Artificial intelligence in financial services is entering a new phase. The conversation has moved beyond productivity gains and chatbot pilots toward a more fundamental question:
How can organisations safely govern AI systems that act autonomously?
In Singapore, regulators are already addressing this challenge.
The recent release of the Model AI Governance Framework for Agentic AI by the Infocomm Media Development Authority (IMDA) announced at the World Economic Forum in January 2026 and the world’s first governance framework designed specifically for agentic AI and the AI Risk Management Toolkit by the Monetary Authority of Singapore (MAS), developed collaboratively with 24 leading financial institutions under Project MindForge, signals a clear shift in regulatory expectations.
Organisations are expected to move beyond managing AI models to governing AI agents capable of reasoning, making decisions, and executing actions across business systems. For CIOs, CFOs, and finance leaders, this represents both an opportunity and an urgent governance challenge.
Why Agentic AI Changes the Governance Equation
Traditional AI systems analyse information and generate recommendations. Agentic AI goes further.
AI agents can access enterprise data, trigger workflows, interact with applications, and complete multi-step tasks with limited human intervention. This creates significant opportunities for finance teams, including accelerating accounts payable and receivable processes, automating reconciliation and exception handling, streamlining vendor onboarding and compliance checks, supporting financial planning and analysis, and assisting with regulatory reporting and audit preparation.
However, greater autonomy introduces new risks. Finance leaders must answer critical questions:
- Which systems can an AI agent access?
- What actions can it perform autonomously?
- How are decisions monitored and audited?
- When should humans remain in the approval loop?
- How can organisations ensure compliance with internal controls and regulatory expectations?
Agentic AI shifts risk from “wrong answers” to “wrong actions.” These are precisely the challenges Singapore’s latest frameworks aim to address.
What Singapore’s AI Frameworks Mean for Financial Institutions
IMDA’s Agentic AI Framework: Four Dimensions of Governance
The MGF organises its requirements across four dimensions that every deploying organisation must address.
Assess and bound risks upfront. Before an AI agent is deployed, organisations must conduct use-case-specific risk assessments, evaluating autonomy level, access to sensitive data, task complexity, and the reversibility of actions the agent can take. In finance, where agents may touch payment systems, reconciliation workflows, or vendor databases, this is non-negotiable. The framework recommends limiting early deployment by user group, tool access, or system exposure, then expanding gradually.
Make humans meaningfully accountable. This is the dimension most organisations misread. IMDA doesn’t simply ask you to put a human in the loop, it asks you to prove the human is genuinely reviewing, not rubber-stamping. The framework explicitly flags automation bias: a reviewer approving hundreds of AI-generated assessments in seconds is not exercising meaningful oversight. Clear accountability allocation, escalation criteria, defined override rights, and measurable human override rates are what actual compliance looks like.
Implement technical controls across the full lifecycle. From development through pre-deployment testing to ongoing monitoring, the framework requires structural controls, rule-based guardrails, least-privilege access, and audit trails demonstrating the system is performing as intended. This covers the full agent lifecycle, not just the point of deployment.
Enable end-user responsibility. Finance teams using AI agents must understand what those agents can and cannot do, who to escalate to when something goes wrong, and how to preserve the human skills the agent is augmenting. Training and transparency are governance requirements, not nice-to-haves.
Critically, the MGF makes clear that organisations remain accountable for the actions of AI agents regardless of the level of autonomy involved. Delegating execution to an agent does not delegate accountability.
MAS Guidelines and MindForge: The Financial Institution Lens
MAS goes further for regulated entities. Its proposed AI Risk Management Guidelines, built on the FEAT principles (Fairness, Ethics, Accountability, Transparency) and informed by the MindForge Toolkit developed with Standard Chartered, Income Insurance, Julius Baer, and others, establish expectations that extend to every AI tool a financial institution uses, including third-party systems.
The guidelines apply to all MAS-regulated financial institutions and set expectations across governance structure, AI risk management systems, lifecycle controls, and operational capability. Board-level AI risk oversight is an explicit requirement, not a technical-team function.
Together, these frameworks reinforce one message: successful AI adoption requires governance by design, not governance as an afterthought.
The Governance Challenge Facing Finance Teams
Finance functions operate in highly regulated environments with established controls, approval workflows, and audit requirements.
Introducing AI agents without appropriate guardrails can create concerns around unauthorised financial transactions, data privacy and confidentiality risks, inaccurate or non-compliant outputs, limited visibility into AI-driven decisions, and difficulties demonstrating accountability during audits.
As organisations move from experimentation to production deployment, these concerns become board-level issues.
The challenge is not whether finance teams should adopt agentic AI. The challenge is how to deploy it responsibly at scale, with the audit trail, access controls, and human oversight mechanisms that regulators now expect to see.
In our work advising financial institutions across Asia on AI governance readiness, the most common gap we encounter is not ambition — it is architecture. Organisations have policies on paper but lack the technical infrastructure to enforce them at the point of execution. The governance layer has to be built into how AI agents operate, not added as a compliance layer afterward.
Operationalising AI Governance with Workato
This is where AI orchestration becomes the clear differentiator.
Rather than allowing AI agents to interact directly with enterprise systems, organisations need a governed execution layer that embeds security, oversight, and policy enforcement into every workflow. Workato, the global leader in agentic orchestration, provides this governance layer through its Enterprise Model Context Protocol (MCP) capabilities and AI agent framework, enabling organisations to operationalise the governance principles outlined by IMDA and MAS.
1. Controlled Access to Enterprise Systems
AI agents should only access the data and applications necessary to complete a specific task. Workato’s Enterprise MCP capabilities provide secure, governed access to enterprise applications, APIs, and business data while enforcing predefined permissions and policies. Scoped tokens, environment isolation, and identity-aware routing ensure the right agents reach the right systems with the right level of access, and no more.
This directly supports IMDA’s least-privilege requirement and MAS’s expectation of defined access controls across the AI lifecycle.
2. Human-in-the-Loop Controls
Not every decision should be fully autonomous. Finance teams can configure approval workflows that require human intervention for high-risk activities, payment releases, policy exceptions, vendor onboarding approvals. For high-stakes decisions, approvers can be required to provide written justification before proceeding, creating the documented oversight trail that IMDA’s framework explicitly calls for.
Override rates can be tracked over time, providing evidence that human review is meaningful rather than nominal.
3. End-to-End Auditability
Regulators increasingly expect organisations to demonstrate how AI systems arrive at decisions and what actions they take. Workato captures workflow execution details, decision points, approvals, and system interactions to create comprehensive audit trails, giving finance leaders the evidence base needed for internal controls, regulatory reporting, and audit preparation.
Workato’s platform architecture includes full traceability by default across MCP-deployed workflows. This is a meaningful differentiator for organisations needing to evidence compliance.
4. Governance by Design, Not Afterthought
With Workato, organisations can define the rules that govern approved data sources, permitted actions, escalation thresholds, approval requirements, and exception handling procedures before agents are deployed into production workflows.
This ensures AI agents operate within established business policies rather than creating parallel processes outside existing governance frameworks. IT teams manage and govern every agent from a single console, with consistent controls applied across agents regardless of which AI provider powers them.
5. Continuous Monitoring and Risk Management
Agentic AI requires ongoing oversight, not a one-time deployment review. Workato provides visibility into workflow performance, agent activities, and exceptions, enabling organisations to identify anomalies, trigger risk-based interventions, and continuously improve governance controls aligned with MAS’s emphasis on lifecycle risk management and IMDA’s requirement for post-deployment monitoring.
A Practical Example: AI-Powered Accounts Payable
Consider an AI agent supporting invoice processing.
The agent can retrieve invoices from multiple channels, validate supplier information, match invoices against purchase orders, identify exceptions, route approvals to relevant stakeholders, and trigger payment workflows.
However, governance controls ensure the agent accesses only authorised financial data, escalates exceptions for human review, complies with segregation-of-duty policies, maintains a complete audit trail, and operates within predefined spending thresholds.
This is the difference between automation and governed automation and it is precisely what both IMDA and MAS are asking organisations to demonstrate.
The Next Competitive Advantage: Trusted AI
Singapore’s regulatory direction is clear. The future of AI adoption will not be defined by which organisations deploy agents first, but by which organisations deploy them responsibly.
While IMDA’s framework is voluntary today, it is already functioning as a procurement and audit baseline. Organisations are expected to defend their governance posture in risk reviews, vendor assessments, and board discussions now, not when formal regulation arrives.
For CIOs and finance leaders, the priority should not simply be introducing AI into existing processes. It should be establishing the governance foundations that enable AI to scale safely, securely, and compliantly.
Organisations that embed governance into their AI architecture today will be better positioned to unlock the full value of agentic AI tomorrow.
The question is no longer whether your finance function will adopt AI agents. The question is whether your governance model is ready for them — and whether it is built to last.
