AI governance for BSP-supervised financial institutions

The Bangko Sentral ng Pilipinas (BSP) has told the institutions it supervises what good AI governance looks like. Memorandum No. M-2026-031, dated 24 June 2026 and signed by Deputy Governor Lyn I. Javier, issued a guidance paper titled Governance Principles for Artificial Intelligence in Financial Services. It is voluntary. It also sets out the BSP’s minimum supervisory expectations, which is how supervisors will read it. This guide takes the principles in turn and sets out the work each one implies for a bank, e-money issuer or other supervised institution.

What the BSP issued

The principles apply to all BSP-supervised financial institutions (BSFIs) and extend to outsourced service providers that support AI activities, under a shared responsibility model. They complement the BSP’s existing IT risk management framework, which already covers information security, outsourcing and project management, and forthcoming model risk management rules. The BSP says they align with OECD and ASEAN guidance and draw on the work of the Financial Stability Institute.

Implementation is meant to be proportionate to each institution’s risk profile, operational complexity and the scale of its AI. A rural bank using one vendor chatbot and a universal bank running credit models in-house will build different programs, but both start from the same questions.

The five STARS principles

Sustainability. AI should deliver long-term value and take account of its environmental and social impact.

Transparency. Institutions should keep a current inventory of their AI systems, disclose when AI influences a decision, keep documentation sufficient for audit and be able to explain how their AI reaches its outputs.

Accountability. Humans remain responsible for AI-assisted decisions, with roles assigned across the lifecycle from technical teams up to the board.

Responsibility. Framed as social fairness, this asks institutions to avoid harmful or discriminatory outcomes, guard against biased training data, handle data properly, uphold data subject rights and make sure AI does not shut vulnerable groups out of financial services.

Security. Institutions should apply strong cybersecurity and data quality controls, test regularly for AI-specific threats such as adversarial attacks and data poisoning, and monitor continuously, including for bias and hallucination.

Start with the inventory

Transparency names the AI inventory, and every other principle depends on it. Accountability needs a named owner for each system. Responsibility needs to know which systems make decisions about customers. Security needs to know what to test. An institution cannot meet any of them for AI it does not know it runs.

In practice the inventory is a register of every AI system and agent, each with an owner, a vendor where there is one, the data it uses, whether it affects customers and its lifecycle status. It has to reach AI that arrives inside software the institution already uses, not only projects that are declared. The AI system register guide covers how to build one.

Work through the lifecycle

The guidance paper maps five stages, namely planning, development, validation, deployment and monitoring, with controls at each. Those controls are risk-based assessment, documentation, independent validation, testing and continuous monitoring. The BSP treats the lifecycle as iterative, so the work repeats as systems change and new risks emerge.

Two points stand out. Validation should be independent, with a team separate from the developers reviewing AI outputs before deployment. And monitoring does not stop at go-live. Credit scoring, fraud detection and customer service models drift as data and behavior change, so an institution needs limits on the measures it tracks and alerts when they are crossed. The drift monitoring guide covers how to set that up.

Fairness and data privacy

The responsibility principle overlaps with the Data Privacy Act. AI that scores or screens customers will often engage the Act’s rules on automated processing, including the right to information on automated decisions and to object to profiling. The National Privacy Commission’s Advisory No. 2024-04 expects privacy impact assessments, meaningful human intervention and bias controls for AI. One assessment per system that covers both the BSP’s fairness expectations and the NPC’s privacy expectations avoids doing the work twice. The bias assessment guide covers the fairness side.

Vendors and shared responsibility

Most AI in Philippine financial institutions is bought, not built. The shared responsibility model means the institution must be clear, contract by contract, which controls the vendor owns and which stay in-house, and must have oversight over the vendor’s part. Start from the inventory, because every system in it names the company behind it. The vendor AI risk guide covers due diligence and contract terms.

Where Aicura fits

Aicura’s AI Register holds every system and agent with its owners, vendor, data, status and version history, and discovery surfaces AI in use that is not yet on it. Risk assessments start when a system is registered and run again on material change. At Pro, impact assessments, vendor records, security testing and monitoring sit on the same records. Aicura does not assess against BSP issuances. Its register, assessments and monitoring support the work the principles describe, and the judgment stays with your institution.

A note on this page

This is Aicura’s reading of public reporting on BSP Memorandum No. M-2026-031 and its guidance paper. It is not legal or regulatory advice. Read the memorandum on the BSP’s website, and take advice on how it applies to your institution.