From MRM to ModelOps

What is MLOps

Over the past few years, the discipline of MLOps has emerged in response to the growing complexity of deploying machine learning models at scale. Drawing inspiration from DevOps, it introduced structured standards and processes for managing the lifecycle of ML models, becoming an essential philosophy and toolkit for data-driven organisations.

MLOps embodies a set of principles that are becoming widely recognised:  

1. Portability: Models can move seamlessly across tools and environments

2. Explainability: Outputs can be justified to users, stakeholders, and regulators

3. Deployment: Automated through controlled workflows, not manual

4. Reproducibility: Model runs can be recreated exactly

5. Versioning: Code, data, and models are version-controlled and traceable

6. Monitoring: Performance, drift, and stability are continuously tracked

7. Retraining: Models are retrained automatically or on defined trigger.

8. Segregation: Development, staging, and production are clearly isolated

9. Access Control: Only authorised users can deploy, modify, or promote models

10. Auditability: Every action, decision, and change is logged 

Beyond MRM

But here’s the challenge. In banking and financial services, many of the most critical models are not based on machine-learning; they're structural, deterministic and highly governed — powering credit decisions, financial forecasts, and regulatory submissions. The dominant framework applied to these models - model risk management (MRM) – pre-dates MLOps but takes a compliance-first approach to the model lifecycle. While essential, MRM only indirectly addresses operational capabilities. As a result, many institutions with mature MRM functions still rely on manual deployments, spreadsheet tracking, and fragmented tooling.

ModelOps — the broader discipline that extends MLOps principles to all model types — must now be prioritised alongside MRM as a core pillar of the modern model lifecycle. Not only does it improve efficiency and velocity, but it also makes managing model risk easier. The traditional approach of overlaying MRM controls onto creaking systems and legacy processes is becoming unsustainable.