ING and Agentic AI: What COOs Should Learn from the Operating-Model Push

For a COO or CFO at a European bank, the most important question is not whether agentic AI is coming, but how to deploy it so that the result is measurable efficiency rather than another pilot-graveyard story. On 18 March 2026, at the 22nd Morgan Stanley European Financials Conference in London, ING gave an unusually concrete glimpse of how the bank is approaching this. The title of the presentation: "Driving value through digitalisation, scalability and (Gen)AI". The figures cited are impressive. That is precisely why it is worth contextualising them cleanly before transferring them into your own board deck.

The temptation is strong to label ING the "first major European bank with an agentic-AI operating model" and to attribute the cited savings figures one-to-one to agentic AI. Neither holds up to scrutiny. Anyone who does so builds on a secondary-source causal bridge and hands their own board a narrative that collapses at the first critical question. This article therefore consistently separates the evidenced from the asserted – and distils from it what is genuinely transferable to your own organisation.

What ING Announced on 18 March – and What Lies Behind It

Let us start with what was actually on the table. At the conference, ING set out how digitalisation, scalability and (Gen)AI are intended to create value together. It is in this context that the much-cited figures sit: an FTE reduction of around 1,250 roles in 2026, of which about 950 in the Netherlands with a focus on AML and KYC functions, and incremental cost savings in the order of EUR 350 million.

Here lies the first and most important clarification. These figures belong to a broad digitalisation and efficiency programme, not to an isolated agentic-AI initiative. Agentic AI is one of the drivers, but not the only one, and in terms of the concrete FTE effect in 2026 not even the dominant one. The largest lever sits in the streamlining of the AML and KYC functions, an area in which ING employs around 6,000 staff globally and in which AI increasingly takes over transaction reviews. Anyone who sells the 1,250 roles or the EUR 350 million in their own board deck as an "agentic-AI saving" builds a causal bridge the primary source does not support. The honest formulation is: agentic AI is part of an efficiency programme whose figures cannot be cleanly attributed to a single technology.

The label "first major European bank with a public agentic-AI operating-model commitment" does not hold either. BBVA already publicly communicated a dedicated C-suite AI transformation unit in 2025 and early 2026. ING is therefore not the first, but rather the front-runner with the most concrete public efficiency commitment. This distinction is not pedantry: it determines whether the analysis withstands critical scrutiny or is exposed as overstated benchmarking.

What is genuinely remarkable about the ING presentation is less the size of the savings than the implementation rate it cites. ING puts the share of its (Gen)AI pilots that make it into production at around 90 percent – against an often-cited industry average of about 30 percent. That is the truly interesting figure, because it speaks not of the technology but of the discipline with which an organisation steers its pilots. More on that shortly.

The Agentic Mortgage Assistant: Where It Really Stands

The most-cited single example is the agentic mortgage assistant in the Netherlands. Here, too, a close look pays off, because the distance between what is running in production and what has been announced is greater than many reports suggest.

In production today, the assistant operates in the back office. It extracts documents, runs policy checks and prepares files for further processing. That is a real, measurable productivity lever – but it is an assist scenario, not a fully autonomous application process. The complete agentic processing of mortgage applications, in the Netherlands and in Germany, is announced for 2026, not live. This dividing line is decisive for your own planning: a back-office assist can be reproduced today, whereas a fully autonomous application pipeline is a goal whose feasibility still has to be demonstrated.

COO Marnix van Stiphout has described the mechanism aptly. It is not about replacement but about leverage – the case handler is given a small team alongside, and that is exactly what shortens turnaround times.

This choice of framing is not accidental. It describes the operating model that can genuinely be implemented for regulated processes such as mortgage processing: agentic AI as an amplifier of case handling, with humans in the decision loop. An organisation that turns this into "autonomous lending decisions" walks knowingly into a regulatory and reputational risk – and overlooks that even ING has only announced the fully autonomous step, not completed it.

Agentic AI vs. RPA: Complement, Not Revolution

A narrative recurs in the reporting that agentic AI "downgrades" classic robotic process automation (RPA). This is an editorial overstatement that the sources do not support. What the sources show is a hybrid model: RPA remains the execution layer for deterministic, rule-based tasks, while agentic AI is added as a reasoning layer for complex, unstructured tasks. The one does not replace the other; it complements it.

For a COO, this distinction is more than a question of terminology. It determines the architecture. Anyone who understands agentic AI as a replacement for RPA is planning a rip-and-replace that is expensive, dismantles existing automation pipelines and – particularly delicate in a regulated environment – forces a complete reassessment of established, approved processes. Anyone who understands agentic AI as an additional layer above the existing automation estate keeps deterministic reliability where it is needed and adds judgement where rules reach their limits. Van Stiphout's statement that virtually all product fulfilment will be touched by agentic AI is, in this sense, not a plea for wholesale clearance but for gradual penetration.

A concrete data point from ING's statements: van Stiphout cites a productivity gain of around 25 percent in a single operations process. That is an indication of the order of magnitude of what appears realistic in the hybrid model – not of blanket scaling across all processes. In addition, ING is testing voice agents in Spain and Germany. That, too, is a pilot rather than a production state, and should be contextualised accordingly.

What European Banks Should Do Now

From the ING case, four measures can be derived, staggered by time horizon. They follow one principle: adopt the blueprint, not the headline figures.

Risks and Open Questions

Three reservations belong to an honest contextualisation. First, the attribution of the figures. The 1,250 FTE and the EUR 350 million are effects of a digital programme whose dominant driver is AML/KYC efficiency, not agentic AI as an isolated technology. Anyone who attributes them wholesale to agentic AI distorts their own business case and exposes themselves to embarrassment at the first follow-up question.

Second, the RPA narrative. A "downgrading" of RPA by agentic AI is not evidenced; the reality is a hybrid model. Anyone who aligns the architecture to a replacement story is planning a project that ING precisely is not running.

Third, the maturity level. The agentic mortgage assistant is in production as a back-office assist; the fully autonomous application processing is announced, not live. Voice agents are in testing. Anyone who infers production readiness from announcements is planning for a state that even the front-runner has not yet reached.

The strategic consequence for European banks is thus clearly outlined. The value of the ING case lies not in its figures, which cannot be cleanly transferred, but in its blueprint: a tightly scoped entry point with measurable ROI, a hybrid architecture rather than rip-and-replace, a production gate that forces pilots into production, and workforce planning that runs ahead of the technology. Anyone who adopts this blueprint and leaves the headline figures aside draws from ING precisely the lesson that holds.

Christian Schablitzki

Strategy & Management Consultant · Agentic AI expert for financial institutions

Over 20 years in investment banking and derivatives trading, followed by more than 10 years advising financial institutions. Currently a Partner at Infosys Consulting in Germany. Certified in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.

LinkedIn profile →