In the coming years, more experienced staff than ever before will be retiring from banks, Sparkassen (savings banks) and insurance companies. As they leave, knowledge that is not documented anywhere will also leave the organization. Artificial intelligence (AI) promises to preserve precisely this knowledge. Some of this is already a reality, whilst other aspects remain merely a promise for now.
A study by Robert Half and Protiviti, involving 245 executives from German banks and insurance companies, shows just how seriously the situation is already being taken. 54.3 per cent of those surveyed rank demographic change as the most important factor influencing their HR strategy up to 2030. 42.1 per cent expect that, by then, between 10 and 20 per cent of their positions will remain permanently unfilled.
An example from the Sparkassen sector illustrates just how tangible this situation can be. In January 2026, Ulrich Reuter, President of the Sparkassen, told the news agency Bloomberg that of the approximately 195,000 employees of the Sparkassen, around 120,000 are likely to retire over the next ten years. At present, this can still be offset by new recruitment. However, the situation becomes more difficult when it comes to specialized roles: "The more specialized the role, the harder it is to find suitable candidates," said Reuter.
This pattern can also be observed internationally. In many financial institutions, core applications continue to run on COBOL-based systems that are decades old. According to research by Reuters, this programming language – which is over 60 years old – continues to process a large proportion of banking transactions in the US, and the number of people capable of maintaining these systems is also falling steadily there.
Why documentation and mentoring are not enough for knowledge transfer
Knowledge can be broadly divided into two forms. Explicit knowledge can be easily put into words and documented: facts, rules, process descriptions, figures. It is found in manuals, wikis or databases and can be passed on independently of the person who originally possessed it.
However, much of what experienced staff actually know falls into a second category: implicit or tacit knowledge. This includes patterns of experience, rules of thumb and a feel for exceptional cases, which develop over many years and are rarely found in manuals. Those who possess this knowledge often do not consider it anything special themselves because it feels like common sense, yet it is the result of years of practical experience. It is precisely this knowledge that often determines how a complex special case is handled or how an exception is correctly categorized within the system.
Traditional knowledge management tools only partially address this problem. Knowledge databases and documentation do a good job of preserving explicit knowledge, but they are hardly capable of capturing implicit knowledge. Mentoring programmes and work shadowing transfer implicit knowledge in person, but they are resource-intensive and heavily dependent on individual people. For this reason, a third option is coming to the fore in professional circles: AI-based systems.
What AI-powered knowledge management actually achieves
An article by two Accenture executives in the California Management Review, published by the University of California, Berkeley, describes key building blocks that enable companies to make implicit knowledge usable for AI systems. First, it becomes clear where critical knowledge is actually located, for example, by analysing which colleagues are most frequently approached for informal advice. This knowledge is then mapped onto so-called knowledge graphs: structures that not only store documents but also capture relationships, exceptions and context.
The final building block is crucial: if the system encounters a case it cannot classify with certainty, it is referred to a specialist. The specialist's decision is fed back into the system, keeping it up to date.
This principle can already be observed in the financial service industry. Morgan Stanley has collaborated with OpenAI to build an internal system called AI @ Morgan Stanley Assistant. It uses GPT-4 to make around 100,000 in-house research documents and analysts' expertise searchable – knowledge that has been accumulated within the company over decades. According to the joint case study by OpenAI and Morgan Stanley, over 98 per cent of advisory teams now use the system on a daily basis. Access to relevant documents has risen from 20 per cent to 80 per cent.
Jeff McMillan, Head of Firmwide AI at Morgan Stanley, describes the effect as follows: "This technology makes you as smart as the smartest person in the organization." Previously, the system answered around 7,000 predefined questions; today, it can answer virtually any question drawn from the entire document corpus, as David Wu, Head of Firmwide AI Product & Architecture Strategy at Morgan Stanley, reports.
COBOL and legacy systems: Even older systems can benefit from AI
A second area of application concerns the technical side: legacy systems such as COBOL-based banking applications. According to industry observers, generative AI can read and document historically developed program code and translate it into understandable, modern language.
This makes a body of knowledge accessible that was previously held only by a handful of specialists, who understand exactly why an application functions in a particular way at a specific point. For institutions that will still be running old and new systems in parallel for years to come, this is a practical starting point.
The limits of AI in knowledge management
Not every figure from this market environment stands up to close scrutiny. Numerous providers of knowledge management software tout impressive success stories that often cannot be independently verified. Cases such as Morgan Stanley's, where named executives from a large, regulated financial institution publicly account for results or independent research such as the California Management Review article are more reliable.
It is also important to note that AI-supported knowledge management replaces neither mentoring nor documentation, but rather complements both. The Accenture authors expressly emphasize that executives should position AI as a partner to human judgement, not as a substitute for it. Without feedback from experienced staff, any system is only as good as the data it was last trained on.
The next steps for banks, Sparkassen and insurance companies
The generational shift won't happen overnight, but the clock is already ticking. Institutions that start capturing their experienced staff's knowledge systematically now, instead of waiting until gaps appear, will gain a head start. Three steps provide a realistic starting point:
- Map critical knowledge before the gap arises: Who possesses specialist knowledge that is not documented anywhere?
- Start with a clearly defined use case, rather than attempting to cover the entire organization straight away.
- Involve experts from the outset so that the system remains capable of learning, rather than becoming obsolete after implementation.
Those who take these steps now are not only safeguarding knowledge. They are buying themselves time, which will become the scarcest resource during the generational transition.
Sources
Robert Half: Fachkräfte vs. Fähigkeiten: Woran es der Finanzbranche wirklich mangelt (Study „Future Workforce 2030“ with Protiviti, 20.02.2025, updated 11.04.2025).
FONDS professionell: „Wir raten davon ab“: Sparkassenpräsident zu Kryptos für die Rente (15.01.2026, Bloomberg-Interview with Ulrich Reuter).
CAST Software: Why COBOL Still Dominates Banking – and How to Modernize (16.05.2024).
Teresa Tung, Philippe Roussiere: Tacit Knowledge Is Your Next Competitive Moat, California Management Review Insights, UC Berkeley Haas (16.03.2026).
OpenAI: Morgan Stanley uses AI evals to shape the future of financial services (Study with Morgan Stanley).
AInvest: Banks Face Looming COBOL Talent Gap – Mainframe Modernization Is No Longer Optional (13.04.2026).