
Singapore FinTech Festival 2026 and the MAS AI Guidelines: Can You Show Your AI Works for Customers?
The MAS Guidelines on Artificial Intelligence Risk Management, issued on 7 October 2026, take effect for MAS-regulated financial institutions on 7 October 2027, with life cycle controls (including third-party AI) and AI capability and capacity expected by 7 October 2028. Institutions remain accountable for third-party AI and should test it in the context of their own use cases.
On Wednesday, in his keynote at the Singapore FinTech Festival, Deputy Prime Minister Gan Kim Yong announced three National AI Missions for Financial Services: strengthening financial crime prevention, improving access to financing, and enabling safe agentic transactions, meaning AI agents that carry out transactions on a customer's behalf.
He was just as clear about the condition attached: "Financial institutions must remain accountable, with strong human oversight and effective safeguards." [3], [4]
That line sums up SFF 2026 for anyone deploying AI in a bank or insurer. The festival's theme, "The Coming Rewiring of the Financial System: New Risks. New Rules. New Growth", puts all three side by side. The new rules already exist. MAS issued its Guidelines on Artificial Intelligence Risk Management on 7 October, and they take effect on 7 October 2027, less than 11 months from now.
I'm at Singapore EXPO this week, and the question that matters in these meeting rooms is not whether to deploy AI. It is whether you can show that your AI works for the customers it serves. [1], [5]
From Pilot to Production
SFF's technology theme this year looks at how agentic AI, tokenisation, digital money and quantum technologies are reshaping financial services. The opportunity is real: AI can support customer service, automate operational workflows, strengthen risk management and enable new financial products. [5]
Moving from pilot to production raises questions that model performance alone cannot answer.
I spent 20 years running software businesses for capital markets. A system can pass every acceptance test and still fail the people who use it. Take a bank's AI assistant explaining the early-withdrawal fee on a fixed deposit. It scores well on internal benchmarks and retrieves the right product terms. Now put it in front of customers who write in Malay, Mandarin, Tamil or Singlish, including a retiree new to digital banking. Does each of them come away understanding the fee? Or did a technically correct answer lead one of them to a decision they would not otherwise have made?
These are deployment-level questions. Benchmarks tell you how the model performs. They do not tell you whether your customers understood it.
A broader approach to AI assurance asks three questions:
Does the AI work for its intended audience? Can real users understand, interact with and appropriately rely on its outputs?
Does it achieve the intended business outcomes within established guardrails? Does it do its job without creating unacceptable operational, customer or business risks?
Can the institution show that its AI is governed and defensible? Is there enough evidence to support its decisions, controls and oversight?
Those questions sit close to what MAS's new Guidelines expect.
What the MAS AI Risk Management Guidelines Require
On 7 October 2026, MAS issued its Guidelines on Artificial Intelligence Risk Management for Financial Institutions. They apply to all MAS-regulated financial institutions, in proportion to each institution's size and risk profile, and cover all forms of AI, including generative AI and AI agents. The expectations come in two stages.
Date | Expectations |
7 October 2027 (from) | The Guidelines take effect. Section 3, AI oversight, and Section 4, AI identification, AI inventory and risk materiality assessment. |
7 October 2028 (by) | Section 5, AI life cycle controls (including third-party AI management, evaluation and testing, and post-deployment monitoring), and Section 6, AI capability and capacity. |
2027 | MAS intends to consult the financial sector on what additional guidance on agentic AI would be useful. |
[1], [2]
Beyond Vendor Testing: Why Real-World Validation Matters
Much of the AI financial institutions deploy will be bought rather than built. A system may pass a provider's technical benchmarks and still behave differently inside an institution's own environment, particularly when it talks directly to customers or influences financial decisions.
Consider a third-party assistant introduced to support insurance customers. The vendor shows that it retrieves policy information accurately and performs well across a set of predefined test scenarios. Those results do not show whether customers understand the answers, whether explanations stay clear across different audience groups, or how the assistant behaves when it meets questions the vendor never tested. In a specific market it will meet different languages, communication styles, levels of financial literacy and expectations.
A vendor's test report tells you how the system performed for the vendor. It does not tell you how it will perform for your customers.
MAS is explicit that financial institutions remain accountable for AI used in the services they deliver, including AI from third parties. The Guidelines put it plainly: "The decision to onboard and use third-party AI is ultimately a decision of the FI." They expect contracts that give the institution a risk-appropriate degree of visibility over the AI and changes to it, testing of third-party AI in the context of the institution's own use cases, and documentation of how it was tested and deemed suitable. Where residual risk sits outside the institution's risk appetite, it should consider limiting or suspending the service, or replacing the provider.
Two more lines are worth noting. Testing should cover conditions "from real-world scenarios to edge cases". And when selecting AI systems, "The FI should also consider incorporating reviews by domain experts or users". [1], [2]
The aim is not to replace automated testing. It is to answer what automated evaluation cannot: how real people interpret, experience and respond to an AI system in practice.
Life cycle controls: evidence that keeps up with change
AI risk management does not end when a system is approved for deployment. An assistant may behave differently after a model update, a change to its knowledge base or a change made by a third-party provider. An agentic system may meet new risks as its access to tools, workflows and decisions expands.
So testing evidence collected before launch goes stale. MAS says the same: "Significant changes should trigger appropriate reviews and re-approval processes before implementation", and AI should be re-validated where monitoring or change calls for it. [1]
The gaps are easy to miss. An institution may have completed its AI inventory and documented the materiality of a customer-facing assistant, yet have little evidence that different customer groups can understand its responses or complete the tasks it was built for.
Where Adrosaga fits: testing AI with the people it serves
Internal QA and automated evaluation show how an AI system performs technically. What they cannot show is whether real people interpret its outputs correctly, whether interactions produce the intended outcomes, and whether performance differs between customer groups.
Adrosaga's AI Deployment Testing answers those questions with human-led testing in realistic deployment scenarios. Adrosaga recruits testers who reflect your customers, by country, language, dialect, fluency, age and accessibility needs, and can put several groups in the same test so you can compare how each experienced your AI. Testers work with the specific version of the assistant or agent in scope. The report shows the objective, methodology, responses and findings, can be filtered by audience, and the raw data can be exported for your own risk and compliance review.
For a bank or insurer, that means finding the cases where a technically acceptable answer still leaves a customer confused or heading for the wrong decision, before customers find them. Tests can be repeated after a model, knowledge base or vendor change to check whether earlier issues are fixed and whether new ones have appeared.
Adrosaga provides testing evidence that institutions feed into their own governance and assurance processes. Regulatory assessments, compliance decisions and supervisory submissions stay with the financial institution.
After SFF: pick one use case
SFF puts the next phase of financial innovation on show. The harder part comes afterwards: knowing whether those technologies work once they meet real customers, operating environments and business processes.
You will leave Singapore EXPO with plenty of ideas. Take one AI use case home and ask a simple question: could you show today that it works for the customers it serves? If the answer is not a clear yes, book a free 30-minute demo with Adrosaga and we will show you what testing evidence for that use case would look like, with no commitment.
I'm at SFF until the festival closes on Friday. If you would rather talk in person, message me on LinkedIn.
— Joe Nash
Chief Revenue Officer, Adrosaga
AI-assisted. Human-authorised.
Frequently asked questions
What AI announcements were made at SFF 2026?
On 18 November 2026, DPM Gan Kim Yong announced three National AI Missions for Financial Services: strengthening financial crime prevention, improving access to financing and enabling safe agentic transactions. MAS and the Future of Finance Institute also set out supporting initiatives, including an Agentic Money Sandbox for testing how AI agents can transact.
When do the MAS AI Risk Management Guidelines take effect?
The Guidelines take effect on 7 October 2027. Financial institutions may meet Sections 3 and 4, covering AI oversight, AI identification, inventory and risk materiality assessment, from that date, and Sections 5 and 6, covering AI life cycle controls and AI capability and capacity, by 7 October 2028.
Who do the Guidelines apply to?
All financial institutions regulated by MAS, covering all forms of AI, including generative AI and AI agents. They are applied proportionately, according to each institution's size and risk profile and the materiality of each AI use case.
Are financial institutions responsible for AI provided by third-party vendors?
Yes. Financial institutions remain accountable for AI used in the services they deliver, including third-party AI. They should test third-party AI in the context of their own use cases and keep documentation of how it was tested and deemed suitable for use.
Why is human-led AI deployment testing important?
It shows how AI systems perform with representative users in realistic scenarios. It complements automated evaluation by finding issues of comprehension, interpretation, usability and outcomes that technical benchmarks may not capture.
Do the MAS Guidelines cover agentic AI?
Yes. AI agents fall within the Guidelines' general scope. MAS also intends to consult the financial sector in 2027 on what additional guidance on agentic AI would be useful.
Citations
Related News

Our Contributors Are Not a Resource. They Are Our Advantage.
Adrosaga treats contributor experience as a primary driver of AI data quality. Contributors who get clear tasks, fair and prompt pay, accessible support and fair quality review…

Ground Truth Is the New AI Advantage: Why Better Models Need Real Human Data
Ground truth is a trusted reference used to train, test or judge an AI system. As more new web content is AI-generated and some reliable sources restrict crawlers, the advantage…

The AI Trust Gap Is Now the Real Barrier to Deployment
AI is moving fast. Trust is not. Christian Bates, Head of Sales (ANZ/NZ), on why organisations need human-validated evidence before their AI reaches real customers, and what that…


