MAS-Aligned AI Governance Framework for a Tier-1 BFSI Client
Client profile: Tier-1 Asian bank
60+ production models, no central inventory, MAS Veritas assessment in 9 months. We built the unified model risk framework, control library, and audit pack. Veritas passed at first attempt.
Headline metrics
- Models brought into governance
- 60+
- Time to deploy reduced
- 45%
- MAS Veritas pass rate
- 100%
- Months to operating model
- 9
Summary
Challenge
A leading Asian bank operating across Singapore, Hong Kong, and India had over 60 production AI and ML models with inconsistent governance, no central inventory, and an upcoming MAS Veritas assessment. Internal audit had flagged the gap; the chief risk officer needed a defensible programme within nine months.
Approach
Magnis AI built a unified model risk management framework aligned to MAS FEAT, the Veritas methodology, and the bank's existing SR 11-7 controls. We delivered the AI inventory, tiering taxonomy, control library, evaluation gates, and audit evidence pack. Cross-functional governance committee charter was approved by the board within four months.
Outcome
MAS Veritas assessment passed at the first attempt. Model deployment time reduced by 45 percent through pre-approved control templates. Three highest-risk legacy models retired and replaced with compliant successors. Bank's internal AI governance function operating independently within nine months.
Engagement narrative
The starting position
The bank's AI footprint had grown organically over six years. Fraud detection, credit decisioning, customer intelligence, AML alert prioritisation, and contact-centre AI were all in production across three jurisdictions. There was no central inventory, no consistent risk classification, no shared validation methodology, and no audit-ready evidence pack. Internal audit had raised a finding. MAS Veritas was 9 months out. The CRO needed to know what would happen if the regulator asked tomorrow.
Diagnostic phase (weeks 1 to 4)
We ran a regulatory exposure scan against MAS FEAT, RBI model risk guidance, and EU AI Act implications. We catalogued every production model and shadow AI use case across the bank. Output was a single AI register with 64 systems classified by risk tier, a gap analysis against MAS Veritas, and a prioritised 9-month remediation roadmap shared with the board.
Design phase (weeks 5 to 12)
Target operating model defined: a model risk committee chaired by the CRO, model owner roles across business units, validation conducted by an independent team. Policy library written covering acceptable use, third-party AI, generative AI, and model lifecycle. Control catalogue produced with 47 controls mapped to FEAT principles, with each control specified at engineering level (bias-test CI specs, monitoring thresholds, documentation templates).
Deploy phase (weeks 13 to 28)
Controls implemented in the existing MLOps stack. Model registry consolidated. Bias-test pipeline integrated into CI. Evaluation gates added to model approval flow. Monitoring dashboards built per model tier. Audit evidence pack assembled for the 12 highest-risk models prioritised for Veritas evidence.
Veritas assessment outcome
Magnis AI prepared the bank's submission and rehearsed regulator-style questioning with the model risk committee. Submission was made on schedule. MAS Veritas assessment passed at first attempt. The bank's internal governance function took over operations from month 9 onwards, with quarterly assurance reviews supported by Magnis AI for the first year.
Client testimonial
“We started with a regulator deadline and an inconsistent AI portfolio. We ended with a defensible operating model the board signed off on and the regulator validated. The Magnis AI team brought a level of senior depth and regulator fluency we had not seen from larger firms.”
Chief Risk Officer, Tier-1 Asian bank