What MAS's 2026 Financial Stability Review tells us about the changing relationship between AI, data, technology and organisational resilience. 

When AI Becomes an Enterprise Dependency

AI usually sits in the growth column of an annual report. It is associated with productivity, automation, innovation and new business models.

MAS's latest Financial Stability Review places it in a more complicated context.

The Monetary Authority of Singapore's 2026 review does not suggest that Singapore's financial system is facing an AI-driven crisis. In fact, MAS's assessment is that Singapore's firms, households and financial institutions remain well buffered against a range of shocks. What is striking is how often AI now appears across different parts of the risk landscape, from cyber and operational risk to third-party concentration and financial market stress.

For anyone responsible for data, technology, analytics and AI, this raises a broader question: how should organisations govern a technology that is becoming embedded in more parts of the enterprise?

AI is showing up across the risk landscape

MAS's 2026 Systemic Risk Survey provides a useful snapshot.

Among 57 financial institutions surveyed, 80% identified cyber and operational risks as a key concern. MAS links these concerns to AI-assisted cyberattacks as well as concentration risks involving critical third-party vendors. Financial market stresses, including AI-driven market fragilities, were cited by 57% of respondents. Geopolitical risks were cited by 77%, while 55% cited macroeconomic uncertainty.

The figures are notable because they show how broadly AI-related concerns are beginning to touch financial institutions.

For financial institutions, AI is becoming relevant to cybersecurity, operational resilience, third-party risk and financial markets at the same time.

That does not mean these risks will materialise, and MAS does not suggest that they inevitably will. The findings indicate that institutions are beginning to account for the ways AI could interact with existing sources of risk as adoption and investment grow.

For organisations outside financial services, there is a useful lesson here. The deeper AI becomes embedded in an enterprise, the harder it becomes to assign AI risk to a single technology or innovation team.

The concentration question

One of the less discussed issues in the review is the potential concentration risk created by common reliance on cloud and AI infrastructure providers. MAS notes that financial institutions are concerned about their shared dependence on these providers.

This is fundamentally an architectural question.

Organisations have traditionally assessed technology vendors individually. Is the vendor secure? Is the service resilient? What happens if the provider experiences an outage?

AI introduces another dimension.

When many organisations depend on the same underlying cloud, model or AI infrastructure, an incident affecting that provider can have consequences beyond a single organisation. The issue becomes one of dependency and concentration as much as vendor risk.

For data and technology leaders, that brings architecture much closer to governance. It raises practical considerations around portability, redundancy, third-party dependencies, data access, resilience and the ability to maintain critical operations if an AI service becomes unavailable.

These are familiar technology management concerns. AI is increasing their significance.

The cyber equation is changing

The same dynamic is visible in cybersecurity.

MAS's review notes that advances in frontier AI could make vulnerability discovery and exploitation easier, while generative AI could enable more sophisticated fraud and social engineering.

At the same time, financial institutions are looking at AI as part of their defensive response.

This creates a complicated feedback loop. Organisations can use AI to strengthen security while attackers use similar capabilities to increase the speed and sophistication of attacks.

The practical challenge is therefore less about deciding whether AI is positive or negative for cybersecurity. It is about how quickly security controls, monitoring and organisational capabilities can adapt as the technology develops.

That becomes particularly important when AI systems gain access to enterprise data, applications and other tools. Security controls designed around human users need to account for systems that can increasingly operate with a degree of autonomy.

What does this mean for AI governance?

This is where MAS's review becomes particularly relevant to the wider enterprise.

AI governance is sometimes treated primarily as a policy exercise: establish principles, define acceptable use, assess models and put human oversight in place.

Those remain important. But the risks identified in the MAS review point towards a broader governance model, one that connects AI governance with data governance, cybersecurity, enterprise architecture, third-party risk and operational resilience.

Singapore's work on agentic AI provides a useful example.

IMDA's updated Model AI Governance Framework for Agentic AI incorporates feedback from more than 60 organisations and addresses risks associated with multi-agent systems, third-party agents and automation bias. The framework emphasises human accountability and recommends measures such as bounding the risks of autonomous actions, implementing technical controls and maintaining meaningful human oversight.

The practical examples are particularly useful.

GovTech Singapore, for instance, took a phased approach to deploying agentic coding assistants. Its initial deployment was limited to internal employees, excluded external tools and focused on lower-risk systems. During that phase, the organisation built controls including central logging, monitoring and approved connections to external tools before expanding the deployment.

This is governance expressed through architecture and operating practice rather than governance existing only as a policy document.

As AI systems gain access to enterprise data and greater ability to act on behalf of users, governance has to travel with the system. Controls need to exist across the data, model, identity, application and infrastructure layers.

That has implications for who owns AI governance as well.

A data office cannot address all of these questions alone. Neither can an AI team, security function or enterprise architecture team. The governance model increasingly has to connect these disciplines.

The AI investment question

The other major AI theme in MAS's review is financial exposure.

MAS conducted a stress test involving a severe downturn in AI-related investment and revenue across the AI supply chain. Under that scenario, around 32% of Singapore-listed firms were assessed as being at risk, representing about 16% of overall corporate debt. The firms most affected were concentrated among highly leveraged, capital-intensive companies and those more reliant on working-capital financing.

Importantly, MAS also found that most firms were able to weather the scenario, supported by earnings and cash reserves, and described corporate balance sheets as generally sound.

The purpose of a stress test is to understand where vulnerabilities could emerge under adverse conditions. It is not a forecast.

For technology and AI leaders, the exercise adds another dimension to the conversation about value. AI investments need to be considered alongside the dependencies and commitments that support them, including infrastructure, vendors, data, financing and business processes.

The question of AI value therefore sits within a wider question of organisational resilience.

A longer-term data security question

Quantum computing receives considerably less attention in the public conversation, but MAS's treatment of it is worth noting.

The concern is that advances in quantum decryption could eventually undermine cryptographic systems currently used to protect sensitive information. This includes the possibility of data being stolen today and decrypted in the future.

MAS has been working with the financial industry on quantum-resilient cryptography and is working towards full quantum resilience for the financial sector before 2030.

For data leaders, this introduces a different dimension to data governance.

Protecting information is partly a question of time. Data that needs to remain confidential for ten, twenty or thirty years may need protection against technologies that do not yet exist at scale.

That makes cryptographic resilience part of the broader data lifecycle conversation.

The governance challenge is becoming more connected

What I find most interesting about MAS's review is the way several traditionally separate conversations begin to converge around AI.

AI connects to cybersecurity because attackers can use it to increase the speed and sophistication of attacks.

It connects to architecture because organisations increasingly depend on shared cloud, model and AI infrastructure.

It connects to data governance because AI systems require access to information and, increasingly, the ability to act on that information.

It connects to operational resilience because organisations may become dependent on AI-enabled systems for critical processes.

And it connects to financial risk because the scale of AI investment has become large enough to influence capital markets and corporate balance sheets.

For organisations adopting AI, this does not mean putting the brakes on innovation. It does mean thinking carefully about the dependencies being created alongside the benefits.

The governance question is therefore becoming much more practical.

Where does AI governance sit? Who has authority over the data an AI system can access? What controls determine what it can do? How dependent is the organisation on a particular provider? What happens when the model, service or underlying infrastructure becomes unavailable? And how are these risks incorporated into existing enterprise risk and resilience frameworks?

These questions sit across the CDAO, CIO, CTO, CISO, risk and business functions.

That may be the most useful takeaway from MAS's latest review.

AI is becoming part of the enterprise's operating environment, and its risk profile is becoming intertwined with the systems, data, infrastructure and decisions that organisations already depend on.

For data and AI leaders, the challenge is to make sure governance evolves at the same pace.

The question is not simply how much AI an organisation can deploy. It is whether its data, architecture, governance and operating models are ready for the dependencies that come with it.

Join us at CDAIO, Enterprise AI, and Data & AI Architecture Singapore 2027, 13-14 April to learn more from data analytics and AI leaders. Reach out to Eleen Meleng for more information.