Start from your job, not from the law.

Start with the binding law: PDPA 2012

Every AI use of personal data sits under this Act, it is the only economy-wide binding instrument in the stack.

Consent Obligation (s 13)

If your AI feature collects or uses personal data, you need consent or a schedule exception before anything else.

Notification Obligation (s 20)

Individuals must be told the purposes of collection, vague "product improvement" wording will not carry AI training uses.

Purpose limitation (s 18)

AI uses must stay within purposes a reasonable person would consider appropriate, check this before re-purposing data for training.

Know when vendors are data intermediaries

Your AI vendor's obligations differ from yours; misallocating them is the most common contract gap I see.

PDPC AI advisory (1 Mar 2024)

This is how PDPC reads the PDPA for ML recommendation and decision systems, your primary interpretive text.

Business improvement exception

Often the only realistic basis for training on existing customer data without fresh consent, learn its conditions cold.

Data minimisation in development

PDPC expects you to train on the minimum attributes and volume, document why you needed what you used.

PDPC GenAI advisory (final, 20 Jul 2026)

The July 2026 finals govern GenAI development and deployment, they run alongside, not instead of, the 2024 guidelines.

Publicly-available exception for web-scraped data

Scraping is not automatically lawful; the exception analysis now turns on digital barriers.

Document your publicly-available assessment

If you rely on the exception for data behind digital barriers, a DPIA or written record producible to PDPC is now expected.

AI-specific notifications

General privacy notices are insufficient for training consent, plan explicit AI-specific wording and an opt-out path.

Breach notification (Part 6A)

A poisoned or leaking training set can be a notifiable breach, wire this into your AI incident runbook.

Penalty exposure: 10% of Singapore turnover

Board-level number: enforcement now carries turnover-based penalties, which is how you get budget for the programme above.

MGF 2.0: the base operating model

The four areas are the mental model every Singapore framework reuses, internal governance, human involvement, operations, stakeholder comms.

Operations management

Data quality, lineage, bias minimisation and post-deployment monitoring are your engineering checklist, verbatim from the framework.

Separate training and test datasets

A basic the auditors will ask about first, keep the split clean and documented.

Pick the human-involvement model deliberately

In/over/out-of-the-loop is a design decision you must justify by severity and probability of harm, not a default.

MGF for Generative AI: nine dimensions

If you build or fine-tune GenAI, these nine dimensions are the canonical risk map your compliance team will quiz you against.

"Food label" disclosures

Ship transparency artefacts with your model, downstream deployers depend on your disclosures for their own compliance.

MGF for Agentic AI

The January 2026 framework is the reference design for governing agents, read it before wiring tools to an LLM.

Bound agent limits upfront

Minimum tools and data access per task is the single highest-leverage control you can implement.

Baseline testing before deployment

Execution accuracy, policy adherence and tool use are the new test dimensions, add them to your eval harness.

Gradual rollout with continuous monitoring

Ship agents behind staged rollouts; production behaviour, not offline evals, is where agent risk shows up.

CSA Guidelines on Securing AI Systems

Security across five lifecycle stages including end-of-life, the companion guide turns each stage into concrete controls.

CSA Agentic Addendum (final, 17 Jun 2026)

Threat modelling, taint tracing and rogue-action controls purpose-built for agentic systems, your security review template.

AI Verify toolkit

Government-built, open-source tests mapped to 11 principles, run it and keep the report; it is your evidence.

Project Moonshot

Benchmarking plus red teaming for LLMs, use it for pre-deployment safety baselining of any GenAI feature.

MAS TRM Notices: the binding baseline

These notices under the FSM Act are law for your institution, AI systems sit inside this perimeter, not outside it.

The 1-hour incident notification rule

A relevant incident on a critical system must reach MAS within one hour of discovery, test this clock with AI failures in scope.

Max 4 hours unscheduled downtime per 12 months

If an AI system is critical, its downtime budget is four hours a year, architecture decisions follow from this.

FEAT Principles

MAS's fairness, ethics, accountability and transparency expectations, the standard your AIDA use is informally measured against.

Board awareness of AIDA use

FEAT puts AIDA on the board agenda; your reporting line should make AI risk visible upward, not buried in model inventories.

Veritas methodology and toolkit

The MAS-industry FEAT assessment methodology, use it to evidence fairness testing on credit, marketing and fraud models.

MAS P017-2025: proposed AI Risk Management Guidelines

Still at consultation as of 17 Aug 2026, but this is the shape of coming supervisory expectations; gap-assess now.

Proposed: board and senior management oversight

The draft guidelines put AI oversight squarely on the board and senior management, prepare your governance charter early.

Proposed: AI inventory

A current inventory of AI use cases is the first thing supervisors will ask for; build it before the guidelines finalise.

Proposed: risk materiality assessment

The proposed proportionality model, materiality determines control depth, so your assessment methodology matters.

MindForge AI Risk Management Toolkit (Mar 2026)

The industry-built handbook aligned to the proposed guidelines, your most practical implementation reference.

Implementation examples (DBS, Julius Baer, Prudential)

Peer case studies show what "good" looks like to the consortium, benchmark your controls against them.

ABS GenAI Guardrails Handbook

Nine guardrail approaches for banking GenAI, an important input to the MAS toolkit, practical for vendor and use-case reviews.

Watch P012-2026: TRM amendments

Proposed amendments to all 11 TRM notices closed 31 Jul 2026, final text could tighten incident and change management for AI.

MGF 2.0 in one page

Singapore's baseline framework, its four areas are the agenda for your AI governance committee.

Internal governance structures

Roles, SOPs, training and periodic review, this is the management system you are personally answerable for.

AI solutions should be human-centric

One of two guiding principles, the tone-setter for every policy your organisation writes.

PDPA: where the legal risk lives

The only binding economy-wide law here; AI misuse of personal data is enforced with turnover-based penalties.

10% of Singapore turnover

The number that makes AI data governance a board topic, not an IT topic.

PDPC GenAI finals (20 Jul 2026)

Fresh, final guidance, ask your DPO for a one-page gap assessment against its four tightened positions.

Agentic AI: the 2026 frontier

The world's first agentic framework came from IMDA, if your strategy includes agents, this is your governance baseline.

Make humans meaningfully accountable

The core ask: named humans accountable for agent behaviour, with approval checkpoints on irreversible actions.

CSA agentic security addendum

Security guidance your CISO should already have mapped, rogue actions and data disclosure are board-visible incidents.

MAS direction of travel

For financial groups: proposed guidelines signal board-level AI oversight is becoming supervisory expectation.

FEAT: board awareness

MAS expects the board itself to be aware of AIDA use, a standing agenda item satisfies this.

SS ISO/IEC 42001 certification

The certifiable management-system standard, a credible signal to regulators, partners and customers.

AI Verify as your evidence layer

Government-built testing that turns "trust us" into reports, fund it before you need it.

The regional frame

ASEAN's roadmap and DEFA will shape cross-border AI business, strategy teams should track the November 2026 signature target.

PDPA 2012: the binding anchor

All AI-personal-data advice starts here; note Parts 3-6 obligations and the Part 9C enforcement toolkit.

Enforcement: directions and financial penalties

s 48I directions and s 48J penalties are what your risk advice quantifies against.

MAS TRM Notices: binding sectoral law

For FI clients, these notices carry statutory force under the FSM Act, the 1-hour rule is the classic exam question.

Advisory guidelines are not law

PDPC's own words: advisory, not legally binding, calibrate client advice and contract wording accordingly.

GenAI finals of 20 Jul 2026

Final text supersedes the June draft analysis; the notify-the-source proposal was dropped, replaced by documented assessment.

Consent cannot be a condition of service

s 14(2)(a) PDPA applied to AI training, review client consent flows for overreach.

Anonymisation as an alternative

The finals accept anonymisation as an alternative to minimisation, a practical out for training-data design.

P017-2025 is still a proposal

As of 17 Aug 2026 there is no final guideline, advise clients on trajectory, not compliance obligations, here.

CSA addendum: final but voluntary

Published 17 Jun 2026 as guidance, expressly non-mandatory, relevant to duty-of-care arguments, not statutory compliance.

SS ISO/IEC 42001 status

Voluntary standard unless incorporated by contract or regulation, check procurement documents for incorporation by reference.

Official NIST crosswalk

For clients with US touchpoints, the IMDA-NIST mapping is the interoperability evidence.

Official ISO 42001 crosswalk

AI Verify ↔ ISO/IEC 42001 mapping supports dual-compliance arguments in certification discussions.

Contrast: Vietnam's binding AI law

ASEAN's first binding AI statute in force 1 Mar 2026, relevant for regional clients and for "Singapore is not the EU" advice.

DEFA: the coming regional layer

Negotiations concluded 29 May 2026; signature targeted November 2026, monitor for cross-border AI and data provisions.

Accept the shape of the problem

There is no single federal AI law to comply with. Anyone who tells you otherwise is selling something.

Section 5 of the FTC Act is your general purpose exposure

Every claim you make about your AI is an advertising claim, and every tool you ship is capable of being a means of deception.

Then go to the states

This is where binding AI specific obligation lives today. Work the 50-State AI Law Atlas for the jurisdictions you operate in.

Watch the preemption fight, do not plan around it

The federal government is litigating against state AI laws. Until a court rules, the state law still binds you.

Adopt the AI RMF because your counterparties will ask for it

It is voluntary, and it is what contracts, insurers and state statutes point at.

M-25-22 is how your buyer is required to buy

Competitive marketplace, measurable performance, cross functional review. Your proposal is read against it.

M-25-21 is what your buyer must be able to say about your system

Chief AI Officer sign off, use case inventory entry, and the minimum practices if the use is high impact.

Know whether your use case is high impact

If it is, and the practices cannot be met, the agency is directed to stop using it. That is a contract risk, not a policy risk.

If you sell an LLM, the Unbiased AI Principles are contract terms

Truth seeking and ideological neutrality, written into new contracts and retrofitted into existing ones.

Expect your deployment to appear in a public inventory

Agencies publish their AI use cases annually with risk determinations attached.

Title VII binds you, tool or no tool

You are liable for the selection procedure you use, including one a vendor built and scores.

The four fifths rule is the arithmetic

Adverse impact analysis is a 1978 regulation, and it is what your AI screening tool will be measured against.

Withdrawn guidance is not repealed law

The EEOC removed its AI documents in January 2025. Nothing about your exposure changed.

The ADA is the second exposure, and the one people miss

Screen out risk, and the duty to offer an alternative format on request.

Then check your states

Several states and cities regulate automated employment decision tools directly. The Atlas has them.

ECOA and Regulation B are the binding floor

Specific and accurate principal reasons for adverse action. The rule does not care how complex the model is.

Read what was actually withdrawn

Circulars 2022-03 and 2023-03 went in the bulk withdrawal of 12 May 2025. The regulation behind them did not.

Use the AI RMF as your documented method

When an examiner asks how you govern model risk, a recognised framework is a better answer than an internal memo.

State law is where the AI specific duties are

Check the Atlas for the states you lend in.

Start with the CISA joint guidance

AI data security and secure deployment, voluntary and widely cited.

Know the Treasury clearinghouse exists

Voluntary coordination on AI driven vulnerability scanning with industry and critical infrastructure operators.

Understand covered frontier models

Classified benchmarking of model cyber capability, and a voluntary access framework for developers.

Read the bar on licensing before you assume a regime exists

The order expressly disclaims mandatory licensing, preclearance or permitting.

Track the agent standards work

Agent identity and authorisation is the open control gap, and this is where the US answer is being written.