MAS AI risk management guidelines: what a Singapore fund manager needs by October 2027.

MAS issued the Guidelines on Artificial Intelligence Risk Management on 7 October 2026. Licensed fund management companies are on the list of institutions they apply to. The question for a Singapore manager is not whether, but how much.

This guide answers for Singapore, and for the managers licensed there: hedge funds, private equity and venture firms, and multi-strategy houses. A small manager will qualify for the lighter path on most of its AI use. It will still have a year of work ahead.

There are two dates, and the second one is the expensive half.

Paragraph 1.8 splits the commencement. The Guidelines take effect on 7 October 2027, and a firm may meet Sections 3 and 4 from that date. Sections 5 and 6 can wait until 7 October 2028.

Sections 3 and 4 are oversight, identification, inventory and risk materiality assessment. Sections 5 and 6 are the life cycle controls and the capacity to run them. The first year produces the list. The second year is what you must then do to everything on it.

The lighter path is a test on each use, not an exemption for the firm.

Paragraph 2.3 offers a basic set of AI governance policies and procedures instead of the full framework. It is available where poor performance or unavailability of the firm's AI services is unlikely to have a material adverse impact on the firm, its customers or other stakeholders.

Paragraph 2.4 gives six examples that would generally meet the criteria. They are drafting or proofreading emails to customers, summarising documents or meeting notes, and the initial review or analysis of documents. Then generating formulas or charts, image generation for marketing material, and a chatbot that helps staff find internal policies. Three things narrow that list.

Basic is not nothing. Paragraph 2.5 sets six minimum items. A named senior manager accountable for AI oversight. Permitted and prohibited uses, with defined points of mandatory human review. An approved tool list with a request process, staff education, regular compliance checks, and that review in (f). Footnote 13 adds a condition under the examples themselves: the humans in those use cases should be checking outputs before using them.

Identification is the real work, because the AI is already in your stack.

Paragraph 4.2 asks for consistent identification of AI use across all business and functional areas, and then reaches further. Where AI is incorporated within, or used in the delivery of, third-party services, the process should at minimum cover AI in the services of material third-party providers. Footnote 21 is explicit: software sold as something other than AI, but carrying AI features the firm uses as part of the service.

For a fund manager that is the portfolio and order management system, the CRM, the document platform, the research terminal and the fund administrator's portal. AI features tend to arrive in those systems through a release note rather than a contract variation.

Paragraph 1.3 draws the boundary: calculators and tools whose outputs rest solely on predefined logic are not AI, and footnote 6 puts rule-based systems and robotic process automation outside as well. The waterfall macro in the distribution model is out of scope. The assistant reading the data room is in.

Paragraph 4.3 asks for a designated control function responsible for identification, with the final say on what counts as AI use. At a twenty-person manager that is a named person, not a department. Paragraph 4.4 then expects residual risk from unidentified AI use to sit within risk appetite, with mitigating measures where poor vendor disclosure or shadow AI makes that impractical.

The inventory and the materiality assessment are one piece of work.

Paragraph 4.7 lists the inventory attributes. Purpose, approved scope of use including jurisdiction, model type, data used, dependencies and life cycle status. Then the assigned risk materiality rating, the named owner and links to the essential documentation.

Paragraph 4.12 sets the dimensions the assessment must cover at minimum. Impact means the consequences of failure for the firm and its stakeholders, and the sensitivity of the data involved. Complexity includes how much visibility the firm has into a third party's technology. Reliance covers the autonomy granted to the system and the degree of human involvement in the process it supports.

Reliance is the dimension that moves as a firm matures. An assistant that drafts a paper for a human to sign rates lower than an agent that updates a monitoring record on its own. That is why paragraph 4.10 assesses the use case, not the product.

Your vendor's assurance is not your control.

Paragraph 5.11 is blunt. The decision to onboard third-party AI is the firm's, and the firm retains primary accountability for its use, including the effect on its own ability to comply with regulatory requirements. Where a provider's disclosures are thin, paragraph 5.11(a) points to external assessments by independent and competent parties, and says in terms that self-attestations do not count.

Two items belong in procurement now. Paragraph 5.11(f) expects legal agreements to cover performance, data protection, a right to audit, and notification when AI is introduced or updated. Paragraph 5.10 expects the firm to test third-party AI against its own use cases, using its own data, and to document why it judged the service suitable. A vendor questionnaire answered once at onboarding satisfies neither, and the retention terms are where that usually comes apart.

Paragraph 5.23(d) then asks for records of monitoring activity, results, incidents and remediation, with access to systems and data logged. For generative AI and agents it suggests logging prompts and model responses, together with the model versions used. That record cannot be made retrospectively. Where AI arrived tool by tool, nobody can say which system touched which memo, or which model version produced a number an investor has already seen.

Our position is that a Singapore manager should stop debating the perimeter and go and find the AI. The classification arguments resolve quickly once the list exists. Where AI runs through a single layer that records the system, the model and the material behind each output, the inventory and the monitoring record are read off the firm's own logs. Nothing in the Guidelines requires that architecture. Every section assumes somebody can answer. APRA made the same move at home through CPS 230, and the European Union through the DORA register.

Primary sources: the Guidelines on Artificial Intelligence Risk Management and the MAS media release of 7 October 2026. This guide is general information, not legal advice. Whether a particular use meets the paragraph 2.3 criteria depends on the facts and on the position your supervisor takes.