Sygnum runs live on-chain transactions via AI agent

Sygnum ran live on-chain transactions through an AI agent in a supervised pilot, becoming the first regulated Swiss bank to do so. The bank used its in-house Model Context Protocol (MCP) server to link its systems with Anthropic’s Claude model. Transactions were real and processed under bank oversight while clients retained custody of their private keys on their own devices.

In the pilot, clients submitted plain-language instructions to the AI agent. The agent mapped transaction steps, examined relevant smart contracts and flagged potential risks. It returned the prepared transaction to the client for approval; every signature was completed on the client’s device with a self-custodial wallet so private keys never left the owner’s control.

Sygnum’s AI@Sygnum team developed the MCP server to manage interactions between the bank’s systems and the Claude model. The configuration allowed the agent to handle planning and contract review while preserving client control over signing and custody.

Use cases tested in the pilot included moving stablecoins, executing asset swaps, on-chain lending, wrapping tokens and adding liquidity.

Thomas Frei, head of AI and Data Analytics and lead of AI@Sygnum, described connecting AI agents to wallets as “foundational to where finance is heading,” and said he expects agents to transact, settle and interact with markets on behalf of clients over the next decade.

The pilot aligns with a broader industry effort to move AI agents from advisory roles into transaction execution. In May, Anchorage Digital introduced a platform to let AI agents move funds through regulated banking rails, and FIS has worked with Anthropic to develop agentic AI tools focused initially on anti-money-laundering workflows.

Sygnum framed the project as a controlled test of agent-driven execution in a regulated setting while keeping private keys under direct client control. The bank did not disclose how many clients participated or the total transaction volume processed.