner gap, measured: AWS-commissioned research by Strand Partners, Singapore, 2026 (pointer, statistical evidence). Published 6 May 2026, this study of 1,500 Singapore businesses across financial services, healthcare, and manufacturing found that just under 30% of AI-adopting SMEs have a clearly defined person responsible for overseeing AI accuracy, just under 40% have no formal process for escalating AI outputs employees are unsure about, and six in ten would face significant or moderate disruption if the primary person responsible for AI left (about one in ten say AI initiatives would likely stop altogether). This is Dimension 2 (meaningful human accountability) measured empirically, at scale, in the same region and the same year this topic teaches the dimension.

A real financial-sector build-out of Dimension 3, the technical-controls side: MAS MindForge AI Risk Management Toolkit, 2026 (pointer). Singapore's Monetary Authority (MAS) published the MindForge AI Risk Management Toolkit on 20 March 2026, developed with a consortium of 24 banks, insurers, and capital markets firms, including an operationalisation handbook and case-study supplement (see Topic 6.8). It shows what "implement technical controls and processes" looks like when an entire sector builds it together rather than one organization alone.

An industry body's own guardrail document: Handbook on Generative AI Guardrails in Banking, Association of Banks in Singapore, 2026 (pointer). Published 24 March 2026 by the Association of Banks in Singapore's Standing Committee on Data Management, drawn from more than 30 member use cases. Another real, dated, sector-specific instance of organizations converging on the same governance requirements this topic's framework states generally.

The co-publisher whose name recurs across this family: the AI Verify Foundation (pointer). IMDA co-published the Model AI Governance Framework for Generative AI with the AI Verify Foundation on 30 May 2024, the immediate predecessor to the Agentic AI edition this topic teaches (see Topic 6.7). Seeing the same co-publisher's name across successive editions of the MGF family is itself evidence of the kind of institutional continuity a defensibility argument can point to: this is not a one-off document from an unfamiliar source, it is the latest release from an authority and a technical partner with a multi-year, publicly documented track record in exactly this area.

The incident that made the accountability gap visible: Google Gemini "please die," 2024 (pointer, owned by Topic 7.7). The anchor for your Topic 7.7 policy remains relevant here: Google's account of the incident named a violated policy and safety filters but no accountable owner, gate, or consequence (see Topic 7.7). Grade that same incident, hypothetically, against Dimension 4: even a perfect internal policy with a named owner would not, by itself, have told the student or his sister anything about the AI's limits, given them a way to recognize the failure as a system limitation rather than a targeted threat, or offered a path to a human at the moment it mattered. The Agentic MGF's fourth dimension names precisely the piece a strong internal policy alone does not cover.

The framework's own lineage of maturity, one year earlier: the Model AI Governance Framework for Generative AI, 2024 (pointer, owned by Topic 6.7). Published by IMDA and the AI Verify Foundation on 30 May 2024 with nine policy dimensions covering generative systems broadly, this earlier edition shows the same authority iterating toward the agent-specific instrument this topic teaches; comparing the generative edition's nine broader dimensions to the agentic edition's four sharper ones is itself a lesson in how a framework narrows and sharpens as a technology category matures from "produces content" to "takes action."

A real bright line already drawn in this region: Singapore's deepfake election ban, 2024 (pointer, owned elsewhere, do not centerpiece). Singapore's Elections (Integrity of Online Advertising) Act, passed in 2024, prohibits deepfake content of election candidates during campaign periods (see Topic 6.9). It shows a case where Singapore chose statute, not voluntary guidance, for a specific, narrow, high-stakes harm, a useful contrast to the voluntary status of the Agentic MGF discussed at length in this topic: the same jurisdiction reaches for both tools, guidance for the broad and evolving case, law for the narrow and settled one.

Notice the spread on purpose: a national voluntary framework (the Agentic MGF), its own lineage (the MGF family), two independent statistical studies measuring the exact gap this topic addresses, a sector-wide financial toolkit, an industry body's own handbook, a pointer back to the incident that motivated Module 7's whole policy artifact, and a contrast with an area where the same government chose binding law instead. Applied AI governance in this region is not one document; it is a fast-moving stack of national frameworks, sector toolkits, industry handbooks, and a small set of hard statutes, and the skill this topic teaches, grading your own rule against a published external standard, is the skill that lets you navigate all of them, not just this one framework.

Where people go wrong

Mistake 1: Treating "voluntary" as "ignorable." The Agentic MGF is not a statute, and some readers stop there. The framework's own text states organisations remain legally accountable for their agents' actions regardless of adoption; voluntary describes authorship, not consequence. Treating voluntary guidance as safe to skip misreads exactly what makes it worth following.

Mistake 2: Rewriting the whole policy instead of grading it. Facing a new framework, the instinct is to start over. This wastes the seven parts you already built and tested (see Topic 7.7). The correct move is grading: map what exists, find the real gaps, fix only those.

Mistake 3: Marking a dimension "fully addressed" because a line exists. A line that gestures at a requirement without naming the owner, trigger, and consequence the requirement actually needs is partial credit, not full credit, the same three-part test from the last topic applies to grading against this one.

Mistake 4: Skipping the evidence column. "I think we cover accountability" is not a grade; it is a guess. Every grade in a gap map must cite the specific line, or state plainly that no line exists.

Mistake 5: Assuming an internally strong policy automatically covers end-user responsibility. The most common single gap this topic surfaces. A policy built entirely to govern the organization's own behavior can be excellent by every internal measure and still say nothing about the person on the other side of the interaction, because that was never what it was built to say.

Mistake 6: Confusing v1.0 and v1.5 content. Version 1.5 (20 May 2026, updated 5 June 2026) added case studies and best practices; the four core dimensions did not change between versions. Citing "v1.5's new fifth dimension" or similar would be a fabrication; grade against the four dimensions that both versions share.

Mistake 7: Claiming certification or endorsement. The Agentic MGF has no certification body and issues no compliance seal. A gap map that says "Tidepoint is Agentic MGF certified" claims something that does not exist. The honest claim is "graded against the framework's four dimensions, with evidence," never "certified."

Mistake 8: Treating the framework as a checklist to tick rather than a bar to meet. Writing a thin, aspirational line for each of the four dimensions just to have something in every cell of the table repeats the exact aspiration-versus-rule failure this program has already taught (see Topic 7.7). A weak line that technically mentions end-user disclosure without an owner, trigger, or consequence does not earn "fully addressed."

Mistake 9: Forgetting the sector layer. A financial institution's agents face MAS guidance on top of the general Agentic MGF (see Topic 6.8); grading against this one framework is necessary, not sufficient, for a regulated sector.

Mistake 10: Never re-grading. A gap map produced once and filed decays exactly as an unrevised policy decays (see Topic 7.7). A new framework version, a new agent capability, or a real incident should trigger a re-grade, not just a policy revision.

Mistake 11: Using the framework's language instead of your own enforceable form. Writing "we enable end-user responsibility" as a policy line, copying the framework's own dimension title, is the aspiration trap again: it names no owner, trigger, or consequence. Translate the framework's requirement into your program's enforceable form every time.

Mistake 12: Believing a full gap map means the agents are safe. The gap map shows your policy's words match the framework's words. It does not verify the policy is operated, logged, and enforced (see Topic 10.2). Grading the page is not the same as proving the page held.

Mistake 13: Grading a dimension "Full" because a related, but narrower, line exists. A prohibition against one specific harmful output is not the same as a full accountability line, even though both relate loosely to governance. Check that the line's actual scope matches what the dimension asks, not merely that the line is in the same general neighborhood.

Mistake 14: Assuming the end-user fix belongs to the same owner as the internal controls. The agent's technical owner may have no authority over customer-facing product language or escalation design. Name the owner who actually controls what the fix requires, or the new lines will pass the three-part test on paper and still never ship.

Mistake 15: Grading against a general impression of "how the industry usually handles this" instead of the framework's actual published dimensions. Substituting a vague sense of best practice for the four specific, published dimension names produces a gap map that cannot be checked against anything concrete. Quote the dimension's own wording, not a paraphrase of what you assume it probably means.

Mistake 16: Treating a strong policy grade as proof every agent it covers is actually compliant. A gap map grades the policy's language; whether each agent the policy governs actually follows that language in practice is a separate question for enforcement and logs (Section 3M; (see Topic 10.2)). A "Fully addressed" grade on paper and an uninspected agent quietly out of compliance can coexist, and only an audit, not the gap map, catches the second.

Questions people ask

What is Model AI Governance Framework (MGF) for Agentic AI?
Singapore's voluntary governance framework for AI agents, published by the Infocomm Media Development Authority (IMDA); Version 1.0 launched 22 January 2026 at the World Economic Forum, Version 1.5 published 20 May 2026 (updated 5 June 2026). Four dimensions: assess and bound risks upfront, meaningful human accountability, technical controls and processes, end-user responsibility. The world's first published agentic-specific governance framework, and the reference point this topic's gap map uses.
What is IMDA (Infocomm Media Development Authority)?
The Singapore statutory board that publishes the Model AI Governance Framework family, including the Agentic AI edition this topic teaches. More on IMDA (Infocomm Media Development Authority)
What is gap map?
The artifact this topic produces: a dimension-by-dimension comparison of an existing policy against an external framework's requirements, with a grade (Full, Partial, Not Addressed) and quoted evidence for each dimension.
What is fully addressed?
A gap-map grade meaning a specific policy line states the requirement as an enforceable rule (owner, trigger or gate, consequence) with scope that genuinely matches the dimension's requirement.
What is partially addressed?
A gap-map grade meaning a policy line exists and gestures at the requirement but is missing one of the three enforceable parts, or covers only part of what the dimension requires.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →