Docs

Appearance

Principles ​

FEAT succeeds when it becomes operational: controls that teams execute and evidence that auditors can follow.

The FEAT pillars ​

  • Fairness: decisions should not systematically disadvantage groups without justification
  • Ethics: AIDA outcomes should align to ethical standards (not lower than human decisions)
  • Accountability: roles, approvals, and recourse exist and are used
  • Transparency: disclose AIDA use and provide meaningful explanations when requested

The 14 FEAT principles (grouped) ​

FEAT provides a set of principles that teams typically implement as a combination of policy, controls, and monitoring signals:

Fairness (4 principles) ​

  • avoid systematic disadvantage unless justified by legitimate factors
  • ensure use of personal attributes is justified and appropriate
  • regularly review and validate data/models for accuracy, relevance, and bias risks
  • regularly review AIDA-driven decisions to ensure models behave as intended

Ethics (2 principles) ​

  • align AIDA with the firm’s ethical standards, values, and codes of conduct
  • ensure AIDA-driven decisions meet at least the same ethical standards as human decisions

Accountability (5 principles) ​

  • require appropriate internal approval for AIDA use (commensurate with materiality)
  • stay accountable for both internally developed and externally sourced models
  • proactively raise management/board awareness of material AIDA risks
  • provide channels for enquiry, appeal, and review of AIDA-driven decisions
  • consider verified supplementary data when reviewing decisions (where appropriate)

Transparency (3 principles) ​

  • proactively disclose AIDA use (as appropriate)
  • provide, on request, clear explanations about what data is used and how it affects decisions
  • provide, on request, information about consequences of AIDA-driven decisions

How to implement FEAT (control themes) ​

Fairness: define, measure, and remediate ​

Implementation themes:

  • define fairness metrics and thresholds per use case (not “one metric fits all”)
  • evaluate on a cadence and on material changes (data, model, population)
  • remediate systematically (root cause → mitigation → re-test)

Ethics: define unacceptable outcomes ​

Implementation themes:

  • define unacceptable outcomes and misuse scenarios
  • require escalation and review for edge cases and high-impact changes
  • document tradeoffs explicitly (performance vs harm reduction)

Accountability: approvals and recourse ​

Implementation themes:

  • assign owners and reviewers with decision rights
  • implement approval gates for launch and major changes
  • operationalize recourse and appeals (workflow, SLAs, audit trail)

Transparency: disclosures and explanations ​

Implementation themes:

  • publish appropriate disclosures of AIDA usage
  • provide explanation patterns appropriate to the decision context
  • document limitations and “safe use” guidance for operators and customers

What evidence is typically expected ​

Evidence is strongest when it is linked to controls and kept current:

  • fairness evaluation methodology + historical results
  • decision thresholds and approval records
  • user guidance/disclosures + versions
  • incident/complaint handling records and remediation outcomes
Operationalizing in Modulos

Turn FEAT principles into controls, evidence, and testing signals

Risk treatment

Track mitigations, ownership, and residual risk acceptance

Testing

Schedule evaluations and retain history

Disclaimer

This page is for general informational purposes and does not constitute legal advice.