The Singapore stack: voluntary on paper, adopted in practice
The short answer
Voluntary and adopted are not opposites
OCBC published its own internal Source of Wealth agent pipeline inside a government framework it was never legally required to follow, because the published record is worth more to the bank than the privacy of keeping it internal.
What you will be able to do
- Distinguish voluntary best-practice guidance from statute, correctly labeling every instrument in the Singapore Model AI Governance Framework (MGF) family as non-binding while still explaining why an organization remains legally accountable for what its AI systems, including its AI agents, actually do.
- Trace the three-instrument arc (MGF Second Edition, 2020; MGF for Generative AI, 2024; MGF for Agentic AI, 2026) as one evolving story, naming the publisher, date, and defining feature of each, rather than treating them as three unconnected documents.
- Analyze why organizations with zero legal exposure for ignoring a voluntary framework adopt it anyway, using the "defensibility standard" argument: a regulator or a court, after an incident, asks what a careful organization would have done, and a published framework is the documented answer.
- Apply the Model AI Governance Framework for Agentic AI's four dimensions (assess and bound risks upfront; make humans meaningfully accountable; implement technical controls and processes; enable end-user responsibility) to a real AI system, using the OCBC and Dayos case studies published inside the framework itself as worked models.
- Evaluate a published case study (such as OCBC's Source of Wealth memo system or Dayos's IT-ticketing agent) for which specific framework expectations it satisfies and which it leaves open, rather than accepting a company's own framing at face value.
- Recognize a "living document" governance model: a framework that is deliberately incomplete on publication, revised on evidence from real deployments, and explicitly invites the case studies and feedback that will shape its next version.
- Produce a gap map: an assessment of one of your own organization's AI systems against the MGF family, naming which framework speaks to it, which of that framework's expectations it currently meets, and which gaps remain.
- Defend the practical argument for adopting voluntary guidance against the objection "why bother if nothing forces us," using the cost asymmetry between building to a published standard now and reconstructing your reasoning after an incident with no defensible record.
The ideas, one by one
The MGF family is three instruments, not one
The Second Edition (PDPC with IMDA, 2020) governs classical predictive and classification AI; the Generative AI framework (IMDA and the AI Verify Foundation, 2024, nine dimensions) governs content-generating models; the Agentic AI framework (IMDA, Version 1.0 January 2026, Version 1.5 May 2026, four dimensions) governs systems that take actions. Name the right one for your system before you cite anything.
The defensibility standard is why voluntary guidance gets adopted
A regulator or a court asks what a careful organization would have done; a widely adopted, government-convened framework is a pre-written, credible answer to that question.
Real case studies teach the expectations better than the principles do
Dayos's tiered risk-scoring for IT tickets and OCBC's eight-step, human-bounded Source of Wealth pipeline are both published inside the Agentic AI framework itself and are the model this topic's exercise asks you to follow.
Voluntary does not mean legally weightless
Organizations remain fully accountable, under binding law, for what their AI agents actually do; a voluntary governance framework and real legal exposure for an agent's real-world actions are two separate facts that coexist.
A human review step is not automatically meaningful accountability
Alderwood's broker rubber-stamped every declaration; the Version 1.5 update's own new guidance on automation bias and override-rate monitoring names exactly this failure mode.
The family is a living document, deliberately
Version 1.0 to Version 1.5 took four months and incorporated feedback from over sixty companies; always confirm the version and date of any citation before you rely on it.
The three MGF instruments complement NIST and ISO/IEC 42001, they do not replace them
NIST gives an operating rhythm, ISO/IEC 42001 gives a certifiable management system, the MGF family gives technology-stage-specific expectations illustrated with real, dated case studies.
A gap map names what exists, not what is intended
The discipline that makes a gap map defensible is describing current design and observed practice honestly, the way Tyler discovered the broker review was not what it looked like on the org chart.
The gap map is a living artifact, not a one-time exercise
Given how fast the family revises, re-check the map against the current version the way Topic 6.6 taught for a shipping decision, and feed it forward into the Topic 6.10 evidence-reuse lab and the Topic 7.8 agentic-policy grading.
Listen to it
The same lesson, talked through by two hosts as episode 47 of the podcast.
Real cases
Example 1: OCBC and Bank of Singapore bound an agent's autonomy by design, not by policy (2026). Bank of Singapore, OCBC's wholly owned private banking subsidiary, built an eight-step agentic AI pipeline to draft Source of Wealth memos, the documents a relationship manager and compliance team use to establish where a private-banking client's money came from. The pipeline names a human owner at two distinct stages (a human-in-the-loop checkpoint after document extraction, and human final approval before the memo is used) and restricts the agents themselves to task-level autonomy only: extraction, benchmarking, drafting, proofreading, and formatting, never a credit, onboarding, or risk decision (IMDA, "Model AI Governance Framework for Agentic AI," Version 1.5, case study, p. 20). This is Dimension 2 of the framework, meaningful human accountability, made concrete: the design itself, not a written policy layered on afterward, is what keeps the decision with a human.
Example 2: Dayos scores every task before deciding how much autonomy it gets (2026). Dayos, a Singapore-headquartered enterprise automation company, replaced its own IT ticketing system with an agent built on its Hero platform, cutting licensing costs by 121,000 US dollars a year in a forty-five day rollout. Before launch, every category of IT ticket was scored on severity of impact, reversibility, and the feasibility of human review, and the score determined the tier: full automation for low-severity reversible tickets (sixty percent of volume) with biweekly audits, human sign-off required for moderate-severity tickets (thirty percent), and no autonomous execution at all for the highest-severity, least-reversible tickets (ten percent) (IMDA, "Model AI Governance Framework for Agentic AI," Version 1.5, case study, p. 18). This is Dimension 1, assessing and bounding risk upfront, with a real cost figure and a real tiering method attached rather than a general commitment to "be careful."
Example 3: The framework family widens with each generation of AI, tracking the technology rather than getting ahead of it (2020 to 2026). The Model AI Governance Framework, Second Edition (PDPC with IMDA, 21 January 2020) governed predictive and classification AI. The Model AI Governance Framework for Generative AI (IMDA and the AI Verify Foundation, 30 May 2024) added nine dimensions for models that generate novel content. The Model AI Governance Framework for Agentic AI (IMDA, Version 1.0, 22 January 2026; Version 1.5, 20 May 2026) added four dimensions for systems that take actions. Each instrument was published roughly when the technology it governs reached meaningful organizational adoption, not years ahead of it and not years behind. The lesson for a governance professional tracking currency in any jurisdiction: expect the next MGF instrument, whatever technology prompts it, to arrive on a similar rhythm, close behind real deployment rather than in front of it.
Example 4: A four-month revision cycle built entirely from real deployments (2026). Version 1.0 of the Agentic AI framework launched 22 January 2026. Version 1.5 followed 20 May 2026, updated again 5 June 2026, incorporating feedback from more than sixty companies and adding named case studies across every one of the four dimensions, including OCBC, Dayos, MSD, PwC, Tencent, X0PA, Google, GovTech, Workday, and others (IMDA, "Model AI Governance Framework for Agentic AI," Version 1.5, "What's New in This Version"). No statute anywhere revises this fast; the speed is only possible because the framework is voluntary, low-friction to contribute to, and explicitly built as a living document rather than a fixed rule. For a governance professional, the takeaway is a habit: re-check the version and date of any MGF citation before you rely on it in a memo, an audit response, or a board paper, the same discipline Topic 6.6 taught for a shipping decision that can go stale in weeks. (see Topic 6.6)
Example 5: The contributor list is international, not just local (2026). Alongside Singapore-headquartered organizations such as OCBC, Bank of Singapore, and Dayos, the Agentic AI framework's published case studies and its "What's New in This Version" acknowledgments name multinational contributors including AWS, Google, Workday, and Tencent, plus GovTech Singapore and professional-services firm PwC (IMDA, "Model AI Governance Framework for Agentic AI," Version 1.5, "What's New in This Version"; publicly reported case-study coverage, verified 2026-08-15). None of these organizations is legally compelled to feed real deployment evidence back into a Singapore government agency's voluntary document. Their participation is the clearest available evidence that the defensibility-standard argument from Section 3C operates beyond Singapore's own borders: global technology and professional-services firms treat contributing to, and being able to point to, this specific framework as valuable enough to be worth the disclosure, wherever they happen to be headquartered.
Example 6: A high-impact use case with human checkpoints named at the two points that matter, not sprinkled everywhere (2026). X0PA, a business-to-business software company that provides AI-powered recruitment solutions, deployed its AI Agentic Platform across an eight-stage hiring workflow: candidate application, screening and matching (parsing resumes and ranking candidates against job requirements), interview scheduling, interview execution (conducting and transcribing interviews), candidate communication, and analytics, with agents handling every stage except two. At shortlisting and at the final hiring decision, the two points where a wrong call does the most damage to a real person's career and to the hiring organization's fairness obligations, a human checkpoint reviews, approves, modifies, or rejects the agent's recommendation. X0PA enables that review to be genuine rather than nominal through explainable outputs (matching scores and ranking rationale that let a reviewer see the basis of a recommendation, not just its conclusion) and specific, named failure-mode training (recruiters are trained on situations where the agent misinterprets a non-standard career path, or produces lower-confidence evaluations for candidates with limited structured data, so the human knows exactly when to look harder rather than trusting the score by default) (IMDA, "Model AI Governance Framework for Agentic AI," Version 1.5, case study "X0PA: Enabling effective human approvals in high-impact recruitment workflows," p. 30). The lesson for a gap map: a human checkpoint does not need to appear at every stage to be meaningful, X0PA automates six of eight stages, but it does need to sit at the specific points where the decision is genuinely high-stakes, and it needs the reviewer trained on exactly how the agent tends to be wrong, not simply told to "review carefully."
Where people go wrong
- "Voluntary means we can skip it with no consequence." The consequence of ignoring the MGF family is not a fine; it is the absence of a defensible record and a missed competitive signal. Alderwood's near-miss with its enterprise client shows the cost directly: no framework mapping meant no honest answer to a real buyer's question until Tyler built one under deadline pressure.
- "Voluntary means no legal stakes." The framework itself states organizations remain legally accountable for their agents' actions. A voluntary governance framework and binding legal exposure for an agent's real-world conduct (under the PDPA, contract law, or tort law) are two separate facts; the first does not cancel the second.
- "A human review step is the same as meaningful human accountability." Alderwood's broker "reviewed" every declaration and had, in practice, stopped genuinely checking any of them: automation bias, the exact failure the Version 1.5 update added guidance about. A checkpoint that exists on an org chart but not in observed behavior does not satisfy Dimension 2.
- "The MGF family is one document." Treating "the Singapore framework" as a single thing collapses three separate instruments, covering three different technology generations, published four and six years apart, into one blur. A learner who cannot name which of the three applies to a given system has not actually located the relevant guidance.
- "An old citation is close enough." Because the family is a living document, citing Version 1.0 language, or an outdated summary, after Version 1.5 has revised it is a real error, not a rounding error. Always confirm the version and publication date before relying on a citation.
- "A published case study proves the company is fully compliant." OCBC's own case study is an illustration of Dimension 2, not a certification that Bank of Singapore has satisfied all four dimensions perfectly and permanently. Read a case study for what specific expectation it illustrates, not as a blanket seal of approval.
- "Since it's not a statute, a court or regulator would never look at it." A diligence inquiry after an incident routinely asks what a careful organization in the defendant's position would have done, and a widely adopted, government-convened voluntary framework is precisely the kind of evidence used to answer that question, for or against the organization being examined.
- "Bounding an agent's risk is a policy document, not a design choice." Dayos's tiering was implemented in the agent's actual permissions and workflow, not written in a policy binder nobody consulted. A gap map that lists "we have a policy" as evidence of Dimension 1 compliance has not actually checked whether the system's design enforces the policy.
- "If the technology isn't agentic, none of this applies to us." An organization using a classification model, a generative chatbot, and an agentic workflow needs all three MGF instruments mapped, not just the newest one. Skipping the 2020 or 2024 instrument because "we're focused on agents now" leaves the classical and generative systems ungoverned by the family's own standard.
- "A gap map is a one-time exercise." Given how fast the family revises (four months from Version 1.0 to Version 1.5, with a further update three weeks after that), a gap map built once and never revisited will misstate the current expectations within a single business quarter. Treat it the way Topic 6.6 treated a shipping decision: a living document needs a living map, with a re-check trigger tied to the next framework revision. (see Topic 6.6)
- "Only Singapore-headquartered organizations should bother engaging with a Singapore framework." The Agentic AI framework's own contributor list refutes this directly: multinational organizations including AWS, Google, Workday, and Tencent all contributed real case studies to Version 1.5, none of them Singapore-headquartered, none of them legally required to. A framework's relevance is decided by what it governs (a technology and its risks) and where an organization is connected (any AI system touching Singapore users, data, or operations), not by where the framework's publisher happens to sit.
- "A dollar figure like Dayos's $121,000 saving proves the governance was worth it." The saving and the governance are related but distinct facts: Dayos reported the saving as a business outcome of the automation project, and separately reported the risk-tiering as how it bounded that project's risk. Citing the saving as evidence the tiering "worked" conflates a return-on-investment claim with a risk-management claim; the tiering is defensible on its own terms (a documented, upfront, severity-and-reversibility-based method), whether or not a specific dollar figure ever attaches to your own organization's version of it.
- "There is one correct way to allocate Dimension 2 accountability, and every gap map should look like OCBC's." OCBC's pipeline-step model and PwC Singapore's role-based model both satisfy Dimension 2, by different, equally defensible mechanisms suited to different situations (one workflow versus a repeatable internal capability). A gap map that penalizes a system for not looking exactly like OCBC's diagram, when a role-based allocation would fit the system better, has confused one worked example for the only acceptable answer.
- "Automating most of a workflow and leaving only a couple of human checkpoints is under-governed." X0PA automates six of eight hiring stages and is, on the framework's own account, a defensible design precisely because its two remaining checkpoints sit at the highest-stakes points and are equipped to be substantive (explainable outputs, named failure-mode training). The right question is never "how many stages are automated" but "are the checkpoints that remain in the right place and genuinely functional."
Questions people ask
- What is Model AI Governance Framework (MGF)?
- Singapore's family of voluntary, non-binding AI governance guidance, published by IMDA (alone or with the PDPC or the AI Verify Foundation). It comprises three instruments covering three technology generations: the Second Edition (2020, classical AI), the Generative AI framework (2024, nine dimensions), and the Agentic AI framework (2026, four dimensions). More on Model AI Governance Framework (MGF)
- What is Infocomm Media Development Authority (IMDA)?
- The Singapore government statutory board responsible for developing the country's infocomm media sector, and the publisher (alone or jointly) of all three MGF instruments. More on Infocomm Media Development Authority (IMDA)
- What is Personal Data Protection Commission (PDPC)?
- Singapore's data-protection regulator, which enforces the binding Personal Data Protection Act (PDPA) and co-published the Model AI Governance Framework, Second Edition, with IMDA in 2020. More on Personal Data Protection Commission (PDPC)
- What is AI Verify Foundation?
- A Singapore-based non-profit that builds AI governance testing tools and co-published the Model AI Governance Framework for Generative AI with IMDA in 2024. More on AI Verify Foundation
- What is voluntary guidance?
- Non-binding recommendations that carry no statutory penalty for non-adoption, distinct from a statute (such as the PDPA) that a regulator can enforce. All three MGF instruments are voluntary guidance; the topic's honesty rule requires labeling them as such, using "expects" or "recommends" rather than "requires."
Keep going
This lesson builds Singapore and APAC AI governance frameworks, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
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 →