Agentic AI is quickly becoming the next focal point in banking’s AI conversation. After two years of experimentation with copilots, chat interfaces and internal productivity tools, the discussion is shifting. The question is no longer whether banks can deploy generative AI. It is whether they can redesign work, customer journeys and decision-making in a way that allows AI to do more than assist. It is whether they can build systems that can act.

That was the central theme in a recent discussion with Sára Hanniker, Head of AI Solutions and Research, and Josef Dvořák, Chief AI Officer (CAIO) at Finshape.

What emerged was not a story about rapid disruption or easy wins. It was a much more useful conversation. Their view was that agentic AI in banking is real, but the route to value is slower, more structural and more demanding than the current noise around the market suggests. For banks, the challenge is not simply adding smarter technology to existing processes. It is about redesigning workflows, rethinking customer journeys, clarifying accountability and deciding where human judgement should become more valuable rather than less.

Moving from helpful tools to delegated workflows

One of the clearest distinctions in the conversation was between helpful AI and genuinely agentic AI. Sára gave a practical example. Using AI to summarise a LinkedIn profile or prepare a note is useful, but it does not amount to transformation. What becomes agentic is the point at which a workflow is split into multiple tasks, information is gathered from several places, decisions are delegated and AI starts acting as a chain of specialised agents rather than as a single assistant.

That distinction matters because it reframes the question of maturity. A bank does not become agentic because it has rolled out a copilot or launched a chatbot. It becomes agentic when it allows AI to participate in workflows involving sequencing, judgment, and action.

This is a far more demanding shift than the market sometimes implies, because it means changing the way work is designed in the first place.

Sára proposed that the management mindset has to change first. Without that, process redesign does not stick. In her view, banks need not only new technology but also redesigned knowledge inputs, analysts who can work with new workflows and teams willing to accept new ways of operating. In other words, the obstacle is not simply technical readiness. It is organisational willingness to rebuild how work gets done.

The biggest mistake banks make is underestimating the timescale

Josef was especially pointed on the gap between market perception and operational reality. He pointed out that anyone watching AI through LinkedIn might conclude that everything is already in production and scaling nicely. In practice, that is not what most institutions face. Getting from experiment to production remains hard, especially in environments where risk, legacy systems and fragmented ownership all slow progress. He suggested that the most overhyped aspect of AI today is not its capability, but the timeline attached to it.

Many institutions are still measuring AI maturity against public narratives rather than internal reality. The result is a form of strategic distortion. Teams feel pressure to present progress, vendors feel pressure to position every pilot as a breakthrough, and leadership ends up expecting transformation on a timetable that bears little resemblance to how large financial institutions actually change.

He made the point in very practical terms. If it took a bank years to roll out something like Microsoft Teams to a high level of adoption, it is unrealistic to imagine that agentic AI will be embedded overnight. The technology may move quickly, but adoption within banks still hinges on culture, incentives, politics, and process change.

Incentives and trust remain a real bottleneck

One of the more useful moments in the discussion came when the conversation moved away from familiar barriers such as data quality and legacy architecture. Sára acknowledged those issues but argued that agentic AI introduces a new category of problem: If AI is going to act, then banks need answers to uncomfortable questions. Who delegates decisions to it? Who is accountable for those decisions? How is trust built in a system that may never be perfect?

This feels like the real heart of the current challenge. Banks know how to govern traditional software and control human decision-making through process, policy, and oversight. Agentic AI sits awkwardly between the two. It behaves neither like a normal application nor like a normal employee. That creates friction between product, risk, compliance and operations, and it explains why many AI initiatives appear to stall after early enthusiasm.

Josef added another dimension that many institutions will recognise instantly, even if they rarely state it openly. Anti-innovation incentives are often built into the organisation itself. If status, budget, and influence are still tied to headcount or siloed structures, then leaders are not always rewarded for redesigning work to make teams leaner or decision-making more distributed.

The technology may be new, but the institutional resistance is familiar.

Do not bolt AI onto broken processes

Another strong theme from the interview was the rejection of “incrementalism” for its own sake. Josef argued that trying to introduce AI by slowly modifying existing people-based processes can be painfully slow. In Finshape’s own experience, attempting to layer AI onto running projects and existing ways of working created frustration and poor momentum. The reason is straightforward. If a process was designed for humans in a pre-AI world, simply adding AI to it does not remove the underlying complexity. It often preserves it.

His alternative was to build in parallel. Instead of endlessly tweaking the current process, define the end state, build a semi-automated version beside the old one and switch when it is genuinely better.

That is a more radical recommendation than many banks are comfortable with, but it is strategically important. It suggests that the institutions most likely to create real value from agentic AI will not be the ones that make the smallest safe improvements. They will be the ones willing to design for the target operating model rather than the inherited one.

The real battleground is customer relationships

For all the current emphasis on productivity, both speakers made clear that the bigger long-term question is customer relevance. Josef believes that as agents become widely available to consumers, people will increasingly rely on AI to research, compare and potentially optimise financial choices for them. If that happens, banks face a new strategic question. Where does the customer relationship sit when proactive digital agents start mediating more of financial life?

That is a much more disruptive proposition than using AI to help staff write emails or summarise policy documents. It suggests that banks may need to compete not only on products and channels, but on whether they can remain visible and valuable in a world where the customer’s first point of interaction is an intelligent intermediary.

His practical recommendation was to start not with major life events, but with the everyday interactions customers have through their mobile devices.

Banks have spent years trying to increase engagement, but engagement alone is not enough if what they present is irrelevant. His argument was simple: if AI makes it possible to offer every customer something closer to a private-banking experience, then relevance in daily interactions becomes a much more plausible strategic goal.

Seamless journeys matter more than product silos

Sára made perhaps the most important structural point in the interview. She stated that competitive advantage is not about model choice. Models differ in cost and capability, but they are already good enough for many use cases. The differentiator is increasingly how banks orchestrate specialised agents and how well that orchestration supports a seamless customer journey.

This is where the conversation becomes uncomfortable for many incumbents. Most banks are still organised by product, channel or function.

Customers are not.

A customer may move from a current account issue to a mortgage question to an investment topic without recognising those internal boundaries. Yet those boundaries still shape the journey. Sára’s point was that agentic systems force banks to confront this mismatch because an orchestrator that is supposed to support the whole journey cannot simply hand off at every organisational seam. Banks will need teams that think end-to-end about customer intent, and those teams will need to evolve before the organisation chart does.

Josef reinforced this by noting that AI tends to amplify what is already broken. Weak data quality becomes more obvious. Friction between silos becomes harder to hide. He illustrated this with an example of journey-mapping, where a bank discovered that thousands of customers were receiving an unnecessarily aggressive missed-payment message. That kind of issue can persist for years when institutions operate silo by silo. Once journeys are examined end-to-end, the problems become visible very quickly.

Governance needs standard rules, not bespoke panic

There was also a pragmatic view on governance. Sára did not dismiss the challenge. On the contrary, she described governance as essential. But her answer was not to create a unique governance framework for every workflow or every agent. Her preference was for common standards that apply across AI use cases: traceability, monitoring, evaluation, data quality controls and compliance with requirements such as GDPR and the EU AI Act. Specialised agents tied to specialised domains make this easier because the data boundaries and responsibilities are clearer.

That approach feels sensible for banks that risk overcomplicating AI oversight before they have even scaled the first generation of systems. The real requirement is not bespoke governance theatre. It is a workable set of rules that can be applied consistently and audited properly.

What bankers should take away

The Finshape discussion offers a more credible view of agentic AI than much of the current market commentary. It suggests three practical conclusions for banks:

• Agentic AI is not the same as deploying a copilot. It starts when banks redesign workflows so AI can sequence, judge and act.
• The hardest barriers are organisational. Incentives, accountability and trust are central. The winning institutions will be those that orchestrate specialised agents around customer journeys.
• Banks should be wary of using AI only as a cost-cutting tool. That view is too narrow and will age badly.

For financial services leaders, the most useful message is also the simplest. Agentic AI will not reward cosmetic change. It will expose weak journeys, weak incentives and weak organisational design. That may be uncomfortable, but it is also where the real opportunity lies.

The banks that treat this as a workflow and operating-model challenge, not just a technology project, are most likely to create lasting value.

Join us on May 28 in Brussels to dig deeper into the agentic future with the Finshape team on their booth, where their experts will be ready and waiting to answer all your questions.

You can hear first-hand from the experts in the video below or listen on your favourite podcast platform here.