AI agents in business banking are moving from slideware to working software, and the first design question is not what an agent can do. It is where it should live.
In the second episode of TreasurUp Talks, our CTO Marien van Baren sat down with Joost Kevelam to work through that question, building on the agentic business banking whitepaper we published last week.
The short version: an agent is only as useful as the context it sits in, and the safest place for that context is inside the bank. Here is how Marien frames it.
What an AI agent actually does
Start at zero. An agent is software that does not wait to be told every step. It sees the situation, works out what is needed, and prepares the work. Put that next to business banking, the payments, FX, cash and liquidity that companies run every day, and you get agentic business banking: the agent does the legwork around a treasury decision and brings it to a person for approval. The decision stays human. The groundwork does not.
Three layers: composable, intelligent, agentic
TreasurUp’s three-layer platform: composable foundation, intelligence engine, agentic layer.
TreasurUp builds this in order. Layer one is the Composable Banking Platform, the foundation TreasurUp has offered to banking clients for years. A bank selects what it needs: a full suite such as the Liquidity Suite or Risk/FX Suite, a single module, or one service to start with, delivered via web, mobile, embedded in an ERP platform, or as APIs. You can enter anywhere and expand from there.
Layer two is the Intelligence Engine, domain AI offered as a shared platform capability: forecasting, optimization, smart insights, and a natural-language query layer that returns a sourced answer. It is not a dashboard bolted on top. It is baked into every module and service the platform delivers.
Layer three is Agentic Daily Business Banking, the agents themselves. They prepare the action, and a person approves it. The approval gate is not a setting that can be switched off; it is part of the architecture.
Where AI agents in business banking should live
Marien’s principle is blunt: whoever hosts the agent owns the relationship. Put the intelligence inside a treasury management system, and the TMS owns a slice of the company’s financial life. Put it in a generic orchestration tool, and the accountability question gets murky the moment something goes wrong. Who owns the automation flow? Who answers for it?
The bank is already the trust layer. The data, the controls, and the accountability live there. So that is where the agent belongs, the bank’s agent under the bank’s name. That is the white-label vision: the bank owns the agents, the data, the model, and the deployment.
Domain AI, not generic AI
The larger platforms lead with a broad, generic model. They are serious players. TreasurUp’s argument is different. With the treasury domain knowledge built up over years, plus the composable platform and the intelligence layer underneath, an agent can be far more specific, moving a business client from intent to outcome with the right context rather than a generic guess.
Flexibility on the model matters too. The bank picks the brains: a commercial large language model, the bank’s own private model, or a mix. TreasurUp’s orchestration plugs into whatever the bank chooses, across every deployment, because banks are building AI themselves and the agent has to sit inside their existing setup.
Security first: approval-gated execution
Every action with financial, regulatory or accounting impact passes through a logged human approval.
Approval-gated execution is what makes AI agents in business banking safe to deploy, and it rests on three principles.
Approval gates on every material action. The agent proposes; the human disposes. TreasurUp does not expose APIs to execute autonomously.
Model risk built in. Every agent has a well-defined specification. Banks validate it independently, monitor for drift, and can disable, override, or roll back any agent’s decision on their own.
Audit and explainability end to end. Every step is captured. Logs can be bank-owned, and the agent’s reasoning is shown in plain language at the moment of the decision.
Governance is treated as a feature from day one, designed around frameworks like the EU AI Act and DORA. In regulated banking, if governance comes last, you have already lost.
The first use case: relationship manager intelligence
How an RM’s week shifts from admin and preparation toward client dialogue.
The first thing TreasurUp is building is a relationship manager intelligence layer. It watches client behaviour across the platform and tells the relationship manager who needs them and why. Because it sits inside the platform’s data, it sees what outside tools cannot: activity that never becomes a trade, a forecast saved but never executed, a treasury manager who quietly stopped logging in. Those signals reach the RM with the context to act on them.
This matters because of where RM time goes. In many commercial banks, relationship managers spend only 25 to 30 percent of their week in client dialogue; the rest goes to admin, preparation, and internal meetings. Banks putting agentic AI in the frontline report returning 10 to 12 hours per banker per week and lifting client dialogue toward 55 percent. The reported effects: higher revenue per relationship manager, lower cost to serve, and higher coverage ratios. More on this in our note on AI in business banking.
Where banks should start in 2026
Marien’s advice is to build in order. Composable first, so the bank picks where the agent fits. Then intelligence, so the agent becomes useful. Then agentic, because the work moves from clicks to outcomes.
The pressure is real. Treasurers and finance managers are already using AI tools outside the bank, and every one of those is a relationship drifting away. The bank that holds the outcome keeps the relationship. The hackathon TreasurUp ran with its engineering partner Coera made the point concrete, five working agents in 48 hours, but the strategic call is simpler than the engineering. Start experimenting with agentic use cases in 2026, and build them where they belong.
Take it further
Read the full argument in the agentic business banking whitepaper, or book a 45-minute session with TreasurUp to map where an agent could sit in your own channels.
This conversation is from TreasurUp Talks, episode 2, with CTO Marien van Baren. Listen to the full episode.