Nematix
Generative Ai
 MAS FEAT Principles for GenAI in Financial Services
Feb 15, 2025

MAS FEAT Principles for GenAI in Financial Services

MAS FEAT predates large language models. Here is how it applies to GenAI in financial services and what regulated firms need to demonstrate.

The Monetary Authority of Singapore published the FEAT Principles — Fairness, Ethics, Accountability, Transparency — in November 2018. They were designed to set expectations for the use of AI and data analytics in financial services, and they have done that job well. Singapore’s financial sector has used FEAT as a governance anchor for AI adoption for more than five years. By influence, the framework has shaped how regulators and compliance teams across Southeast Asia think about responsible AI.

The framework predates large language models by several years. It was written in the context of traditional machine learning: gradient-boosted trees for credit scoring, logistic regression for fraud detection, recommendation engines for product cross-sell. GenAI introduces a set of capabilities and failure modes that FEAT’s drafters did not have in front of them when they wrote it.

That does not make FEAT inapplicable to GenAI. FEAT was deliberately written as a principles-based framework — technology-agnostic by design. Each of the four principles applies to GenAI deployments, in some cases with greater urgency than they applied to the traditional ML models they were originally written to govern.

F — Fairness

The fairness principle requires that AI-assisted decisions do not result in discriminatory outcomes — that protected characteristics such as race, gender, age, and religion do not, directly or through proxies, determine who receives credit, what premium they are quoted, or how they are treated by a financial institution.

For traditional ML models, fairness assessment is technically tractable. You can inspect the features the model uses, evaluate its output across demographic groups, and quantify disparate impact through established metrics. Regulators and firms have years of practice doing this for credit scoring models.

For LLMs, the picture is substantially more complex. Foundation models are trained on internet-scale text corpora. The composition of those corpora is not publicly disclosed in detail, and the demographic biases embedded in that data are not fully characterised. When an LLM is used in a customer-facing application — generating personalised product recommendations, drafting credit decision rationale letters, responding to customer service inquiries — biases in the training data can surface in ways that are difficult to detect through standard testing.

MAS expects firms to assess AI models for discriminatory outcomes, test output across demographic groups, and maintain audit trails demonstrating that protected characteristics did not drive decisions. For GenAI, meeting this expectation requires a design choice: constrain GenAI to tasks where it handles narrative and presentation rather than the underlying decision.

The practical architecture is a separation of concerns. Structured data models — with their inspectable features and testable outputs — make the consequential decision: credit approved or declined, claim accepted or queried, customer risk tier assigned. The LLM then takes that structured output and generates a customer communication, a case summary, or an audit record. Fairness assessment is applied to the structured model, where it is tractable. The LLM is in the loop for communication quality, not for the decision itself.

Where GenAI must be involved in consequential decisions — which is sometimes unavoidable when the input is unstructured text — fairness testing should include adversarial probing with demographic proxies: names that signal ethnicity, gendered pronouns, postcodes associated with particular communities. If output distributions shift based on these proxies, that is a fairness finding that must be addressed before deployment.

E — Ethics

The ethics principle covers several dimensions that apply directly to GenAI.

The first is data use. Training or fine-tuning a model on customer data requires a lawful basis for that processing. For most financial institutions, customer data was collected under terms that permitted its use for the delivery of financial services, not for AI model training. Using that data to fine-tune an LLM without revisiting the consent and data governance basis is an ethics failure — and in jurisdictions with active data protection enforcement, a legal one.

The second is synthetic data. GenAI enables financial institutions to generate synthetic datasets for model testing, scenario planning, and system development. Synthetic data generated from real customer records carries re-identification risk: a sufficiently sophisticated adversary may be able to recover information about real individuals from synthetic records. The ethics principle requires that this risk is assessed and controlled. Synthetic data pipelines used in financial services should include re-identification testing before those datasets are used in anything other than fully air-gapped development environments.

The third is customer interaction transparency. When a customer interacts with a GenAI-powered chatbot or virtual assistant, they have a reasonable interest in knowing they are interacting with a system rather than a person. MAS’s ethics principle covers the use of AI in customer interactions — the expectation is that customers are not misled about the nature of what they are communicating with. Disclosure requirements for AI-assisted customer service are a design requirement, not a post-deployment consideration.

A — Accountability

The accountability principle is among the most operationally specific of the four. It requires that there be clear human ownership of every AI-assisted decision — that a named individual or function is responsible for the model, its outputs, and the consequences of those outputs.

For GenAI, this translates into three requirements. The first is named ownership: for every GenAI application in production, a named individual — typically a senior manager in the business function using the application — is accountable for its performance, its compliance with FEAT, and any remediation required if it produces harmful outputs.

The second is documented approval: a deployment record for each production GenAI application that evidences human review of the application before deployment. This is not a tick-box exercise. The approval record should demonstrate that the approver understood what the application does, its known limitations, and the human oversight mechanisms in place. Model cards — a structured document covering what the application does, what data it was trained or configured on, its known failure modes, and what it is explicitly not designed to do — are the mechanism for making this documentation structured and reviewable.

The third is escalation procedures: clear processes for what happens when the model’s output is contested or uncertain. If a customer disputes a decision that was influenced by a GenAI system, who reviews that dispute, how do they access the relevant audit trail, and what authority do they have to override the system’s output? These procedures need to be designed before deployment, not after the first complaint.

T — Transparency

The transparency principle requires that customers and regulators can understand, at an appropriate level, how AI-assisted decisions were reached. The level of detail required varies by context: a retail customer asking why their loan application was declined needs a plain-language explanation; a regulator conducting a supervisory examination of the firm’s AI risk management needs access to model documentation, testing records, and audit logs.

For traditional ML, explainability has a reasonably developed toolkit. SHAP (SHapley Additive exPlanations) values quantify the contribution of each input feature to a model’s output. LIME (Local Interpretable Model-agnostic Explanations) provides local approximations. These tools are imperfect but are accepted by regulators as a good-faith approach to explainability.

LLMs present a harder explainability challenge. The mechanism by which a large language model generates a particular output — which tokens in the context window influenced which tokens in the output, and through what pathway — is not inspectable in the way that feature importance for a decision tree is inspectable.

Practical approaches for GenAI transparency include: where the LLM’s input includes structured data features, SHAP-value attribution on those features provides the explainability story; retrieval-augmented generation (RAG) systems can log the specific documents or data chunks that were included in the retrieval context for each generation, providing a retrievable record of what information influenced the response; for customer-facing explanations, the prompt itself can be designed to instruct the LLM to produce a reasoning chain — a human-readable account of the factors that influenced the output.

Audit logging is the floor, not the ceiling, of transparency. Every production GenAI interaction in a regulated use case should produce a log record including the input, the model version, the retrieval context if applicable, and the output. These records are the evidence that supports any subsequent regulatory examination.

The FEAT Assessment Framework

MAS published the FEAT Assessment Framework (FAAF) in December 2019 to provide a structured methodology for assessing AI applications against the FEAT Principles. The FAAF is a structured questionnaire and assessment process that takes a firm through a systematic evaluation of each principle as applied to a specific AI application.

For GenAI applications, the FAAF remains the appropriate starting point, but several questions require GenAI-specific interpretation. Fairness questions about feature selection and model training assume a traditional ML architecture; for LLMs, the equivalent question is about the composition of the foundation model’s training data and the content of any fine-tuning dataset. Accountability questions about model documentation should be interpreted to require model cards covering not just the fine-tuned application but also the foundation model on which it is built and the constraints applied through system prompting and retrieval configuration.

Firms that have applied FAAF to traditional ML models but have not revisited it for their GenAI applications should treat each new GenAI deployment as a fresh FAAF assessment. The framework is technology-agnostic enough to accommodate GenAI, but the answers to its questions look different from the answers for a gradient-boosted credit scoring model.

The Veritas Consortium

The Veritas Consortium is an industry-led collaboration convened by MAS to develop open-source tools for assessing AI against the FEAT Principles. The consortium, which includes major Singapore-licensed banks, insurers, and fintech companies, has produced the Veritas Methodology and associated toolkits covering fairness assessment for credit risk, customer marketing, and fraud detection use cases.

The Fairness Assessment Methodology (FAMS) toolkit provides practical implementation guidance and open-source code for computing standard fairness metrics on model outputs. For regulated financial institutions applying FEAT to GenAI deployments, the Veritas toolkit is not a complete solution — it was not designed for LLMs — but the methodological framework it provides for scoping a fairness assessment, selecting appropriate metrics, and documenting findings is directly applicable.

The Veritas Consortium materials are publicly available through the MAS website. Compliance and AI governance teams at financial institutions in Singapore and across the region should treat these as a primary reference alongside the FEAT Principles text and the FAAF.

Applying FEAT Proactively to GenAI

The firms that are ahead of the curve are not the ones waiting for MAS to issue specific GenAI guidance. They are the ones that are applying the existing FEAT framework to every GenAI deployment they make now — documenting their fairness assessments, assigning accountability, building the audit trail, and designing transparency mechanisms at the point of design rather than retrofitting them after deployment.

FEAT was written to be enduring. Its four principles — Fairness, Ethics, Accountability, Transparency — are not defined in terms of any particular technology, which means they survive the transition from traditional ML to GenAI. The application of the principles to GenAI requires technical and organisational judgment that goes beyond what the original framework documents specify, but the principles themselves are the right starting point.

The regulatory expectation in Singapore, and increasingly across the region, is that financial institutions deploying AI — including GenAI — have a structured governance process for those deployments. FEAT is that process. Applying it to GenAI now, before specific GenAI guidance arrives, is both the responsible and the prudent course.

Find out how Nematix’s Strategy & Transformation practice can align your technology investments to business outcomes.