AI regulation

Is Singapore Leading the Globe with New AI Identification Rules? What It Means for You

Singapore's financial regulator will soon expect every bank, insurer and fund manager it supervises to know where AI is running inside the business, agents included. Most organisations in Singapore, and globally, don't have that information.
16 November 20267 min read

This week, more than 70,000 people will walk into Singapore EXPO for the Singapore FinTech Festival, where this year's theme is "New Risks. New Rules. New Growth." PromptLeash will be there, and we believe "New Rules" will be the key talking point.

On 7 October 2026, the Monetary Authority of Singapore (MAS) published its final Guidelines on Artificial Intelligence Risk Management. They apply to every financial institution MAS supervises and to every form of AI those institutions use. Section 4 is particularly notable to us, as it requires firms to find all of their AI and keep an accurate list of it.

Can you identify every single AI system running in your organisation right now?

What Are These Guidelines?

MAS released the draft guidelines for consultation on 13 November 2025, the same week as last year's festival. Consultation closed at the end of January, and the final document was released last month.

The guidelines take effect on 7 October 2027. Firms have until then to meet the oversight, identification and inventory expectations, and until October 2028 for the heavier life cycle controls and capability building. That leaves eleven months to address three core expectations.

Identify it. Firms need consistent processes for finding AI across every business and functional area. That includes AI embedded in services from material third-party providers, and even SaaS tools with AI features that aren't marketed as AI. Someone independent of the teams using AI must own how the firm finds AI, including the definitions, checks and records, and make the final call when it's unclear whether a tool counts.

Inventory it. Firms must keep an accurate inventory of AI use cases, systems and models, and update it as things are added, changed or retired. Suggested fields include purpose, scope of use, data used, dependencies, life cycle status, owners and a risk materiality rating. For agents, MAS suggests the inventory capture agent identifiers, the tools and systems each agent can access, and the agent's guardrails. The inventory also has to link to other registers, such as data and third-party registers.

Rate it. Every use case gets a materiality assessment across three dimensions: impact, complexity and reliance. Reliance covers how much autonomy the AI has and how much human involvement remains. It is a very important factor, and we will break down the levels of reliance in a future post, because we believe that increased AI autonomy does not equate to increased productivity.

So, Is Singapore Leading the Way?

Yes, at least in some respects. Singapore isn't the first regulator to expect an AI inventory, but its guidelines are the most explicit we've seen about finding all AI used throughout the business.

Below are the countries and regions with guidelines or regulations relating to the visibility of AI within a firm. Links to each are in the sources for this post.

RegulatorDocumentStatusWhat it asks for on AI visibility
Singapore (MAS)Guidelines on AI Risk Management, 7 Oct 2026Final (in effect Oct 2027)Firm-wide identification, an AI inventory and a materiality rating for every use case.
Canada (OSFI)Guideline E-23 Model Risk Management, 11 Sep 2025Final (in effect May 2027)An inventory of models that carry non-negligible inherent risk.
Hong Kong (HKMA)ART Framework and updated supervisory guidance, 15 Oct 2026Supervisory guidanceNine principles across Accountability, Resilience and Trust.
Australia (APRA)Letter to industry on AI, 30 Apr 2026Expectations under existing prudential standardsGovernance isn't keeping pace with adoption. Expects firms to map their full AI supply chain, including fourth parties.
European Union (EU AI Act)Digital Omnibus amendments, agreed 2026Binding law (high-risk obligations delayed)The heaviest duties apply to high-risk systems such as credit scoring, delayed until December 2027.

A few jurisdictions moved before Singapore, notably Canada and the EU, and are ahead in certain respects. Canada is ahead on the inventory itself: its guideline comes into effect five months before MAS's and expects an inventory of the models that carry meaningful risk. The EU has the bigger law, as the AI Act is binding legislation, but its most demanding obligations were pushed back by 16 months. In contrast to both, Singapore's guidelines make the board responsible for all AI in a regulated firm, not only the high-risk tier.

Australia and Hong Kong are heading the same way. APRA's letter was its first set of AI-specific expectations, drawn from a deep-dive review of large banks, insurers and superannuation trustees. Its central finding was that AI deployment is outpacing the maturity of risk frameworks.

The EU is the only one with actual legislation in place. However, the "Guidelines" title doesn't mean optional. MAS guidelines aren't legislation, but they set the supervisory expectations that supervisors will check against.

The Hurdles of Meeting the New Guidelines

The intuitive way to build an AI register is to send a survey down the hierarchy, asking people to list the AI tools they use. However, multiple surveys suggest that around half of all employees conceal their AI use or are uncomfortable admitting to it. Others will forget some of their use cases, and the definition of what counts as AI may be misunderstood.

That misalignment produces three main groups of error.

Shadow AI. These are tools employees use without approval, such as a personal ChatGPT account, a browser extension or a free trial. MAS expects firms to identify the risk from unidentified AI, including shadow AI, and to keep that residual risk within appetite. That makes shadow AI a supervisory question, not only an IT one.

Embedded AI. Your CRM, document platform and HR system have added AI features in the last year. Some were switched on by default, and few were announced as "new AI". MAS states plainly that SaaS with AI features not marketed as AI is in scope.

Agents. An approved model running inside an approved platform can still be wired into an agent that reads email, queries databases and takes actions. The inventory question for an agent includes its scope and influence. Few organisations can answer that today across every team it affects.

What It Means for You

If you're a financial institution in Singapore, the clock is running. You have until October 2027 to have identification processes, an inventory and materiality ratings in place.

If you're part of a regional or global group, MAS lets Singapore entities lean on group-wide AI risk frameworks, as long as those frameworks meet its expectations. In practice, the Singapore standard is likely to become the bar for the whole group.

If you're in Australia or elsewhere in the region, treat this as a preview. APRA has already said AI governance isn't keeping pace with adoption, and HKMA is building its framework around agents. More scrutiny is coming, and the firms that can answer it as they grow their AI use will have a far smoother ride.

Wherever you are, the work for the next eleven months looks much the same.

  1. Agree what counts as AI.
  2. Find usage from evidence, instead of relying on surveys alone.
  3. Build the inventory and keep it current.
  4. Rate what you find for impact, complexity and reliance.
  5. Start on capability now. The board needs to understand AI well enough to challenge it, and that sits in the oversight expectations due in October 2027. The requirement that the people who build, deploy and use AI are competent follows in October 2028. Both take longer than an inventory, so starting early pays off. It's also where course completions alone fall short.

More Is Coming Soon

An AI inventory answers the regulator's first question: what AI do you have? The next questions are harder. Who is using it, for what work, and is it delivering value?

This is why we built PromptLeash: to show how people and agents actually use AI across your organisation, beyond licences and logins. It links that use to roles and workflows, so you can see which teams depend on which tools and where agents are operating. Its AI governance view checks usage against your internal policies and gives you an AI risk score.

For MAS-style identification, this means your inventory starts from evidence of real use rather than recollection, and it keeps updating as usage changes. For the capability expectations, the same data shows where role-specific training will make the biggest difference. Because PromptLeash deploys in your private environment, no usage data leaves your organisation.

PromptLeash doesn't meet the guidelines for you on its own. Ownership, materiality ratings and sign-off stay with your control functions, as MAS intends. What we provide is the visibility those decisions depend on.

Meet us at the Singapore FinTech Festival. We'll be at Singapore EXPO from 18 to 20 November 2026. If AI identification, agent governance or proving AI value is on your list for 2027, come and find us, or book a time to meet.

Sources

Discuss the opportunity

Bring your next AI investment question.

Book a demo to see how PromptLeash interprets approved AI activity, and how it could support your next decision.

Where this applies

Explore by role, industry and trust.