A FEAT fairness assessment asks whether an agent's decisions systematically disadvantage any person or group without justification. For an agent, you define the decision it makes or supports, identify the groups and personal attributes at stake, test outcomes on real cases, justify or fix any differences, and record the result with a date to repeat it.
What do the FEAT fairness principles require?
MAS's FEAT principles cover fairness, ethics, accountability and transparency in the use of AI and data analytics in financial services. The fairness principles, in substance:
- Individuals or groups should not be systematically disadvantaged through AI-driven decisions unless the decisions can be justified.
- The use of personal attributes as inputs to AI-driven decisions should be justified.
- Data and models should be regularly reviewed and validated for accuracy and relevance, and to minimise unintentional bias.
- AI-driven decisions should be regularly reviewed so that models behave as designed and intended.
They are principles, not a rulebook. MAS expects financial institutions to apply them in proportion to the materiality of the decision. A broader introduction is in MAS FEAT principles applied to AI agents.
Who does it apply to?
Financial institutions in Singapore using AI and data analytics in decisions about customers, and others that adopt FEAT voluntarily. The assessment matters most where an agent makes or shapes decisions on credit, insurance underwriting and claims, pricing, fraud and account restrictions, and collections.
What makes fairness assessment different for agents?
Scoring models are assessed by testing a formula's outputs across groups. Agents need a wider lens for three reasons.
Agents decide in language. An agent drafting claim decisions may treat two similar claims differently because of how each customer wrote. Customers who write briefly, in a second language or with poor spelling can be disadvantaged without any protected attribute being used.
Agents choose their own evidence. If an agent retrieves more supporting documents for some customers than others, outcomes may differ for reasons nobody designed. Fairness depends on what it looked at, not only on how it judged.
Agents act at several points. Unfairness can appear in who gets escalated to a person, who gets a fast answer, or who gets offered a payment plan, not only in the final decision.
The accountability side of this, who owns the result, is covered in MAS FEAT accountability and AI agent owners.
How do you run the assessment?
- Define the decision. Which capability, what outcome, who is affected, and what a good outcome looks like.
- Identify groups and attributes. Which personal attributes the agent can see or infer, which groups could be disadvantaged, and which proxies might stand in for protected characteristics.
- Justify attributes used. For each personal attribute the agent relies on, record why it is relevant and legitimate.
- Choose fairness measures. Decide how to compare outcomes: approval rates, error rates, escalation rates, time to resolution, all by group.
- Test on real cases. Replay the agent on a representative set of past cases from your operation and measure by group.
- Explain differences. Separate differences that reflect legitimate factors from those that do not.
- Mitigate. Change sources, instructions or limits; route certain cases to people; remove the capability if needed.
- Record and approve. Document the assessment, the decision and the owner's sign-off.
- Review. Repeat on change and on a schedule, and monitor live outcomes in between.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| No unjustified systematic disadvantage | Outcome testing by group on real cases before and after release | Test results by group |
| Justified use of personal attributes | Register of attributes visible to the agent, with justification | Attribute register |
| Proxies controlled | Review of inputs and writing-style effects | Proxy analysis |
| Data and models reviewed | Sources and model changes trigger retesting | Change records; retests |
| Decisions reviewed in operation | Live monitoring of outcomes by group with thresholds | Monitoring reports |
| Proportionality | Depth of assessment set by decision materiality | Materiality rating |
| Documentation | Assessment recorded with owner sign-off | Signed assessment |
What will MAS or an internal auditor ask?
- Which agents make or support decisions about customers, and how material are those decisions?
- Which personal attributes can each agent see, and why?
- How did you test for unfair outcomes, on what cases, and what did you find?
- What did you change as a result?
- How are outcomes monitored after release?
- Who signed off, and when is the next review?
What does the fairness assessment not cover?
It does not settle data protection duties; a data protection impact assessment may also be needed, as discussed in do AI agents need a data protection impact assessment. It does not define a single fairness metric; you choose and justify. It is not a fundamental rights assessment under EU law, although the method overlaps with how to run a fundamental rights impact assessment for an AI agent. For a short answer on scope, see how MAS FEAT principles apply to AI agents, and for the full map, AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Fairness testing needs real cases and repeatable runs. Pryme's Attest replays each agent capability against real cases from your own operation and scores every answer against a floor, so the same case set can be cut by group and rerun after each change. Its stable gate checks that the agent gives the same answer to the same case across runs.
Questions people ask
Do we need demographic data to test fairness?
Often you lack it, and collecting it raises its own issues. Use what you lawfully hold, test for proxy effects such as language or region, and record the limits of the assessment plainly.
Is a difference in outcomes always unfair?
No. Differences that follow from legitimate, relevant factors, such as ability to repay, can be justified. The assessment must show the reasoning.
How often should the assessment be repeated?
On any material change to the agent and on a regular cycle set by materiality. Live monitoring between assessments catches drift sooner.