Sygnum pilots AI agents for on-chain transfers with client custody

Sygnum ran a live pilot in Switzerland where an AI agent prepared and orchestrated on-chain transfers while clients kept custody of their private keys. The bank announced the results on May 18, 2026.

Clients submitted plain-language instructions to the agent. The agent generated a step-by-step plan, reviewed relevant smart contracts and flagged potential risks before creating a transaction proposal for client approval.

Every cryptographic signature required to complete a transfer was executed on the client’s device via a self-custodial wallet. Private keys were not transmitted to Sygnum or its servers.

The AI@Sygnum team routed agent interactions through an in-house Model Context Protocol (MCP) server and used Anthropic’s Claude as the underlying model. The agent prompted client approval at each step and did not sign or settle transactions autonomously.

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

Thomas Frei, head of AI and data analytics at Sygnum, predicted that over the next decade agents will transact, settle and interact with markets on clients’ behalf.

Sygnum described the exercise as a pilot and emphasized that custody, consent and signing control remained with clients. The bank said integration work with self-custodial wallets and protocol development was central to the test.

Other firms have begun testing agent-driven transaction workflows. Examples include Anchorage Digital’s Agentic Banking and a partnership between FIS and Anthropic focused on using agents for anti-money-laundering tasks.

According to the bank, the trial represents an operational test of agent-driven on-chain activity within a regulated Swiss banking environment.