Building the Agentic Bank: what needs to be true

The Agentic Bank isn't a destination. It's a direction

Part three of a three-part series on the agentic bank written by Oper’s CEO and Co-Founder Geert van Kerckhoven.

In the first two parts of this series, I laid out the opportunity: a front office redesigned around agent-to-agent interaction, a mid-office and back-office where AI delivers immediate ROI, and a lifecycle management layer where the cost-of-tokens thesis unlocks product categories that simply didn't exist before.

But a blueprint is only useful if you can execute on it. So let me close with what actually needs to be true.

Model selection: three capabilities that matter

Not all models are created equal for lending. When you're evaluating your AI architecture, three capabilities matter most.

Multimodal data handling. Mortgage origination is inherently multimodal - scanned payslips, handwritten annotations, photographed IDs, structured XML data from registries, free-text emails from brokers. Your model stack needs to handle all of it natively, not through a patchwork of specialized tools.

Reasoning with auditability. In a regulated environment, it's not enough for a model to reach the right conclusion. You need to demonstrate how it got there. Reasoning chains must be auditable, verifiable, and non-discriminatory. This isn't a nice-to-have - it's a supervisory requirement.

Tool use and computer use. The agent needs to interact with your systems - API gateways, core banking platforms, document management systems, external databases. Sometimes it also needs to use vision to navigate mainframe interfaces that were never designed for machine interaction. Models that excel at tool use can actually operate within your technology landscape.

Sovereign vs. Cloud: a European question

European banks face a deployment question that their US counterparts largely don't: where does the data go?

The good news is the open-source model ecosystem has matured to the point where you have genuine choices. Models can run in private European cloud infrastructure with full data residency guarantees. You can mix and match: sovereign deployment for sensitive processing, cloud for non-sensitive tasks, with clear data governance at every boundary.

Your AI roadmap should map which models can be deployed in a sovereign way, which in a non-sovereign cloud, and which tasks require which posture. This isn't a one-time decision - it's a living architecture that evolves as models improve and regulatory requirements clarify.

EU AI Act: when do the rules for credit scoring actually apply?

If you use AI for credit scoring or loan approval in the EU, you are operating a high-risk AI system under Annex III of the EU AI Act. When this article first appeared in May 2026, those obligations were due to apply from 2 August 2026. That date has moved.

The Digital Omnibus on AI, adopted by the Council on 29 June 2026 and published as Regulation (EU) 2026/1744, sets two fixed dates:

The obligations have not changed: conformity assessments, a quality management system, registration in the EU database, documentation of training data, risk management procedures, human oversight, and ongoing monitoring for accuracy, robustness, and non-discrimination. Penalties still reach €35 million or 7% of global turnover.

The EBA has noted that many of these requirements can be integrated into existing supervisory review and evaluation procedures. But "can be integrated" is not the same as "will be easy." Fifteen extra months is preparation time. Banks that file December 2027 under "later" will arrive there in the same state they would have arrived at August 2026.

The bank that moves first

The agentic bank isn't a destination. It's a direction - and a compounding one. Every dossier processed builds institutional knowledge. Every edge case resolved deepens the agent's expertise. Every workflow automated frees capacity for the next initiative. The banks that start now build advantages that compound, while the banks that wait will find the gap widening every quarter.

The blueprint laid out here - from borrower-facing agents through AI credit committees to dynamic loan products - isn't science fiction. The technology exists. The regulatory frameworks, while demanding, are navigable. The economics are overwhelming. What's missing, in most institutions, is the architectural vision and the organizational will to pursue it front-to-back rather than piecemeal.

We see a future where mortgage lending is faster, more flexible, more affordable, and more responsive to borrowers' lives. Where processing costs don't dictate the boundaries of product design. Where agents handle the volume, and humans handle the judgment. Where the front office, the back office, and the lifecycle are connected by a single intelligence layer that learns, adapts, and improves with every interaction.

That future isn't coming. For the institutions that are paying attention, it's already here.