Dutch pension funds and pension providers have spent much of the past few years focusing on the transition to the Future Pensions Act (Wet toekomst pensioenen, Wtp). At the same time, developments in artificial intelligence (AI) have accelerated rapidly.
For organisations that have already transitioned to the new pension system, room is gradually emerging to look beyond the necessary IT-related compliance challenges and focus on further organisational optimisation. This raises a new question: how can AI be used to achieve operational excellence across pension processes while maintaining human control?
Fool with a tool is still a fool
AI is a means, not an end in itself. Before viewing AI as the answer to our challenges, it is important to recognise that AI is an umbrella term covering many different types of systems. Every AI technology has its own strengths and limitations, meaning its capabilities vary depending on the specific application. At the time of writing, for example, Claude excels at analysing long-form documents, while Datarails AI is particularly effective in spreadsheet analysis.
The way an AI system can support operational processes therefore depends entirely on the process that is to be improved. The context of that process is equally important, including the roles involved, stakeholders, data sources and application landscape. Besides the type of AI system, its level of autonomy also determines how it can be deployed within a process.
AI on the agentic spectrum
The application of AI systems within business processes should be viewed in relation to different levels of autonomy. Broadly speaking, three levels can be distinguished: human-in-the-loop (supportive), human-on-the-loop (agentic workflows) and human-out-of-the-loop (autonomous agents). These categories should not be seen as fixed, but rather as points on a spectrum. The greater the decision-making autonomy of an AI application, the broader or more specialised its role can become within processes across the pension value chain.
Today, we mainly see task-oriented AI applications being adopted by pension fund executive offices. These applications support individual tasks while keeping the human firmly in the loop. Examples include drafting meeting minutes using Microsoft Copilot or preparing draft emails with ChatGPT.
When humans move to an on-the-loop or out-of-the-loop role, we enter the domain of agentic systems. An agentic workflow (on-the-loop) can be viewed as a structured sequence of actions, or prompts, in which the AI independently makes decisions that ultimately lead to a logical outcome.
An autonomous agent (out-of-the-loop), by contrast, can be regarded as a digital entity that independently executes processes within predefined boundaries, determines appropriate follow-up actions and completes them without human intervention.
Applying AI
Pension funds and pension providers rely heavily on repetitive processes, making the sector well suited for AI adoption. This applies to both administrative and operational activities.
Chatbots, for example, can be deployed to support participant communication and answer participant-specific questions. A chatbot on a pension fund’s public website requires a different combination of functionalities than one integrated into a personal participant portal, where individual data analysis is also involved. The accuracy and specificity of chatbot responses therefore depend largely on the underlying dataset available to the AI system.
AI also offers opportunities within Governance, Risk and Compliance (GRC) tooling. Pension funds use GRC systems to centrally monitor service providers and demonstrate compliance. AI can simplify and enhance these activities. Human-in-the-loop applications can already read contracts and automatically populate relevant monitoring fields within GRC systems. Within an agentic workflow (human-on-the-loop), AI can analyse evidence reports, perform compliance gap analyses, prepare information requests and formulate tasks within the risk management review cycle. Ultimately, an autonomous agent (human-out-of-the-loop) could independently request information from service providers based on its findings.
Keeping the human in control
As humans become increasingly removed from operational execution and decision-making within business processes, the question of how to maintain human control over operational excellence becomes increasingly important. The answer starts with a thorough understanding of which processes require change and why.
Pension funds and pension providers that have developed a Target Operating Model (TOM) or documented their Administrative Organisation and Internal Control (AO/IC) framework already have a solid foundation for embedding AI into their operations. A clear understanding of the context surrounding each process is essential, as it determines the requirements that should be imposed on an AI system. Equally important is understanding how the autonomous characteristics of a specific AI application align with that process.
Key questions include: Who interacts with the AI application—employees only, or also external service providers? Which data are being used—public information or personal participant data? Which data sources underpin the AI-driven process? And what decision logic governs the AI system?
Operational excellence
Successfully integrating AI starts with identifying the right processes and establishing clear governance frameworks for how AI systems should be deployed and what capabilities they should possess. A Target Operating Model or AO/IC framework provides an excellent starting point for identifying, contextualising and preparing processes for AI implementation.
Achieving operational excellence through AI begins with ensuring that humans remain in control of the processes across the pension value chain.
