Skip to main content

MAS runs the hardest AI regime in APAC, and it has never been a law

The short answer

A guidelines-based regime is not an optional regime

MAS built eight years of AI governance instruments, FEAT, Veritas, the AIRG Consultation Paper, without ever passing an AI statute, and every one of them behaves like a real obligation inside institutions MAS already licenses and examines under other binding law.

What you will be able to do

  • Trace the eight-year MAS arc from the FEAT Principles (2018) through the Veritas consortium's three phases (2021 to 2023) to the AIRG Consultation Paper (2025) and the 2026 industry toolkits, and state correctly, for each instrument, whether it is voluntary guidance, an industry deliverable, or a still-unissued proposal.
  • State accurately the current legal status of the MAS AI Risk Management Guidelines as of 15 August 2026: a Consultation Paper whose public comment period closed 31 January 2026, with final guidelines not yet issued, and a proposed 12-month transition period once they are.
  • Apply the AIRG's four sections, oversight, key AI risk management systems, key AI life cycle controls, and capabilities and capacity, to a real AI system, distinguishing what belongs in each section rather than treating the Guidelines as one undifferentiated checklist.
  • Analyze why proportionality in the AIRG governs the degree of implementation, not whether an institution is in scope at all, and correctly reject the argument that a small or narrow-use institution is exempt.
  • Defend the doctrine that an institution cannot delegate its AI governance responsibility to a third-party vendor, even where the AI itself, a foundation model, a scoring engine, a chatbot platform, is built and operated entirely by someone else.
  • Produce a risk materiality assessment for one AI system that scores it across at least the three dimensions the AIRG names, impact, complexity, and reliance, and reach a proportionate, defensible tier.
  • Produce an AI inventory entry, in the AIRG's shape, that captures a system's purpose, its data, its third-party components, and its materiality tier, ready to sit inside a Board-level inventory a MAS examiner could review.
  • Distinguish the MAS AIRG's finance-specific, guidelines-based supervisory regime from Singapore's cross-sector, general-purpose voluntary frameworks, the Model AI Governance Framework family, so you can correctly route a governance question to the instrument that actually governs it. (see Topic 6.7)
  • Recognize why a guidelines-based regime with no statute behind it can be, for a regulated financial institution, functionally harder to satisfy than a statute with a lower supervisory bar, because MAS pairs the Guidelines with continuous supervisory engagement, examination, and its existing statutory powers over the institutions it already licenses.

The ideas, one by one

The AIRG is a proposal, not yet a rule, and you must say so exactly

Consultation Paper P017-2025 (MAS, 13 November 2025), comment period closed 31 January 2026, final guidelines not issued as of 15 August 2026, a proposed 12-month transition once issued. Label every claim about it accordingly.

The arc is sequential, not redundant

FEAT (2018) named the values with no method. Veritas (2021 to 2023) built and field-tested the method across three phases and seven-bank pilots. The AIRG (2025, proposed) gathers the arc into four structured sections. Each beat exists because the last one was deliberately incomplete on its own.

Proportionality is a dial, not a switch

Every MAS-supervised financial institution is in scope; proportionality scales the intensity of the program to the institution's own risk profile, it does not exempt anyone from having a program at all.

You cannot delegate AI governance to a vendor

A licensed, embedded, or third-party-operated AI system still belongs in your inventory, under your accountability, with the vendor's own assurances treated as an input to your assessment, never a substitute for it.

Materiality is impact, complexity, and reliance, scored separately, tiered by the sharpest one

Averaging the three dimensions softens exactly the score most likely to reveal real risk, and reliance, how much a human genuinely defers to the system rather than merely appearing to review it, is the dimension most often scored too generously.

A guidelines-based regime with real supervisory teeth can be harder than a weaker statute

MAS's existing licensing power, its fit-and-proper regime reaching Board and Senior Management personally, and an industry that has already normed itself (MindForge's 24 institutions, the ABS's 30-plus-use-case Handbook) together close almost all of the practical gap between "proposed" and "binding."

The inventory and the materiality assessment are the concrete deliverable

Section 2's identification, inventory, and materiality assessment requirements are not abstractions; they are the two artifacts this topic's lab produces, ready to sit inside a Board-level record and to be re-expressed across other regimes.

Route the right question to the right instrument

The AIRG is finance-sector-specific and guidelines-based, issued by a supervisor with statutory teeth over the institutions it covers; the Model AI Governance Framework family from Topic 6.7 is cross-sector and voluntary with no supervisory relationship behind it. Confusing the two misroutes a governance decision. (see Topic 6.7)

This is a living record, not a one-time filing

A materiality tier is only as current as the last honest re-check; a silent vendor update, a data-source change, or a shift in how heavily staff defer to a system's output can move the real risk without anyone noticing unless change management catches it.

A document is not a defense; a document that survives being tested is

A MAS-style examiner probes whether a named owner can explain a real decision, whether a reliance score is backed by actual review evidence, and whether a director can state risk appetite unprompted. Build the artifact to survive that test, not to look complete on a first read.

Listen to it

The same lesson, talked through by two hosts as episode 48 of the podcast.

Real cases

These examples trace the MAS arc through its own verified instruments and one verified sector picture of the gap the AIRG is built to close, each with its date and source named.

Example 1: FEAT sets the vocabulary before there is a method (MAS, 2018). In November 2018, MAS published four principles with no assessment method attached to them. That gap was deliberate; MAS chose to name the destination (fairness, ethics, accountability, transparency) before prescribing the route. The lesson for a governance professional: a regulator naming its values years before it operationalizes them is not stalling, it is giving the industry a multi-year runway to build the method collaboratively, which is exactly what happened next.

Example 2: Veritas proves the method works inside real banks, not just on paper (MAS, 2021 to 2023). Phase 3 of Veritas, concluding 26 June 2023, is the example worth sitting with longest, because it did not stop at publishing a Toolkit v2; it paired that publication with integration pilots run by seven financial institutions. A methodology that only exists as a document is a hypothesis. A methodology seven banks have actually run against their own production models is evidence, and it is exactly the kind of evidence a Consultation Paper can lean on two years later to justify moving from voluntary guidance toward guidelines that look and feel far closer to a rule.

Example 3: the AIRG names third-party AI risk as its own line item (MAS, 2025). The Consultation Paper's explicit inclusion of "AI inventory including third-party AI" as part of Section 2, rather than leaving vendor-supplied AI as an implicit afterthought, is itself a real, citable regulatory choice, and it directly answers a question every large financial institution has been quietly avoiding: does licensing a model from a fintech partner or a foundation-model provider transfer the governance burden along with the license. MAS's answer, embedded in the Guidelines' own structure, is no.

Example 4: MindForge shows the supervisor and the industry building the same toolkit in parallel (MAS, 2026). The MindForge consortium, 24 banks, insurers, and capital markets firms working with MAS to build an operationalization handbook, is a real, verified instance of a regulator co-producing its own compliance tooling with the institutions it supervises, months before the underlying Guidelines are final. It is worth naming as its own example because it shows MAS treating "final guidelines" and "practical readiness" as two separate clocks, and starting the second clock early.

Example 5: the industry association moves independently, and faster, than the regulator (ABS, 2026). The Association of Banks in Singapore's Handbook on Generative AI Guardrails, published four days after MindForge and drawn from more than 30 actual use cases across member banks, is a real example of an industry body producing its own guidance ahead of a regulator's final rule, specifically for the generative AI use cases that move fastest and that a multi-year consultation process cannot keep pace with. It is a case study in how a sector organizes itself around an anticipated rule rather than waiting for the rule to arrive.

Example 6: the sector-wide adoption gap the AIRG is built to close (Deloitte AI Institute, 2026). Deloitte's "The State of AI in the Enterprise: The Untapped Edge," published 3 February 2026 from a survey of 3,235 leaders in 24 countries including 75 Singapore respondents fielded in August to September 2025, found that 72 percent of Singapore businesses plan to deploy agentic AI in several operational areas within two years, while only 14 percent report a mature model for agentic AI governance, against a 21 percent global average (Deloitte AI Institute, 2026). Read against the AIRG, this is the exact gap Section 4's capabilities-and-capacity requirement is written to close: a sector racing toward deployment far faster than it is building the governance maturity to manage it, and a Singapore figure running below the global average on readiness even as it plans above-average adoption speed.

Example 7: the agentic layer arrives while the AIRG is still a proposal (IMDA, 2026). IMDA launched Version 1.0 of the Model AI Governance Framework for Agentic AI on 22 January 2026, at the World Economic Forum, with Version 1.5 following just four months later on 20 May 2026, updated again on 5 June 2026 to add case studies and best practices (IMDA, 2026). (see Topic 6.7) That timeline sits almost entirely inside the same window the AIRG itself has spent unfinalized, which is a real, citable example of two Singapore instruments moving at once, on different tracks, toward the same financial institutions that are building agentic features. An institution tracking only the AIRG's status would miss that a second, cross-sector instrument aimed squarely at autonomous action was maturing in parallel, on an even faster cadence.

Example 8: the accountability gap inside financial-sector-adjacent SMEs (AWS-commissioned research by Strand Partners, 2026). Research commissioned by AWS and conducted by Strand Partners, published 6 May 2026 across 1,500 Singapore businesses spanning financial services, healthcare, and manufacturing, found that just under 30 percent of AI-adopting SMEs have a clearly defined person responsible for overseeing AI accuracy, just under 40 percent have no formal process for escalating AI outputs employees are unsure about, and six in ten firms would face significant or moderate disruption, with about one in ten saying AI initiatives would likely stop altogether, if the single person responsible for AI left the organization (AWS-commissioned research by Strand Partners, 2026). This is Section 1's oversight requirement made concrete in negative space: a sector where "the person responsible for AI" is frequently a single individual whose departure would stall governance entirely is exactly the fragility the AIRG's Board-and-Senior-Management oversight duty is designed to prevent from persisting once the Guidelines take effect.

Where people go wrong

  • "The AIRG isn't law yet, so there's nothing to do." This is the single most expensive misreading of the whole topic. The Consultation Paper's comment period closed 31 January 2026 and final guidelines are not yet issued as of 15 August 2026, but MAS already supervises every institution in scope under other, binding statutes, and the industry (MindForge's 24 institutions, the ABS's 30-plus-use-case Handbook) has already normed itself around the AIRG's shape months before it is final. "Not yet law" is a true statement about legal status, not a safe basis for inaction.
  • "Proportionality means small institutions are exempt." Proportionality governs the degree of implementation, not applicability. Every financial institution MAS supervises is in scope; a small institution runs a lighter, proportionate program, it does not run no program. Confusing a dial with a switch is the fastest way to fail Section 2 entirely.
  • "If a vendor built the model, the vendor's compliance covers us." This is Doctrine 2's exact failure mode, and it is a habit inherited from decades of ordinary technology outsourcing where it was often true. The AIRG explicitly requires third-party AI in the institution's own inventory and holds the institution, not the vendor, accountable for that system's risk. A vendor's certifications and documentation are inputs to your assessment, never a substitute for running one.
  • "Our inventory only needs to list the AI we built ourselves." An inventory that excludes licensed, embedded, or vendor-operated AI is incomplete under Section 2 by definition, no matter how thorough it looks for the systems it does cover. The AIRG names third-party AI explicitly precisely because this omission is the most common one.
  • "A human technically reviews every output, so reliance is low." Reliance measures how much a human meaningfully defers to the system, not whether a review step exists on an org chart. A review step that happens at a volume no human can genuinely examine, or that overwhelmingly rubber-stamps the system's default, is high reliance in substance even if a box gets checked in process.
  • "Materiality is the average of impact, complexity, and reliance." Averaging softens exactly the score that should drive the response. A system that is low on two dimensions and extremely high on the third (commonly reliance) can still cause serious, undetected harm at scale, and the proportionate control set should scale to the sharpest dimension, not a blended composite that quietly dilutes it.
  • "FEAT, Veritas, and the AIRG are basically the same thing with different names." They are sequential, not redundant. FEAT stated the values in 2018 with no method attached. Veritas built and field-tested the method across three phases through 2023. The AIRG, still a proposal, is the first attempt to gather that whole arc into a structured, institution-wide set of expectations organized into four sections. Treating them as interchangeable loses the fact that each one exists because the previous one was, by design, incomplete on its own.
  • "The AIRG is Singapore's general AI law, so it covers any AI system in the country." The AIRG is sector-specific: it applies to financial institutions MAS supervises, not to every organization using AI in Singapore. The cross-sector, voluntary instrument is the Model AI Governance Framework family from Topic 6.7; conflating the two misroutes a governance question to the wrong regime entirely. (see Topic 6.7)
  • "Since MAS hasn't finalized the Guidelines, our materiality assessment doesn't need to be defensible yet, just a rough draft will do." A rough, undefensible assessment fails the purpose of building it now, which is to be examination-ready the moment MAS asks, proposed rule or not, and to give the Board something real to exercise oversight over under Section 1. Treating the pre-final period as a low-stakes practice run wastes the exact runway the industry's early movers are using to get it right.
  • "Section 1's oversight duty is really a compliance-department duty." The AIRG places oversight explicitly at the Board and Senior Management level, tied to Singapore's existing fit-and-proper regime for the individuals in those roles. Routing the whole duty down to compliance without ever surfacing it to the Board is a structural failure of Section 1, not a workaround for it.
  • "A materiality tier, once assigned, stays valid until we reassess on our own schedule." The change-management control in Section 3 exists precisely because a vendor update, a data-source change, or a shift in how heavily staff rely on a system's output can silently move a system's real risk without anyone re-running the assessment. A tier is only as current as its last honest re-check against what the system actually does today.
  • "MindForge and the ABS Handbook are official MAS rules we can cite as binding." MindForge is a MAS-convened toolkit built with industry, and the ABS Handbook is an industry association's own publication; neither is the AIRG itself, and neither is binding law. They are strong evidence of where the sector and the supervisor are heading and useful operational tooling, but citing either as "MAS requires" overstates what they are.
  • "Our documents are thorough, so we're examination-ready." A thorough document and a defensible practice are not the same thing. An examiner tests whether a named owner can explain a real recent decision, whether a reliance score is backed by actual evidence of review activity, and whether a director can speak to risk appetite unprompted; a beautifully written entry that nobody behind it can actually defend fails exactly the test that matters.
  • "Explainability means we need to see inside the vendor's model." Full mechanistic transparency into a vendor's proprietary logic is often not available and is not what Section 3's explainability control actually requires. Outcome-level logging, tested monitoring for drift and unfairness, and the ability to trace a specific decision back to its inputs can satisfy the control even where the model's internal weighting is never disclosed; demanding full internal transparency as the only acceptable standard sets a bar the control does not actually set.
  • "An agentic AI system is covered by the AIRG alone, since it's still 'our AI.'" A financial institution running an agentic system, one that acts rather than only recommends, answers to the AIRG's inventory and materiality structure and to the separate agentic-specific expectations in the Model AI Governance Framework for Agentic AI. (see Topic 6.7) Treating one regime as covering both leaves a real gap in the institution's own governance record.
  • "The materiality assessment is a one-time exercise once the system launches." A tier is only as current as its last honest re-check; a vendor update to the underlying model, a shift in transaction volume that overwhelms a review process, or a change in what the system is used for can each move the real risk without a single visible change to the institution's own code, which is exactly why Section 3's change-management control exists.
  • "The Board briefing can reuse compliance's technical language, since the Board just needs the general idea." A director who can only repeat a compliance-prepared slide has not demonstrated the meaningful oversight Section 1 asks for, and an examiner's direct question exposes the gap immediately. A defensible briefing translates the system's risk into language a director can genuinely absorb and later restate unprompted.

Questions people ask

What is Monetary Authority of Singapore (MAS)?
Singapore's central bank and integrated financial regulator, with statutory licensing and examination authority over banks, insurers, capital markets intermediaries, and payment institutions. MAS has built its AI governance regime almost entirely through non-binding instruments, FEAT, Veritas, and the proposed AIRG, rather than a dedicated AI statute. More on Monetary Authority of Singapore (MAS)
What is FEAT (Fairness, Ethics, Accountability, Transparency)?
MAS's four founding principles for the use of AI and data analytics in Singapore's financial sector, published 12 November 2018. Non-binding; it named the values the sector was expected to hold before any assessment method existed to measure them.
What is veritas?
A MAS-led industry consortium that turned FEAT's four principles into a tested assessment methodology and toolkit across three phases, concluding January 2021, February 2022, and June 2023, the last including integration pilots run by seven financial institutions. More on Veritas
What is AIRG (AI Risk Management Guidelines for Financial Institutions)?
MAS's proposed guidelines, released as Consultation Paper P017-2025 on 13 November 2025. The public comment period closed 31 January 2026; final guidelines were not issued as of 15 August 2026, with a proposed 12-month transition period once they are. Organized into four sections: oversight, key AI risk management systems, key AI life cycle controls, and capabilities and capacity.
What is mindForge AI Risk Management Toolkit?
A practical operationalization toolkit, including a handbook and case-study supplement, built by MAS with a 24-institution consortium of banks, insurers, and capital markets firms, launched 20 March 2026, ahead of the AIRG's finalization.

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 →