Agentic AI: More time for what really matters

Caseloads in social insurance are growing - and the pace is picking up. In December 2024, 2,594,700 people were drawing an OASI old-age pension, around 400,000 more than ten years earlier - an increase of 18 per cent. And this is only the beginning: according to the Federal Statistical Office, the number of people over 65 will rise by around 50 per cent by 2055.

At the same time, the balance between people in work and pensioners is shifting. In 2024, there were 38 people over 65 for every 100 people of working age; by 2055, that figure is expected to reach 51. The Swiss National Bank also anticipates that over the next ten years, around 400,000 more people will leave the labour market than young people will enter it.

The conclusion is uncomfortable but clear: the approach that has worked for decades - absorbing growing caseloads by hiring ever more staff - will no longer work. The people who would need to be hired simply aren’t there.

Many organisations have already automated a great deal. Standardised processes with clear rules, stable data and unambiguous decisions have been handled efficiently for years through workflows, interfaces and rule-based systems.

The bottleneck lies elsewhere: in demanding, expertise-driven processes involving many documents, multiple sources of information, exceptions and knowledge built up through years of experience. Wherever information first has to be understood, put into the right context and checked for completeness, traditional automation often reaches its limits.

It is precisely these cases that shape the day-to-day work of many case specialists. They gather documents, compare information, resolve inconsistencies, coordinate follow-up queries, document progress and make sure a file is ready for a decision.

Our experience from client projects shows that around 70 per cent of specialists’ working time goes into these case-related support activities. Often only around 30 per cent is left for the actual expert work - assessing a complex situation, evaluating an exception, ensuring the quality of a decision, or providing personal advice.

Traditional automation (workflow or case management) follows a predefined sequence: when a specific event occurs, a specific step is triggered. This works extremely well as long as the input data is structured and the process follows clear rules.

Generative AI supports specialists on request. It can summarise documents, answer questions or produce drafts.

Agentic AI goes further. Within a clearly defined mandate, an AI agent carries out several process steps largely on its own. For example, it recognises an incoming document, extracts the relevant information, links it to the right case, checks for completeness, and triggers the next scheduled step.

The difference is significant:

• Traditional automation handles clearly defined, structured sequences.
• Generative AI answers questions or creates content on request.
• Agentic AI carries a task through several process steps - even when information from different documents and sources has to be brought together.

This makes it possible to tackle processes that were digitally supported but could rarely be automated end-to-end.

An AI agent is not better than an experienced specialist. The best results are achieved when specialists work with AI support.

The agent can structure large volumes of information quickly and consistently. It can repeat defined checks, monitor deadlines and ensure that no scheduled process step is overlooked. However, it takes on neither professional judgement nor responsibility for legally relevant decisions.

The specialist reviews, assesses and decides. They make corrections where an individual case requires it and maintain the dialogue with the people behind the case. The agent lays the groundwork so that they can perform these tasks with the time and attention they deserve.

Accordingly, recognised AI governance standards emphasise that roles, responsibilities, human oversight and the ongoing monitoring of AI systems should be clearly defined.

In social insurance, deploying AI is therefore not merely a technology project. It is a matter of governance, data protection and accountable, traceable responsibility.

The Swiss Federal Act on Data Protection also applies to AI-supported processing of personal data. The Federal Data Protection and Information Commissioner (FDPIC) points out that the purpose, data sources and functioning must be transparent. In the case of automated individual decisions, data subjects have a right to information and, in principle, the right to have the decision reviewed by a human. edoeb.admin

A professionally designed AI agent therefore needs clear guardrails:

• It only processes the data required for its task.
• Its access rights are precisely defined and role-based.
• Every step it performs is logged and auditable.
• Exceptions, inconsistencies and risk-relevant situations are handed over to a specialist.
• Quality, error rates and corrections are measured on an ongoing basis.
• The data architecture meets the institution’s legal and organisational requirements while enabling, via interfaces (APIs), the controlled integration of surrounding systems whose data is needed for the processes being optimised.

Agentic AI gives specialists back time for the tasks that make the biggest difference: assessing complex situations, making fair and transparent decisions, and advising insured persons and employers personally.

The future of pension provision therefore does not lie in fully automated administration. It lies in a better division of labour: existing automation remains the right choice where processes are clear and stable, while agentic AI additionally unlocks the demanding, expertise-driven processes that until now involved too many documents, variants and too much context.

The agent takes care of everything that can be reliably prepared, checked and coordinated. The specialist remains responsible for whatever calls for expertise, judgement and empathy.