AI Governance for Pension Funds: Dutch Pension Act (WTP) Transition and DNB Expectations

The Dutch pension sector faces one of the most far-reaching transformations in its history. The Dutch Pension Act (WTP) requires pension funds to transition from the current defined benefit system to a defined contribution arrangement with individual pension assets. This transition — with a statutory deadline rapidly approaching — requires recalculation of millions of individual entitlements, fundamental adaptation of communication strategies and extensive revision of risk management frameworks.

At the same time, the application of artificial intelligence in the financial sector is growing rapidly. For pension funds, AI offers concrete opportunities to make the WTP transition more efficient, more accurate and more participant-friendly. But deploying AI in a sector that manages the retirement savings of millions of Dutch citizens requires the utmost care — particularly given the increasing expectations of DNB regarding AI governance.

This article explores how pension funds can responsibly deploy AI during the WTP transition, which governance frameworks are needed, and how DNB's expectations translate into concrete measures.

DNB expectations: from principles to practice

DNB has paid increasing attention in recent years to the risks and opportunities of AI applications at supervised institutions. The Good Practice for the use of AI, published by DNB, describes expectations in the areas of governance, risk management, data quality and explainability of AI models. For pension funds, these expectations translate into five concrete focus areas.

Model governance and ownership. DNB expects pension funds to designate a clear owner at management level for each AI application, who is responsible for the functioning, performance and risks of the model. This goes beyond the traditional model ownership structures that many funds know from their ALM models — AI models require ongoing monitoring and periodic revalidation due to their adaptive nature.

Explainability and transparency. Given the fiduciary responsibility of pension funds towards their participants, DNB places particular importance on the explainability of AI-driven decisions. When AI is deployed in processes that directly affect participant interests — such as the calculation of individual pension assets or communication about expected benefits — the fund must be able to explain how the model arrived at its output.

Data quality and data management. The quality of AI output is directly dependent on the quality of input data. DNB expects pension funds to have a robust data governance framework that ensures the completeness, accuracy, timeliness and consistency of data. In the context of the WTP transition, where historical data spanning decades must be processed, this is a particularly relevant requirement.

Risk management and bias detection. AI models may contain unintended bias that leads to systematic disadvantage for certain groups of participants. DNB expects funds to actively test for such bias and take measures to mitigate it. This is particularly relevant for AI applications in participant communication and segmentation.

Outsourcing and chain responsibility. Many pension funds outsource operational processes to pension administration organisations (PUOs) and asset managers who are increasingly deploying AI. DNB expects the fund board to maintain adequate oversight of AI applications throughout the chain, including insight into the models that chain partners use and the associated risks.

AI use cases for the WTP transition

The WTP transition generates a range of complex challenges for which AI can deliver concrete added value. The following use cases illustrate where the technology can make a difference.

Asset-liability modelling and scenario analysis

The transition to individual pension assets requires a fundamentally different approach to asset-liability management. Instead of collective funding ratios, funds must model individual pension assets, accounting for age-dependent investment mixes, solidarity reserves and the conversion of existing entitlements. AI — particularly machine learning techniques — can significantly enrich the scenario analyses needed for this purpose.

Traditional ALM models work with a limited number of predefined economic scenarios. Machine learning makes it possible to generate a much broader spectrum of scenarios, including tail risks and non-linear correlations that are often underestimated in conventional models. This provides the fund board with richer information for conversion decisions and the design of the solidarity reserve.

Furthermore, AI can be deployed for real-time monitoring of individual pension assets after the transition. Deviations from expected trajectories can be detected early, enabling the fund to intervene proactively rather than reactively.

Participant communication and personalisation

One of the greatest challenges of the WTP transition is communication to participants. Millions of Dutch citizens need to understand what the transition means for their personal pension situation — a message that is by definition different for every individual. Natural language generation (NLG) and generative AI offer pension funds the ability to produce personalised communications at scale.

Concretely, AI can be deployed to generate individual transition documents that explain in understandable language what changes for the specific participant: what was the old entitlement, what will the expected new asset value be, which factors play a role, and where can the participant go with questions. AI-powered chatbots and virtual assistants can additionally be deployed to answer frequently asked questions and guide participants through complex choice scenarios.

The governance challenge here is evident: the generated communication must be factually correct, must not be misleading, and must comply with the requirements that AFM imposes on pension information. This requires robust validation processes and human oversight of AI output, particularly in the initial phase of deployment.

Risk management and fraud detection

The WTP transition introduces new risks that call for enhanced monitoring. AI can be deployed for detecting anomalies in transition calculations, identifying data quality issues in historical participant records, and monitoring operational risks throughout the migration process. Machine learning models are particularly well suited for recognising patterns that indicate errors or inconsistencies in large datasets — precisely the type of control needed when converting millions of individual entitlements.

Additionally, AI can contribute to strengthening fraud detection, particularly for processes related to value transfers between funds and the processing of life events that affect individual pension assets.

A governance framework for AI at pension funds

To deploy AI responsibly — and meet DNB's expectations — pension funds need a specific governance framework that connects to their existing governance structure and the requirements of the Dutch Pensions Act. The following elements are essential.

AI policy at fund level. The fund board should establish an explicit AI policy that describes the frameworks within which AI may be deployed. This policy covers at minimum: the objectives of AI deployment, risk tolerance per application type, requirements for explainability and transparency, roles and responsibilities, and procedures for validation and monitoring. The policy is periodically evaluated and adjusted to advancing insights and regulation, including the EU AI Act.

Risk classification of AI applications. Not every AI application carries the same risk. A model used for internal reporting automation has a different risk profile than a model that calculates individual pension assets. The governance framework should contain a classification methodology that categorises AI applications by risk level, with corresponding governance requirements. The classification aligns with the risk categories from the EU AI Act, supplemented with sector-specific criteria relevant to pension funds.

Validation and monitoring. For each AI application, validation procedures are established that are executed both before and after deployment. Pre-deployment validation includes backtesting, stress testing and bias analysis. Post-deployment monitoring includes performance measurement, drift detection and periodic revalidation. The frequency and depth of monitoring are proportional to the risk level of the application.

Chain governance. Pension funds that outsource AI applications to PUOs or other chain partners should contractually stipulate which governance requirements apply to the AI models deployed on the fund's behalf. This includes reporting obligations, audit rights and escalation procedures for incidents or performance issues.

Compliance considerations: EU AI Act and sectoral regulation

Pension funds operate in an increasingly complex regulatory landscape when it comes to AI. In addition to the DNB Good Practice, the European AI Regulation (EU AI Act) is of growing importance. AI applications deployed for creditworthiness or risk assessment of natural persons are classified as high-risk under the EU AI Act, with extensive requirements in the areas of conformity assessment, documentation and human oversight.

For pension funds, the relevant question is whether AI applications used in calculating individual pension assets or in decisions about the allocation of compensation from the solidarity reserve fall under this high-risk category. A prudent interpretation would confirm this, given the potential impact on the financial interests of individual participants.

Pension funds are therefore well advised to proactively structure their AI governance to the level that the EU AI Act requires for high-risk applications. This includes establishing a risk management system, maintaining technical documentation, implementing data quality management, ensuring human oversight, and implementing procedures for post-market monitoring.

Getting started: priorities for fund directors

The combination of the WTP transition, increasing supervision and growing AI capabilities calls for action. For fund directors and CROs, the following steps are priorities:

1. Map your current AI landscape. Inventory which AI applications are already being deployed — both internally and at chain partners. Many funds discover that there is more AI in their processes than they suspected, particularly in outsourced operational activities.
2. Establish an AI policy at fund level. Do not wait for regulation to catch up with you. A proactive AI policy not only strengthens your compliance position but also provides direction to the organisation when evaluating AI initiatives.
3. Identify WTP-related use cases. Determine where AI can add the most value to your specific transition challenges. Focus on applications that improve both operational efficiency and participant experience.
4. Invest in knowledge and capacity. Ensure that the fund board and key function holders have sufficient understanding of AI risks and opportunities to fulfil their governance responsibility. This requires targeted training, not just awareness.

The pension sector stands at the beginning of a dual transformation: the WTP transition and the emergence of AI as an operational and strategic instrument. Funds that succeed in connecting both transformations — by deploying AI as an accelerator of the WTP transition, within a robust governance framework — create an advantage that translates into lower transition costs, better participant communication and a stronger position with regulators.