Ensuring governance and compliance of AI

PGGM, having developed a number of use cases, identified the need for a robust MLOps setup to ensure transparency, control, and compliance of their AI models. To address these challenges, PGGM partnered with Deeploy to get guidance on how to ensure that AI applications are transparent and explainable, have monitoring capabilities, and align AI governance with its associated risks. PGGM and Deeploy together developed a prototype for the implementation of Deeploy’s platform in PGGM’s MLOps infrastructure.

After full implementation, Deeploy’s platform will allow PGGM to deploy and control their AI models, centralize all relevant documentation, and assign model owners who receive alerts for performance issues. Deeploy will also support compliance and governance by offering compliance assessments and checklists as well as monitoring capabilities for performance, drift, and evaluation.

About PGGM

PGGM is one of the largest pension administrators and asset managers in Europe, offering pension management, asset management, and management advice services.

PGGM manages pensions of various pension funds, the affiliated employers, and their employees. By June 30, 2023, PGGM managed pension assets worth EUR 229 billion euros for 4.3 million participants, primarily in the health and welfare sector, making it one of the biggest pension administrators globally. Together with Pensioenfonds Zorg en Welzijn (PFZW), PGGM supports the financial future of workers in this sector and we contribute towards a healthy and vital sector. In order to help achieve its ambitions, PGGM makes its services, knowledge, and experience available to other clients as well.

In 2019, PGGM started its AI program with the goal of boosting the achievement of its strategic goals. Since then, PGGM has been applying AI in several use cases, started to build its AI competencies, and is setting up a complete AI ecosystem.

MLOps challenges at PGGM

As PGGM’s AI initiatives expand, PGGM recognizes the need for a future-proof MLOps setup. While the organization wishes to leverage the value of AI, keeping control of AI models and complying with internal and external requirements is equally important.

As so, PGGM identified the following challenges:

  • AI applications must be transparent and explainable
  • All AI applications need to be logged in an AI register
  • AI models need to be monitored, as the risks associated with the models can change over time
  • AI governance has to be aligned with the risks the AI application introduces
  • The gap between data science and productionalized applications needs to be bridged

To address these challenges, PGGM decided to collaborate with Deeploy. This was an excellent match, as the challenges Deeploy solves align perfectly with PGGM’s needs: a responsible AI platform that can ensure transparency, control, and compliance with deployed models.

Deeploy AI Management System

Deeploy will enable PGGM to deploy their AI models while maintaining control over time. By partnering with Deeploy, PGGM can benefit from one unified deployment and serving setup, closing the gap between data science and productionalized applications.

Platform overview of Deeploy

Moreover, all models in production, relevant documentation, governance measures and compliance insights are centralized in one platform. This setup allows data science managers, relevant business managers, compliance officers, model validators and risk officers to keep oversight over AI models, with a single source of truth. Furthermore, every model can be assigned to a model owner, who is accountable for its operation in production.

To support the model owner, alerts can be configured in Deeploy. For example, if the performance of a model drops below a defined threshold, the model owner receives a notification in their mailbox.

How Deeploy can enable PGGM to ensure compliance and governance of their AI models

One important aspect for PGGM was to have a solution where all information regarding an AI application is readily available to relevant stakeholders. In addition, it was also vital to have governance features in place that align with the risks specific AI models introduce.

Deeploy provides stakeholders with the ability to view all relevant information about the AI models in one central location.

All relevant model metadata and model cards are presented automatically and all events, such as model updates and prediction requests, are tracked.

Overview of detailed model information

In addition to this tracking of information, which is vital for compliance and governance, Deeploy also provides actionable features on the latter.

To align the governance around AI models with their risks, Deeploy will support PGGM in configuring the compliance assessment required before deploying a model into production.

The assessment highlights the requirements for a model to be deployed according to responsible AI principles.  Plus, more checklists can be done after deployment, facilitating compliance and governance for all stakeholders involved.

Overview of compliance insights in Deeploy

Ensuring observability of AI models at PGGM in the future

Besides ensuring compliance and governance for the deployed  AI models, observability of models at any moment in time is also a vital requirement for PGGM. Therefore, we guided PGGM to ensure transparency and monitoring of AI models.

  • Transparency

    All models are registered in one central location, and each model has an assigned owner who is responsible for maintaining control over its ongoing operation. Furthermore, traceability of model predictions is guaranteed by using Deeploy, as all predictions and events are logged. Lastly, tailored explainability has been developed together with PGGM, to make sure model predictions is understandable by the end user.

  • Monitoring

    To keep control of the deployed models, monitoring capabilities are offered by Deeploy, such as performance monitoring, drift monitoring and evaluation monitoring. By creating configurable alerts on all monitoring metrics, the model owner receives a notification in case any thresholds are exceeded. This allows PGGM to keep close control over their models, maintaining model health and accuracy over time.

Overview of monitoring in Deeploy