MAS FEAT's accountability principle asks two things. Internally, each use of AI must be approved by an appropriate internal authority, and the firm stays accountable whether it built or bought the system. Externally, affected people must be able to ask questions, appeal and seek review. For agents, that means named owners, named approvers and a working appeal route.
What does the accountability principle say?
MAS published the FEAT principles, covering fairness, ethics, accountability and transparency, to guide financial institutions using AI and data analytics in decision-making. They are principles rather than prescriptive rules, and they sit alongside MAS's Technology Risk Management Guidelines, the Personal Data Protection Act 2012 and IMDA's Model AI Governance Framework. Our overview of how the FEAT principles apply to agents covers all four.
The accountability principle has an internal and an external side. In substance:
Internal accountability - Use of AI and data analytics in decision-making is approved by an appropriate internal authority. - Firms are accountable for both internally developed and externally sourced models. - Firms proactively raise management and board awareness of their use of AI.
External accountability - People affected by AI-driven decisions are given channels to enquire about, appeal and request review of those decisions. - Verified and relevant supplementary information from them is taken into account in the review.
Who does it apply to?
Financial institutions in Singapore that use AI and data analytics in decisions affecting customers or the firm: banks, insurers, capital markets firms and payment service providers. The principles are written broadly enough to cover any AI that shapes decisions, which plainly includes agents.
What does accountability mean for an agent that acts?
The principle was written around models that inform decisions. Agents go further: they make or carry out decisions inside processes. That sharpens each element.
Approval by an appropriate internal authority, per capability. Approving "the customer service agent" is too coarse. An agent that can answer questions, waive fees and close accounts carries three different risk levels. The internal authority should approve each capability and its limits. A fee waiver up to a small amount might need a team head's approval; account closure might need a committee's.
Accountability for externally sourced agents. Many firms buy agents from vendors. FEAT is clear that this does not move accountability. The firm must understand and stand behind what the vendor's agent does, which means testing it on the firm's own cases and keeping the ability to limit what it may do.
Board awareness. The board should know which agents run, what they may do, and where they touch customer outcomes. A list of AI projects is not enough; the board needs to see authority.
Appeal channels for agent decisions. When an agent declines a claim, holds a payment or closes a ticket, the customer must be able to ask why and seek a review by a person. The review should consider information the customer provides, not just re-run the agent.
A concrete case: an agent maps the ownership of a corporate customer and flags that it cannot confirm a beneficial owner, so onboarding stalls. Our answer on whether an agent can identify ultimate beneficial owners explains its role. Under FEAT, a named person owns that outcome, the customer can query it, and documents the customer supplies are considered on review.
How do you meet the principle with named owners and approvers?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Approval by appropriate internal authority | Each agent capability approved at a level matching its risk | Approval records per capability |
| Named accountability | Each agent has a named business owner | Ownership register |
| Accountability for bought agents | Vendor agents tested on the firm's cases and bounded by the firm | Test results; bound configuration |
| Board awareness | Board receives regular reporting on agents, their authority and outcomes | Board papers and minutes |
| Enquiry channel | Customers can ask about decisions an agent made or influenced | Enquiry logs |
| Appeal and review | Customers can request review by a person | Review records with outcomes |
| Supplementary data considered | Reviewers consider verified information from the customer | Review notes citing new information |
Fairness and accountability support each other. Our piece on MAS FEAT fairness assessment for agents covers how to test that agent decisions do not disadvantage groups without justification, and accountability is how findings get acted on.
What will MAS or an auditor ask?
- Who approved each capability of this agent, and at what level?
- Who owns this agent, and what do they review?
- For a vendor's agent, how did you satisfy yourselves it works for your customers?
- What does the board know about agents and their authority?
- How can a customer appeal a decision an agent made?
- Show us a review where the customer's new information changed the outcome.
What does the principle not cover?
It does not set technical controls for agents; the Technology Risk Management Guidelines address technology risk. It does not govern personal data handling, which falls under the Personal Data Protection Act 2012. It does not assign individual legal liability. And it is one framework among several; compare NIST AI RMF accountability roles, or read what AI accountability is for the underlying idea. Our guide agent governance is a bound, not a model explains why accountability needs enforced limits.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does FEAT require a person to approve every agent decision?
No. It requires that the use of AI is approved by an appropriate authority and that customers can seek review. Routine actions can run within approved limits, with higher-impact ones held for a person.
Who is the appropriate internal authority?
It depends on the materiality of the decision. Many firms set tiers: a business head for low-impact uses, a risk committee for material ones, and the board for the most significant.
Does accountability apply if the agent only recommends?
Yes. If its recommendation shapes a decision about a customer, the firm is accountable for that influence, and the customer should be able to query the outcome.