Sygnum has executed live, multi-step digital-asset transactions on a public blockchain using an AI agent, becoming the first regulated Swiss bank to test that workflow with client custody and approval controls preserved. The pilot moves AI from advisory support into transaction preparation and execution orchestration, while keeping signing authority with the client.
The architecture is built around the Model Context Protocol, or MCP, with Anthropic’s Claude used as the underlying model. Clients submit plain-text instructions, and the agent plans the transaction path, reviews relevant smart contracts and flags risks before returning each transaction for human approval.
Human Approval Remains the Control Point
The most important design choice is that the AI agent does not hold private keys or sign transactions. Every proposed transaction is signed through the client’s self-custodial wallet on the client’s own device, preserving custody, consent and control.
📣 Sygnum announces the completion of the first live AI-agent driven trial digital asset transactions by a regulated Swiss bank, demonstrating how AI can execute on-chain transactions while clients retain custody, consent and control at every step.
In a first for Swiss banking,… pic.twitter.com/iiqEGPnxaX
— Sygnum Bank (@sygnumofficial) May 18, 2026
The pilot covered stablecoin transfers, asset swaps, on-chain lending positions, token wrapping and liquidity provisioning. That range shows agentic workflows can support real DeFi operations, not only portfolio analysis or chatbot-style banking assistance.
Sygnum’s approach also reflects a clear risk boundary. The agent can translate instructions into executable steps and surface risk signals, but the legal and operational authorization still sits with the client, reducing the chance that an autonomous tool acts beyond the user’s intended mandate.
Banks Face a New Governance Standard for Agentic Finance
The pilot arrives as financial regulators sharpen their focus on AI-driven fraud, operational resilience and accountability. FINMA’s 2026 guidance on digital fraud risk emphasized governance, anti-money-laundering controls and operational risk management for banks using new digital technologies.
European supervisory expectations also point in the same direction: banks and investment firms remain responsible for decisions involving AI tools, including systems developed internally or sourced from third parties.
The Sygnum model creates a practical blueprint. A compliant agentic workflow needs plain-text instruction records, smart-contract review logs, approval evidence, transaction-risk flags and clear separation from custody infrastructure.
The remaining challenge is scale. As agent-mediated transaction volume grows, firms will need real-time monitoring, velocity controls, explainable audit trails and liability frameworks that show who is responsible when an agent misreads instructions, misses a risk or prepares an unsuitable transaction.
Sygnum’s pilot does not remove supervisory risk, but it demonstrates a workable pattern: AI can enter the execution layer without taking custody or bypassing human approval. That distinction will matter as banks, custodians and token issuers prepare for more detailed regulatory scrutiny of agentic finance.