AI for banks: agents that stay within your perimeter

What an agent can prepare for your teams

Repetitive, documented and controllable tasks. The agent does the work of collecting and formatting; the decision remains with the person who makes it today.

Bank secrecy is first and foremost regulated through architecture.

Article 47 of the Banking Act punishes with a prison sentence of up to three years anyone who discloses a secret entrusted to them in their capacity as an officer, employee, or agent of a bank. Therefore, every service provider who sees customer data expands the circle of individuals subject to oversight.

With an on-premise installation, this circle does not widen: the models run on your servers, the requests do not leave your network, and no American provider subject to the CLOUD Act intervenes in the chain. WivenLLM provides the models, AgentHub the agents.

If you prefer external hosting, it will be provided by a Swiss-registered hosting company with servers located in Switzerland. Depending on the function involved, FINMA Circular 2018/3 on outsourcing may then apply: inventory, written contract, audit and access rights for you, your audit firm, and FINMA.

What FINMA expects, and how AgentHub Connect responds

According to the FINMA survey published in April 2025, approximately half of the roughly 400 institutions surveyed already use AI, and 91 of them use generative AI. Data protection and quality, and explainability, are the risks they consider priorities.

In its communication 08/2024, FINMA describes seven areas it examines in its audits. Here's what our tools offer for each:

• Governance. Each agent has a designated manager responsible for scoping. Entries in your systems undergo human validation, and an emergency shutdown cuts off all agents with a single click.
• Inventory and classification of risks. AgentHub Connect lists active agents, their connectors, and their permissions. Risk classification is done with you during the scoping phase, process by process.
• Data quality. The agent works with the sources you provide and cites the document used for each answer. A missing source is reported, not invented.
• Testing and continuous monitoring. Each execution is logged: agent, duration, status, errors. Limits put an agent on pause if it goes outside its parameters.
• Documentation. For each agent: their purpose, the data they read, what they are authorized to write, and their known limitations. A document that your compliance department can use.
• Explainability. The agent indicates the documents and rules on which he relies. The analyst sees why the agent proposes a result, not just the result itself.
• Independent journal. The run log can be made available to your internal audit and your auditing firm. It does not retain customer data.

The responsibility remains with the bank: our tools provide the means of control, not compliance itself. Details are on the page Governance and monitoring.

Start with a scope that your compliance department accepts.

1. Framing. A precise process, the data involved, the level of risk, and what your compliance department needs to validate. The timeline is set together at that time.
2. Installation within your area. WivenLLM and AgentHub Connect on your servers, or with a Swiss-registered hosting provider if you prefer. Initial tests on fictitious or anonymized data.
3. Read-only driver. The agent prepares, your teams review each output. We measure the time saved and the errors on real cases.
4. Extension. If the pilot is convincing, the agent can write in your tools, always with human validation where you want it.