They become four tests every agent capability should pass. MAS's FEAT principles ask financial institutions using AI and data analytics in decisions to be fair, ethical, accountable and transparent. They are principles, not prescriptive rules, so the firm decides how to meet them. For an agent that acts, accountability and transparency weigh most, because the action reaches the customer directly.
What do the FEAT principles ask, in substance?
MAS published FEAT to guide the responsible use of AI and data analytics in financial products and services in Singapore. In substance:
- Fairness. Decisions should not systematically disadvantage individuals or groups without justification. Data and models should be reviewed for unintended bias, and the use of personal attributes justified.
- Ethics. AI-driven decisions should meet the firm's ethical standards and be held to at least the same standard as decisions made by people.
- Accountability. Use of AI should be approved by an appropriate internal authority, and firms remain accountable for it. Customers should have channels to enquire about, appeal and request review of decisions that affect them.
- Transparency. Firms should be open about their use of AI where appropriate, and able to give clear explanations of decisions to affected customers on request.
FEAT sits alongside MAS's Technology Risk Management Guidelines, which cover how systems are built, run and secured, and the Personal Data Protection Act 2012, which governs personal data. A principle-by-principle reading is in MAS FEAT principles applied to AI agents.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What changes when the AI system is an agent?
FEAT was framed around decision models: a score, a classification, a recommendation. An agent makes chains of small decisions and then acts. Each principle shifts accordingly.
Fairness must be tested on outcomes, not only on the model's inputs, because an agent can treat customers differently through which questions it asks or which cases it escalates. Ethics extends to conduct: an agent that is technically accurate but pushes customers towards a product they do not need fails the test. Accountability needs a named internal authority for each capability, not a single sign-off for "the agent". Transparency requires that you can reconstruct why the agent acted, from the sources it read to the action it took.
Because FEAT is principles-based, the design work falls to you, an argument made in principles-based regulation leaves the agent design to you.
Concrete examples: what may the agent do, and what stays with a person?
| Use | Agent | Person | FEAT concern |
|---|---|---|---|
| Answer questions about card fees | Yes | Complaints | Transparency of fees |
| Suggest an investment product | Explains features; no recommendation without suitability | Adviser makes the recommendation | Ethics, fairness |
| Assess an SME loan application | Prepares the assessment | Credit officer decides | Fairness, accountability |
| Raise a credit card limit | Within policy and limits | Exceptions | Fairness across groups |
| Respond to a request to review a decision | Gathers the record | Reviews and decides | Accountability |
An illustrative example
Take a Singapore bank whose agent assesses applications for small business loans. Its fairness review compares outcomes across business types and finds the agent refers sole proprietors to manual review far more often than companies with similar finances. The cause is that sole proprietors' records are less structured, and the agent treats missing structure as risk.
Under FEAT, the difference needs a justification, and "the records were harder to read" is not one. The bank changes the agent to request specific missing items instead of referring the whole case, re-runs its fairness assessment on past applications, and records the result. The approach is set out in MAS FEAT fairness assessment for AI agents.
Which controls show FEAT in practice?
- Approval by a named internal authority for each agent capability, recorded.
- Outcome testing across customer groups before use and on a schedule after.
- A full decision record: sources, reasoning and action, retrievable for any customer.
- Customer channels to ask about, appeal and request review of agent decisions, handled by people.
- Disclosure that an agent is used, written in plain language.
The cross-regulation view is in AI agent regulation, mapped to controls.
Questions people ask
Are FEAT principles binding?
They are principles issued by MAS for financial institutions to apply. They are not a prescriptive rulebook, but MAS expects firms using AI in decisions to be able to show how they meet them.
Do FEAT principles apply to internal agents?
They focus on AI used in decisions about financial products and services, so customer-affecting uses matter most. Internal agents still fall under MAS's technology risk expectations.
How detailed must an explanation to a customer be?
Clear enough for the customer to understand the main reasons for the decision and how to challenge it. That requires the firm to have recorded those reasons when the agent acted.