MAS FEAT principles for agentic AI: a practical implementation guide

Singapore’s MAS FEAT principles — Fairness, Ethics, Accountability, Transparency — are the regulatory anchor for AI in financial services in this jurisdiction. Most published guidance treats them abstractly. This piece is the opposite: it’s the architectural checklist I use with banking clients to translate FEAT into concrete agent design.

Why FEAT, and why now?

FEAT predates the current wave of generative AI. It was published in 2018 for traditional ML in finance. What’s new is that agentic AI introduces multi-step reasoning, autonomous tool calling, and emergent behaviours — and FEAT now applies to all of them.

MAS has been increasingly explicit: if you’re deploying any AI that materially affects customers, employees, or markets, FEAT applies. The grace period for “we’re still figuring it out” is closing.

Fairness — concrete implementation

Fairness in agentic AI is not just about training data. It’s about whether the agent’s decisions — across the population of cases it handles — produce equitable outcomes.

For a banking agent doing customer due diligence refresh, this means:

Disparate impact monitoring. Track agent decision outcomes by customer segment (where legally collected). Flag statistically significant disparities for human review.
Calibrated confidence by segment. The agent’s confidence scores must be reliable across all customer types — not just the majority.
Equal access to review. When the agent escalates, the escalation criteria must be the same across segments. A common failure: agents escalate more aggressively for minority cases, creating downstream review bottlenecks.

Ethics — beyond box-checking

Ethics is the hardest principle to translate into engineering because it’s value-loaded. My working definition for banking agents:

No deception. Agents never represent themselves as humans in customer-facing contexts.
No dark patterns. Agents that interact with customers cannot exploit cognitive biases for institutional benefit.
Refusal envelope. Every agent has explicit instructions on what it will not do, regardless of the prompt. For banking: no advice on regulatory evasion, no fabrication of evidence, no decisions outside delegated authority.

The refusal envelope is the single most underrated control. It’s the agent equivalent of the limit framework for traders.

Accountability — the audit trail

This is where most agentic AI deployments fail their first internal audit. Accountability requires:

Decision logging. Every agent decision logged with: timestamp, model version, prompt version, input data, retrieved context, reasoning chain (where exposed), tools called, output, confidence.
Prompt versioning. Every prompt change tracked like code. No “we fixed it in production” without a versioned commit.
Model versioning. When the underlying model is updated (e.g., Claude Sonnet → Claude Opus), the deployment is treated as a model change requiring re-validation.
Human-in-the-loop checkpoints. For decisions above defined materiality thresholds, a human signs off — with the signature logged and immutable.

Build this from the start. Bolting it on later is two-thirds of a rebuild.

Transparency — the dual interface

Transparency in agentic AI has two audiences:

For the regulator and internal auditor: They need to see the system’s decision criteria, performance metrics, and exception patterns. This is documentation: model cards, system cards, evaluation reports, drift monitoring.

For the customer: They need to know that AI was used in a decision affecting them, what data was considered, and how to challenge the outcome. This is consent, disclosure, and recourse.

The two cannot be conflated. A bank that publishes its internal evaluation report verbatim to customers will create confusion; a bank that hides its evaluation from regulators will create a finding.

The FEAT readiness checklist

For each agentic AI deployment, I run this checklist:

Fairness

Disparate impact tested across customer segments
Confidence calibration by segment validated
Escalation thresholds equal across segments

Ethics

Refusal envelope defined and tested
No deceptive representation to customers
Cognitive bias review by independent reviewer

Accountability

Decision logging architecture in place
Prompt versioning in source control
Model change management process documented
HITL checkpoints defined by materiality

Transparency

Model card and system card published internally
Customer disclosure (where applicable) reviewed by legal
Performance and drift monitoring operational
Challenge/recourse process for affected customers

A deployment that ticks all 14 boxes is auditable. A deployment that ticks 7 is a finding waiting to happen.

The Singapore advantage

Singapore is one of the few jurisdictions where regulators have produced practical, implementable AI governance guidance. The combination of MAS FEAT, IMDA’s Model AI Governance Framework, AI Verify, and PDPC’s PDPA guidance gives Singapore-based deployments a clearer roadmap than competitors operating in fragmented regulatory environments.

For banks deploying agentic AI: the path is laid. The cost of compliance is mostly engineering discipline, not legal interpretation. The window to be early and compliant is now.

If you’re building or scoping an agentic AI system in a MAS-regulated environment, we’d be happy to share our FEAT readiness assessment template. Get in touch.