The MAS FEAT principles ask Singapore financial institutions to make AI-driven decisions fair, ethical, accountable and transparent. Each principle depends on people. Staff who oversee AI agents must be trained to spot unjustified disadvantage, hold agent decisions to the same standard as human ones, handle appeals and requests for review, and explain outcomes to customers in plain terms.
What do the FEAT principles ask of people who oversee agents?
MAS published the principles to promote fairness, ethics, accountability and transparency in the use of artificial intelligence and data analytics in decision-making in Singapore's financial sector. They are principles, not rules, and firms apply them in proportion to the materiality of each use. In substance:
- Fairness. Individuals or groups should not be systematically disadvantaged by AI-driven decisions unless the disadvantage can be justified. Data and models should be reviewed regularly for unintended bias, and the attributes used should be justifiable.
- Ethics. AI-driven decisions should align with the firm's ethical standards, values and codes of conduct, and be held to at least the same ethical standards as decisions made by people.
- Accountability. Use of AI should be approved by an appropriate internal authority, and the firm stays accountable for both internally built and externally sourced systems. Customers should have channels to enquire about, appeal and request review of decisions that affect them.
- Transparency. Firms should proactively disclose their use of AI to customers as part of general communication, and give clear explanations on request about what data was used and how it affected the decision.
Read with agents in mind, the ethics principle sets the training standard. If an agent's decisions must meet the same standard as a person's, the people overseeing it need at least the competence expected of the person who used to make those decisions, plus an understanding of the agent. The application is covered in MAS FEAT principles applied to AI agents.
Who do they apply to?
Financial institutions supervised by MAS that use AI and data analytics in decisions about customers: banks, insurers, capital markets firms, payment service providers and others. Staff in scope include credit, claims and underwriting teams reviewing agent decisions; contact centre staff handling customer enquiries and appeals; fraud and payments operations staff releasing held actions; model and agent owners; and risk and compliance staff reviewing fairness. The short version of how the FEAT principles apply to AI agents is that every decision with customer impact needs a named, competent person somewhere in its chain.
What competence do the principles imply when agents act?
Fairness in practice. Supervisors must recognise when an agent's reasons rest on attributes that are hard to justify, or that act as proxies for protected characteristics. They must read outcome data by customer group and escalate unexplained gaps. MAS FEAT fairness assessment for AI agents covers the assessment itself.
The same ethical standard. A claims handler who would never deny a claim on a technicality should not release an agent's denial that does exactly that. Training must make clear that the agent's draft is held to the handler's standard.
Appeals and reviews. When a customer appeals an agent's decision, the person reviewing it must re-decide with authority and full information. That person needs training on the agent's capabilities, its sources and how it fails.
Explanations. Staff must explain what data was used and how it shaped the outcome, clearly, and in a way they stand behind.
Depth and verdicts. An agent reading data at depth 1 raises little FEAT risk. One drafting decisions at depth 2 puts the weight on the reviewer. One acting at depth 3 within bounds, such as approving routine claims, needs sampling for fairness. Anything that moves money is depth 4 and held. Staff should know which verdict they are looking at: held for them, refused as outside bounds, or withheld at the platform floor.
Provenance. A fact from the core system is sealed; a customer's statement is declared; an emailed document is external. Fair decisions rest on the right evidence.
How do you meet the principles and show it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Fairness | Supervisors trained on proxies; outcomes monitored by group | Fairness reports; escalation records |
| Ethics | Agent decisions reviewed against the same standards as human decisions | Review criteria; sampled decisions |
| Internal accountability | Agent use approved by a named internal authority; owners trained | Approval records; owner training |
| External accountability | Appeal and review channels staffed by trained reviewers | Appeal log with outcomes and reviewer |
| Transparency | Staff trained to explain data used and its effect | Explanation records; QA scores |
| Competent oversight | Reviewers assessed before release rights | Assessment results; access grants |
| Record | Every verdict filed with sources and reviewer | Append-only decision record |
What will MAS or an auditor ask?
- Which customer decisions involve agents, and how material is each?
- Show us fairness monitoring for this agent and the actions it led to.
- How do you hold agent decisions to the same ethical standard as human ones?
- Who handles appeals, and what training did they receive?
- Show us an explanation given to a customer.
- Who approved this agent's use, and do they understand it?
Insurers face these questions often; see how insurers can prove staff are trained to oversee AI agents.
What do the principles not cover?
They are not binding rules, though MAS expects firms to take them seriously. They do not cover technology risk controls, which sit in the MAS Technology Risk Management Guidelines. They do not replace the Personal Data Protection Act 2012, which the PDPC enforces. And they do not prescribe training. IMDA's Model AI Governance Framework offers complementary guidance outside finance. The guide to AI literacy and certification covers the wider model.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Pryme Intelligence Academy's Operator track trains the reviewers who handle held decisions and appeals, with tests set in the learner's own industry and role. Certified Pryme Operator is valid for 12 months, and a public register lists certified people by name, credential and expiry. The argument for certifying staff before go-live is made in the case for certifying payments operations staff.
Questions people ask
Are the FEAT principles mandatory?
They are principles MAS expects firms to apply in proportion to materiality. Supervisors will ask how you applied them.
Does FEAT apply to agents bought from a vendor?
Yes. The accountability principle covers externally sourced systems. The firm remains accountable.
What makes an appeal review meaningful?
A trained person with authority to change the outcome, access to the evidence and the agent's reasons, and no pressure to confirm.