ING: Learning Before Scaling AI at Speed

In this conversation, Marnix van Stiphout of ING reinforces a principle that underpins the bank’s entire approach to artificial intelligence: nothing is scaled until it is fully understood. While excitement around agentic AI continues to build across the industry, van Stiphout is clear that disciplined learning must always come before large-scale deployment — particularly in a highly regulated, operationally complex environment like banking.

He explains that ING’s strategy is deliberately incremental. Rather than rolling out AI broadly and hoping for the best, the bank runs small, well-defined proofs of concept designed to generate specific learning outcomes. If a solution proves effective, controllable, and valuable, it can then be scaled confidently. This approach is not theoretical — it mirrors ING’s earlier journey with machine learning. Years ago, ML models were introduced cautiously; today, they operate at full scale in areas such as behavioural modelling for anti-money laundering, where ING has deep expertise and clarity over how the models behave.

Agentic AI, by contrast, is still in its learning phase. Van Stiphout points to concrete examples in the Netherlands and Germany, where ING is using a limited number of agents in mortgage processes. These agents handle focused tasks such as salary retrieval and document submission and retrieval. The impact is significant — reducing manual work and speeding up processes — but the scope is intentionally constrained. Each agent provides insight into governance, controls, escalation, and operational behaviour.

This “learn first, scale later” mindset extends to customer-facing AI as well. Van Stiphout acknowledges that hallucination risk in chatbots is well understood across the industry. ING’s response has been relentless testing. Chatbot outputs are continuously tested in parallel, with subsets of messages reviewed daily. This ongoing scrutiny ensures issues are detected early and mitigated before they can affect customers or operations.

Crucially, van Stiphout rejects the idea that careful experimentation means slowing innovation. On the contrary, this approach allows ING to move faster with confidence. Once the bank understands how an AI capability behaves — technically, operationally, and from a risk perspective — it can be rolled out at scale without introducing unacceptable risk.

He is also clear-eyed about the future. ING does not see major additional operational risk from AI today precisely because of this cautious, structured approach. But he emphasises the importance of resisting premature conclusions. Every new capability must earn its right to scale through evidence, testing, and learning.

The message is pragmatic and consistent: AI is powerful, but power without understanding creates fragility. By treating AI adoption as a continuous learning journey rather than a race, ING is building a foundation that allows innovation to scale safely, responsibly, and sustainably.