By Lynn Räbsamen, CFA | Advisory Board Member, CFA Institute | Author, Artificial Stupelligence
For over 30 years, software waited. You clicked, it answered. Agentic AI does not wait. It has a task and the authority to decide.
In regulated finance, that last part is a problem. Three questions decide where it becomes your problem.
Singapore Confirmed What Was Already True
On 5 August 2026, Singapore’s financial regulator confirmed that agentic AI falls inside its supervisory guidelines for banks, insurers, and asset managers.
MAS did not write new rules for agentic AI. It confirmed that the existing rules include them.
Those guidelines have been out for consultation since November 2025. They cover board oversight, risk frameworks, and controls across the AI life cycle. They apply to all AI use cases including agentic AI, and they will be finalized soon.
It is the first time a major financial regulator has said it that plainly. The direction is now on the record.
So nothing became compulsory this week. But what became clear is: when the guidelines land, agents will already be inside them.
Who Should Be Reading This
Directly: the banks, insurers, capital markets firms, and fintechs licensed in Singapore. They now know what the final text will cover.
Indirectly: anyone who has been waiting for a definitive agent rulebook before doing anything.
Singapore just showed you what that agentic AI rulebook looks like.
It looks like your existing obligations, with one sentence confirming they apply to agents.
That is worse than a new rule, not better. A new rule arrives with a date and a checklist. This arrives with neither. And it has been in force conceptually for a while.
Washington Went the Other Way and Landed in the Same Place
In April, US banking regulators rewrote model risk guidance for the first time since 2011.
SR 26-2, issued jointly by the Federal Reserve, the OCC, and the FDIC on 17 April 2026, replaced SR 11-7 and the 2021 statement on anti-money laundering systems. 15 years of accumulated practice, superseded in one letter.
Generative and agentic AI are described as novel and rapidly evolving, and placed outside the scope of SR 26-2 entirely, with a note that a bank’s own risk management practices should determine controls for anything not covered.
Translated: this does not cover your generative or agentic AI. Please cover them yourself.
Singapore said the rules include agents. SR 26-2 said they exclude agents. You are accountable in both cases.
Britain Did the Work and Got None of the Headlines
The UK fixed accountability years ago. The FCA’s 2024 AI Update confirmed the existing framework applies: Consumer Duty, the Senior Managers and Certification Regime, operational resilience. No dedicated senior manager function for AI. Delegating a decision to an algorithm does not move it.
In January 2026, the Treasury Committee called this a wait-and-see approach and warned it exposed consumers and the financial system to harm. It then declined to recommend an AI-specific regime and asked the FCA to explain, by the end of 2026, what assurance senior managers actually owe.
The Mills Review: Seven Recommendations, One of Them Awkward
Then in July 2026 the FCA published the Mills Review. Seven recommendations, commissioned by the Board and described by the FCA as the first exercise of its kind by any regulator.
The human role runs from operator, through collaborator, consultant and approver, to observer, as AI takes on more of the work.
SM&CR and the Consumer Duty stay as they are. Firms are told to audit where each deployment sits on the spectrum.
Recommendation six is where it gets interesting: “Build and adopt an AI-enabled agentic supervisory model.”
The FCA’s answer to supervising autonomous systems is to deploy autonomous systems.
Somebody should write that down before it becomes normal.
Switzerland is already there. FINMA has built generative AI tools for market supervision, covering document analysis, market abuse investigations, and crypto exposure monitoring. Its Chair also chairs the IOSCO SupTech Forum and told an audience in Zurich in June that the real revolution is not in finance but in supervision.
Singapore got the landmark coverage. Britain published the actual instruments. Switzerland already built one.
Switzerland Named the Variable, Not the Technology
Here is the part that should interest anyone in Zurich.
FINMA published guidance on AI governance in December 2024. It does not mention agents once. What it says instead is that risk depends on how complex an application is, how much it adapts, how autonomously it acts, and how deeply it sits inside a process.
That ages better than naming a technology. Autonomy is a dial, not a category. An agent is an ordinary application with the dial turned up.
FINMA also named the hard part. Responsibility becomes difficult to allocate when systems act on their own and are hard to explain, and when ownership is spread across the institution.
Switzerland described the problem 18 months early and gave it no name. Which is why nobody called it a landmark.
Then the practical bit. FINMA found that some institutions defined AI narrowly, and that AI inventories were often incomplete because use was scattered across the organization.
That was a warning about scope. It reads differently now. Most agentic workflows did not exist when those inventories were drawn. So the list is not wrong. It is just older than the technology it is supposed to cover.
The Human Kill Switch
Most of the commentary lands on the same place: name a human, give them a kill switch.
That is a poor answer, and not because it is wrong.
An agentic workflow is not one decision. It is a chain of them. Pull the data. Reconcile it. Screen the universe. Draft a rationale. Produce a recommendation. Send it to the client.
The efficiency of these systems comes from continuity. Every human checkpoint is a real cost, paid in the thing you bought the system for.
Insert a human at every stage and you have not built an agentic workflow. You have built an expensive one.
Which means checkpoints have to be placed deliberately, not sprinkled.
The Three Questions You Should Ask
Here is the useful exercise, and it is not about technology.
Take one workflow. Break it into steps. For each step, answer three questions.
- Can this step be fully automated? Data retrieval, reconciliation, screening, formatting, monitoring. Usually yes. Nobody’s regulator has ever asked who personally approved a currency conversion.
- Does this step create a regulated outcome? Suitability. Advice reaching a client. Order execution. Pricing. Anything that would require documentation if a person had done it. These are where the liability lives, and they are typically a small minority of the chain.
- Who owns that specific step? Not the system. The step. A named person, with the standing to stop that step and only that step.
Do this properly and you get something more valuable than a governance policy. You get a map showing exactly where automation is free and where it is expensive, which is the same map your business case should have been built on.
Most firms have not drawn it. They have an AI inventory instead, which is a list of systems, not a list of decisions. Regulators are asking about decisions.
One Final Question for the Board
Not just “who owns our AI.” But:
In our largest agentic workflow, which steps produce a regulated outcome, and who owns each one by name?
If the answer is a system diagram, you have documentation. If the answer is a list of decision points with people attached, you have governance.
Disclosure: This article was partially drafted by AI and reviewed by a human.
For more insights about what AI can or cannot do, check out my book “Artificial Stupelligence: The Hilarious Truth About AI“.
Subscribe here to be the first to receive my insights
