Sygnum said it has completed a pilot involving live digital asset market transactions executed through an AI agent, describing itself as the first regulated Swiss bank to test this setup. The system turns plain-language client instructions into prepared multi-step blockchain transactions on Mainnet, but private keys remain with the client, and final signing is still done through self-custodial wallets on client-controlled devices.
The agent can prepare transfers, swaps and on-chain positions
According to the bank, the AI agent goes beyond analysis. It can interpret text instructions, map out execution paths, and prepare workflows involving stablecoin transfers, asset swaps, token wrapping, liquidity provisioning, and on-chain lending positions. Before anything is submitted for approval, the system reviews smart contracts, evaluates transaction structures, flags potential risks, and presents an execution plan to the client.
Thomas Frei, Head of AI and Data Analytics and AI@Sygnum lead at Sygnum Bank, said connecting AI agents to wallets is foundational to the direction of finance. He said the next decade will bring agents that transact, settle, and interact with markets on behalf of clients, while the harder task is preserving bank-grade consent, custody, and trust.
Human approval stays at the center of the workflow
Sygnum described the architecture as a “human-in-the-loop” model. In practice, that means AI handles workflow preparation, while transaction authorization stays separate and under direct client control. The bank said AI is meant to augment human decision-making rather than replace client oversight or trusted banking relationships. Every on-chain action still requires explicit approval before execution.
That distinction matters in the current regulatory climate. The report notes that regulators and banking supervisors are assessing AI governance, accountability, operational resilience, explainability, and control over automated financial activity. Sygnum said its governance framework covers data management, transparency, accountability, operational resilience, and risk oversight in line with regulated banking standards.
MCP server and Claude power the pilot
The pilot was built on an internal Model Context Protocol, or MCP, server developed by the AI@Sygnum team and powered by Anthropic’s Claude model. MCP is an emerging open standard intended to help AI systems and financial platforms exchange structured operational context more efficiently.
Sygnum said this setup allows the AI system to understand transaction environments, wallet states, smart contract interactions, and workflow requirements while keeping execution safeguards in client hands. The bank also described the infrastructure as model-agnostic and asset-class agnostic, suggesting it could scale across different AI systems and financial products over time.
The pilot points to a broader direction in regulated finance: institutions want AI embedded more deeply into transaction and settlement infrastructure, but without handing over full autonomous control. Sygnum’s approach keeps that boundary clear by letting AI orchestrate blockchain workflows while clients retain control over assets and final execution.