Switzerland-based digital asset banking group Sygnum has announced it has become one of the first Swiss-regulated banks to use an AI agent to test live digital asset market transactions, with clients retaining custody, consent, and control at every step. The pilot was conducted under controlled conditions using an in-house Model Context Protocol (MCP) server built by Sygnum's AI team, with Anthropic's Claude as the underlying AI model.
In the pilot, client-issued plain text instructions were used to execute multi-step live transactions on a blockchain mainnet. The AI agent independently planned and prepared each step, reviewed relevant smart contracts, and flagged potential transaction risks before presenting each transaction to the client for approval. Transactions were signed exclusively through the client's self-custodial wallet on their own device, ensuring private keys did not leave client control at any point.
Scope and technical architecture
The Sygnum AI agent is capable of executing multi-step on-chain transactions covering stablecoin transfers, asset swaps, on-chain lending positions, token wrapping, and liquidity provisioning. The MCP-based architecture is designed to be model-agnostic and asset-class agnostic, allowing the infrastructure to scale as the digital asset ecosystem develops. MCP is an open standard enabling AI and financial platforms to share context and data while maintaining security safeguards and client control.
Human-in-the-loop design and governance
Sygnum's approach differs from architectures in which AI agents hold their own wallets and transact autonomously. Under Sygnum's model, clients own, self-custody, and fully control their wallets and assets at all times, with the AI agent augmenting rather than replacing human decision-making. This human-in-the-loop design is described as central to Sygnum's broader AI strategy, which is structured around four pillars: client experience, regulated innovation, organisational efficiency, and operational excellence.
The governance framework is designed to meet regulatory standards covering data, risk management, transparency, and accountability. Sygnum noted that regulators increasingly view the potential for AI agents to act beyond a client's intended scope as a risk for financial institutions, and that its approach was designed to address that concern directly.
The MCP AI agent is not currently available to clients. Production deployment will be subject to full regulatory, compliance, and security reviews and approvals.
Industry context
The pilot is among the first demonstrations by a regulated bank of an AI agent executing live on-chain transactions within a framework that maintains client custody throughout. As agentic AI capabilities advance across financial services, the question of how to preserve regulatory compliance, client consent, and asset security within agent-driven workflows has become a central challenge for institutions seeking to deploy these capabilities at scale.
Sygnum's approach, which uses client-controlled self-custody wallets and requires explicit approval at each transaction step, offers a model for how regulated financial institutions might integrate AI-driven execution into digital asset operations without ceding control of client assets to an autonomous system.
