Banks do not need another vague conversation about artificial intelligence. They need a practical view of where agentic AI can create value, where it should be constrained, and what foundations have to be in place before more autonomy is introduced. That was the central message from a recent interview with Rachael Hadaway, Vice President of Product Management at FICO, whose perspective combined product strategy, decisioning and fraud expertise with a global view across banking markets. Her comments aligned closely with the themes at the centre of our current research, particularly control, accountability, governance, operational resilience and value realism.

A useful starting point was her rejection of the loose way the term “agentic” is currently being used. In her view, too much of the market is simply relabelling existing automation or chatbot-style capabilities as agentic AI. What matters is not whether a tool sounds intelligent, but whether it can credibly own part of a workflow in a controlled and measurable way.

Where agentic AI can add value in banking today

Rachael argued that the workflows most likely to move from model-assisted to genuinely agent-driven over the next 12 to 36 months share three features:

• They have a tight and well-instrumented decision boundary.
• They contain large volumes of repetitive low-stakes subtasks, and
• They include a clear escalation path for when the system is uncertain.

That framework is important because it shifts the conversation away from broad claims and towards operational fit. Based on that logic, she sees the most credible near-term opportunities in areas such as intake and triage, fraud case enrichment, know your customer remediation queues, collections contact strategy and certain servicing tasks where the system can gather context, apply policy and draft an appropriate next step. In those environments, the AI is not replacing the bank’s judgement. It is helping connect systems, data and people more effectively.

Why high-stakes banking decisions still need human oversight

By contrast, she was far more cautious about high-impact areas such as credit decisions, originations, complex underwriting and complaints adjudication. These functions may become faster and better supported, but she does not yet see a near-term world in which banks should allow agents to make binding decisions independently in those domains.

That distinction matters for banking leaders in the Benelux region, where many firms are under pressure to show productivity gains from AI while also dealing with rising expectations around evidence, governance and explainability. The temptation is to treat automation as a linear maturity curve, with more autonomy always being the end goal.

Rachael’s framing was more disciplined.

Why reversibility matters in AI-driven banking decisions

She suggested that the question is not where to draw a single line between human and machine authority, but how to think about the decision space as a grid. The two key dimensions are reversibility and customer or financial impact. If an action is easy to reverse and has low impact, an agent can often act, log the decision and move on without human involvement. If it is hard to reverse and carries significant impact, it should remain firmly human-owned.

Her emphasis on reversibility was especially striking. It is not always the most discussed criterion in AI governance debates, yet it may be one of the most practical. A wrong fraud decline that can be corrected within minutes is very different from an incorrect account closure that takes weeks to unwind and creates complaints, remediation costs and regulatory exposure.

In other words, the risk of an AI-enabled action cannot be judged only by whether it is right or wrong. It must also be judged by how difficult it is to undo the consequences.

How banks build trust in agentic AI over time

Rachael also offered a realistic perspective on how authority may shift over time. She does expect some use cases to move towards greater autonomy, but not because of a sudden breakthrough or a blanket increase in confidence.

Her answer was more grounded.

Trust grows through repeated exposure, consistency and controlled experience with real cases. Systems need to handle enough real-world situations, under supervision, for banks to understand how they behave when confronted with ambiguity, variation and edge cases.

That view will resonate with practitioners who have already discovered that model performance in controlled testing does not automatically translate into production readiness. Banks may want the end state of lower human intervention, but Rachael’s message was that the route there is incremental, measured and operational rather than dramatic.

The risk of human-in-the-loop becoming a rubber stamp

This connects directly to one of the most difficult questions in the current AI debate: what a meaningful human-in-the-loop arrangement should look like. Rachael acknowledged the risk that human reviewers become little more than rubber stampers, approving machine-generated actions without genuine scrutiny. Her answer was not to abandon human oversight, but to make it more deliberate.

First, review interfaces need to show the agent’s reasoning, not just its output. If reviewers cannot see why a recommendation has been made, they are far more likely to approve it reflexively. Second, banks should design deliberate friction into high-stakes approval processes, whether through second reviewers, mandatory justification fields or similar controls. Third, institutions should monitor override rates by reviewer and by action type. Very low override rates may indicate that reviewers are not engaging properly with the task, while erratic rates may point to calibration issues in the underlying system.

That is a particularly useful insight because it moves the debate beyond the simplistic question of whether a human is present. Presence alone is not enough. What matters is whether the control is functioning as intended. For banks designing AI-enabled workflows, this suggests that management information and oversight metrics will be just as important as the model itself. The control environment has to be engineered, not assumed.

How agentic customers could change banking customer journeys

One of the most thought-provoking parts of the conversation came when Rachael turned from internal operations to what she calls the “agentic customer”. Her view is that many banks are underprepared for a world in which customers are represented by digital agents acting on their behalf. If that happens at scale, a number of familiar assumptions about customer journeys start to break down.

Many current banking experiences still contain friction points that are tolerable only because human customers are inconsistent, time-poor or emotionally influenced. An agent acting for a customer does not respond to that friction in the same way. It will not be nudged by a cross-sell pop-up, discouraged by complex processes, or held in place by unclear product terms. It will simply optimise for the customer’s stated objective and route around anything that does not serve it.

What agentic customers mean for banking competition and product design

This has potentially significant implications for product design, pricing and distribution. Rachael did not suggest that banking will inevitably become a pure commodity business, but she did argue that some traditional sources of margin and retention will come under pressure. Where opacity, complex fee structures or hard-to-exit processes still support economics, those advantages may weaken in a world of customer agents.

Banks that succeed in that environment are likely to be those with clearer, more structured and more machine-readable propositions. In other words, value may need to become more explicit. That does not mean differentiation disappears, but it does mean some forms of hidden friction become less sustainable. For Benelux banks, many of which compete in mature and transparent markets, that is an angle worth paying attention to now rather than later.

What unified decisioning means for banks using AI at scale

This is also where Rachael linked the discussion to FICO’s broader idea of unified decisioning. She described it as an orchestration layer for decisions across fraud, credit, AML, servicing and marketing. In her framing, the main benefit is not simply a single business KPI, but governance clarity and cross-lifecycle consistency. That is important because one of the frustrations customers often face is that different parts of the same bank behave as if they have no relationship with one another.

An offer seen in one channel is not recognised in another. Decisions made in servicing do not align with the bank’s treatment in collections or marketing. A more unified decisioning approach is intended to reduce those disconnects by coordinating data, models and policies in one place.

For financial services professionals, the interesting point here is that unified decisioning can be understood not just as a technology architecture choice, but as a governance model. If the bank increasingly relies on AI to recommend or trigger actions, then the question becomes where policy, oversight and evidence are anchored. A fragmented estate of point solutions may still deliver local efficiencies, but it can also make it harder to explain decisions coherently across the customer lifecycle.

How banks can use AI to fight fraud more effectively

Fraud was another area where Rachael’s comments were especially grounded. She described the challenge in simple terms: the only way to catch more fraud is to work more fraud cases. Agentic capabilities can help institutions do that by increasing speed and scale, but the contest is dynamic because fraudsters are also using AI to become more creative and prolific.

Her description of fraud fighting as an endless process of retraining and expanding the definition of suspicious behaviour felt realistic rather than promotional. It reinforces the view that AI in fraud is not a one-off deployment, but an ongoing operational discipline.

AI in US banks versus European banks: what is different?

Finally, her comparison between the US and Europe offered a useful corrective to some lazy assumptions. She argued that Europe is not simply more cautious. It is more procedural. European banks tend to move more slowly at the experimentation stage because the control expectations are clearer and they build towards them from the outset. US banks often move faster initially but then spend considerable time retrofitting their governance after the fact. As a result, the eventual time to production may be less different than many assume.

Interestingly, she also suggested that the two regions are beginning to converge, with US banks trying to bring more upfront control into experimentation and European banks giving teams somewhat more room to test.

What Benelux banks should do now to prepare for agentic AI

Her closing advice for banks wanting to prepare without getting lost in the hype was practical and specific.

1. First, create an inventory of the decisions the bank actually makes. Many organisations rush towards the technology without understanding their own decision landscape.
2. Second, build the policy and evidence layer before the agent layer. Versioned policies, decision logs and a clean audit view should come before broad deployment ambitions.
3. Third, choose two production use cases rather than launching a long list of pilots. In her view, areas such as fraud case management and servicing workflows are sensible places to start because the data is rich and the failure modes are more bounded.

The future of agentic AI in banking will depend on control and clarity

Taken together, Rachael’s perspective was a useful antidote to both hype and cynicism. She did not dismiss the potential of agentic AI, but neither did she present it as a near-term substitute for judgement in high-stakes banking decisions. Her message was that progress will come through disciplined use case selection, strong evidence design, a serious view of reversibility and a willingness to rethink decisioning as a bank-wide capability rather than a set of disconnected experiments.

For banks in the Benelux region, that may be the most valuable lesson of all.

The question is not whether agentic AI will matter. It is whether institutions can build the control, clarity and decision foundations needed to use it well.

You can watch or listen to the full interview below, or follow along on your favourite podcast platform here (don't forget to subscribe for more insights from industry experts!).