MAS FEAT and the Agentic AI Compliance Gap

The Monetary Authority of Singapore published its FEAT Principles — Fairness, Ethics, Accountability, and Transparency — in 2018 to govern how financial institutions use artificial intelligence. In 2018, that meant algorithmic credit scoring. Automated recommendations. Fraud detection models that flagged transactions for human review.

Human beings still initiated every transaction. AI informed the decision. Humans pressed the button.

That assumption no longer holds.

Globally recognised payment networks and major financial institutions have begun deploying AI agent capabilities that allow autonomous execution of financial transactions — payments, settlements, FX — on behalf of customers and institutions. The instruction, the authorisation, and the execution now happen at machine speed. No human presses the button.

The question facing every MAS-regulated institution in 2026 is whether their compliance infrastructure was designed for this world. Under MAS FEAT, the answer — for most — is no.

What MAS FEAT Actually Requires

The Accountability section of the MAS FEAT Principles requires institutions deploying AI in financial decision-making to demonstrate three things: AI-driven decisions are made for identifiable, auditable reasons; the institution can account for every outcome its AI systems produce; and human oversight mechanisms exist for non-routine decisions.

The Transparency section requires that AI-driven decisions be explainable to regulators and relevant parties on request.

For an AI agent executing a payment, this translates directly into a compliance obligation: every AI-initiated transaction must be traceable to an explainable compliance decision, backed by a complete audit record. Not at the end of a monthly review cycle. For every individual transaction, in a format that can be exported and examined by MAS at any time.

The Architecture Problem

Standard compliance infrastructure was not built for this.

Transaction monitoring systems, AML/CFT screening engines, and fraud detection platforms share one structural characteristic: they operate post-execution. A transaction enters the system, and the compliance check happens after.

For human-initiated transactions at human speed, this is acceptable. A flagged transaction can be investigated. Regulators can be notified. The process works.

For AI-initiated transactions at machine speed, post-execution is too late. An AI agent executes a payment instruction in milliseconds. By the time a post-execution system identifies a concern, the funds have moved. The compliance failure has already occurred.

Post-execution monitoring can detect a problem and trigger a reporting obligation. It cannot produce an explainable, pre-decision compliance record for a transaction that has already executed — which is precisely what the Accountability and Transparency sections of MAS FEAT envision.

The gap between MAS FEAT's explainability expectations and what post-execution monitoring delivers is structural. It cannot be closed by tuning existing systems. It requires a different compliance layer entirely.

What Pre-Execution Compliance Looks Like

A pre-execution compliance layer sits between the AI agent's instruction and the payment rail.

When the AI generates a transaction instruction, the instruction is intercepted before it reaches the institution's payment infrastructure. A compliance assessment runs — checking the instruction against the institution's own policy configuration, MAS AML/CFT thresholds, counterparty screening lists, and risk appetite parameters.

The assessment produces a binary decision: cleared and released to the payment rail, or held and queued for compliance team review.

Every decision — cleared or held — is written to an immutable audit log. The log records the instruction, the assessment parameters, the decision, and the timestamp. It is tamper-proof, exportable on demand, and retained for the period required under MAS AML/CFT Notice obligations.

When MAS examiners ask why the institution's AI executed a particular transaction, the institution can produce a complete, explainable record — directly supporting the Accountability and Transparency expectations under FEAT.

The Obligation Is Already Live

MAS FEAT does not contain a grace period for AI adoption.

The moment an institution's AI system executes its first financial transaction autonomously, the obligation to demonstrate explainability and accountability for that transaction exists. It does not begin when the compliance infrastructure is ready. It begins with the first transaction.

Institutions that wait for an enforcement event to motivate action will not be early adopters building a record of due diligence. They will be respondents.

The window to implement pre-execution compliance infrastructure proactively — before the first examination question, before the first regulatory inquiry — is open. It will not remain open indefinitely.

This article is for informational purposes only and does not constitute legal or regulatory advice. Institutions should assess their specific obligations under MAS FEAT, MAS AML/CFT Notice, and MAS TRM Guidelines with their own legal and compliance counsel. Sidian Pte. Ltd. provides pre-execution AI compliance infrastructure for MAS-regulated institutions in Singapore.