The MAS guidelines and the trader who still owns the trade
On 7 October 2026, the Monetary Authority of Singapore issued its final Guidelines on AI Risk Management. Olga Magomedova on what a supervisory text written for banks says, by implication, to any trader using AI tools from a home office.
A supervisory text with a plain idea
The headline change is procedural. The idea underneath it is not. The Monetary Authority of Singapore issued a set of Guidelines on Artificial Intelligence Risk Management to support responsible AI adoption in Singapore's financial sector on 7 October 2026. The guidelines take effect on October 7, 2027, and apply to all financial institutions and all forms of AI.
What the document actually says, stripped of its administrative packaging, is that the firm using the model remains responsible for what the model does. Firms remain accountable for AI used in the services they provide even when the technology is developed, operated or supplied by a third party. That is not a technology rule. It is a principle about ownership of outcomes.
The principle applies well beyond Singapore's regulated banks.
What the guidelines actually ask for
The guidelines are deliberately practical. Boards and senior management should oversee AI risk, with clear roles, a stated risk appetite, frameworks, policies and procedures, and firms should identify where they use AI and keep an inventory at an appropriate granularity. They should assess risk materiality, meaning how much a failure would matter, and apply proportionate controls across the AI lifecycle: data governance, testing, human oversight, cybersecurity, monitoring and change management.
Implementation runs on a staggered clock. Implementation is phased, with sections 3 and 4 applying from October 2027 and sections 5 and 6 from October 2028. The sequencing is instructive. An institution is expected to know where its AI lives, and how much any single failure would cost, before it is assessed on how well it controls that AI.
For Magomedova, the sequencing is the lesson. Inventory first. Materiality second. Controls third. A retail trader who runs an AI assisted process without being able to list, in writing, which decisions the model touches has skipped the first step.
The agentic question MAS left open
The most interesting part of the release is the part that was not finalised. MAS flagged agentic AI as the area still to be settled, told firms to review their controls as they make greater use of agentic systems that can operate autonomously and access tools, and said it intends to consult the industry in 2027 on what further guidance on agentic AI would be useful.
A regulator writing in October 2026 does not yet know how to supervise a system that acts on its own authority. That admission is honest. It is also a warning. The tools that execute, size, and close positions without a human in the loop are already available to retail traders. The supervisory framework around them is not.
A tool a regulator cannot yet supervise is a tool the trader has to supervise alone.
The FCA found the same shape of problem
The point was made from a different direction a few weeks earlier. On 2 September 2026, the Financial Conduct Authority published a new webpage setting out insights for firms on how frontier artificial intelligence may affect cyber resilience, governance and vulnerability management for firms. The UK's financial regulator warned that the rapid pace of AI development is revealing vulnerabilities in firms' cybersecurity arrangements because their existing controls cannot keep up with the risks new technology can either create or exploit.
Two different jurisdictions, two different documents, one shared observation. Capability is arriving faster than the governance around it. The firm, or the trader, that waits for the rules to catch up will have been operating without them for the interval in between.
What this means at a single desk
Translated to a retail practice, the MAS framework collapses to three habits. Keep an inventory of every AI tool in the process, including the ones embedded in the broker's platform. Decide, in writing, which decisions the tool is allowed to touch and which it is not. Review that inventory on the same cadence used for the trading plan itself.
None of this is sophisticated. It is the same discipline an engineer applies to a system of any complexity. Know the components. Know the failure modes. Know what remains when one component fails. The post-mortem artefact matters more than the opinion about the artefact.
AI is a powerful tool. But tools do not replace responsibility.
The principle that travels
The MAS text was written for institutions with compliance departments and risk committees. Its central principle survives translation to a single desk without any loss. The entity using the model owns the outcome the model produces. The vendor does not. The platform does not. The regulator, when it eventually arrives, will not retroactively absorb the loss.
For Magomedova, that is the quiet point worth taking from the October release. Accountability is not a feature a supervisor can add from the outside. It is a property the operator chooses to hold. The trader who holds it, in writing, before the next tool arrives on the screen, is the trader who will still be running their own process when the next one does.