Sygnum Tests AI Agent for Live On-Chain Transactions

Sygnum Executes Live On-Chain Transactions via AI Agent – Clients Retain Full Custody of Private Keys

Key Takeaways

• Sygnum has completed live on-chain transactions using an AI agent within a regulated Swiss banking framework.
• Clients retain full custody, as private keys never leave their own devices.
• The system uses Sygnum’s in-house Model Context Protocol server powered by Anthropic’s Claude.
• Use cases tested include stablecoin transfers, token swaps, on-chain lending, token wrapping, and liquidity provision.
• Other financial institutions, including Anchorage Digital and FIS, are also integrating agentic AI into transaction and compliance workflows.

Sygnum Deploys AI Agent for Live On-Chain Transactions

Sygnum has announced the completion of what it describes as the first live AI agent driven digital asset transactions conducted by a regulated Swiss bank. The pilot project moved artificial intelligence beyond advisory functions and into the execution layer of blockchain based transactions.

According to the bank, clients initiate the process by submitting a plain language request. The AI agent then structures and plans the required transaction steps. Before execution, it reviews the relevant smart contracts and flags potential risks. The final transaction is returned to the client for approval.

A central feature of the setup is custody control. All signatures take place on the client’s own device through a self custodial wallet. Private keys never leave the client’s possession. This means the AI agent can prepare and structure transactions but cannot independently move funds without user authorization.

For users who interact with blockchain based services, the distinction between transaction automation and key control is significant. The model tested by Sygnum separates transaction planning from key custody, maintaining user control at each stage.

Technology Stack: Model Context Protocol and Claude

The pilot was conducted by the AI@Sygnum team using the bank’s in-house Model Context Protocol, or MCP, server. The underlying artificial intelligence model is Anthropic’s Claude.

Within this framework, the AI agent interprets natural language instructions and translates them into structured blockchain actions. The system is designed to analyze smart contracts before execution and highlight risks to the user. The final step requires explicit approval, with transaction signing carried out locally on the user’s device.

By keeping the signing process client side, Sygnum’s model avoids transferring private keys to bank controlled infrastructure. This structure is intended to ensure that custody, consent, and control remain with the client throughout the transaction lifecycle.

Thomas Frei, Head of AI and Data Analytics and AI@Sygnum lead at Sygnum Bank, stated that connecting AI agents to wallets represents a foundational step for the future of finance, where agents transact and interact with markets on behalf of clients.

Tested Use Cases: From Stablecoins to Liquidity Provision

The live trial covered several types of on-chain transactions. These included moving stablecoins, conducting asset swaps, participating in on-chain lending, wrapping tokens, and adding liquidity.

Each of these activities typically requires interaction with smart contracts and decentralized protocols. In standard workflows, users must manually review contract addresses, confirm transaction details, and manage gas fees. In Sygnum’s pilot, the AI agent handled the planning and contract review stages before presenting the finalized transaction for approval.

For users of crypto based financial services, including those who move funds between wallets, exchanges, or decentralized platforms, automation at this level could alter how transactions are initiated and reviewed. However, in the tested model, the AI did not remove the need for user consent or direct signature.

Banks and Digital Asset Custodians Expand Agentic AI Use

Sygnum’s pilot reflects a broader shift within regulated financial institutions. Over recent months, banks and digital asset custodians have sought to move AI agents from advisory roles into transaction execution.

Anchorage Digital introduced what it calls Agentic Banking in May, a platform designed to allow AI agents to move funds through regulated banking rails. Separately, FIS and Anthropic have partnered to integrate agentic AI into banking processes, beginning with a Financial Crimes AI Agent focused on anti money laundering work.

These developments show that AI integration in financial services is extending beyond data analysis and customer support. Institutions are now testing AI systems that interact directly with transaction systems, whether on blockchain networks or through traditional regulated payment infrastructure.

For users of crypto services, including those evaluating platforms for trading, lending, or betting, the integration of AI agents into transaction layers may affect how orders are prepared, reviewed, and executed. In regulated environments, the emphasis remains on maintaining compliance controls and documented user consent.

Implications for Custody and User Control

A key aspect of Sygnum’s implementation is the preservation of self custody. The bank structured the system so that private keys remain on the client’s device at all times. The AI agent does not independently sign or broadcast transactions.

This architecture addresses one of the primary operational risks in digital asset management, namely the handling of private keys. By separating transaction intelligence from key storage, the pilot demonstrates a model where automation and user control coexist.

For market participants who rely on secure wallet infrastructure, including those engaging in decentralized finance activities, the custody model remains central. The Sygnum trial illustrates how AI agents can be integrated without transferring key ownership to a third party.

Our Assessment

Sygnum has completed a live pilot in which an AI agent executed on-chain transaction workflows within a regulated Swiss banking environment while clients retained full custody of their private keys. The system uses an in-house protocol powered by Anthropic’s Claude and requires user approval for every signed transaction. Similar initiatives by Anchorage Digital and the partnership between FIS and Anthropic indicate that financial institutions are expanding AI applications from advisory and compliance functions into transaction execution layers. The development reflects a broader institutional effort to integrate agentic AI into regulated digital asset operations while maintaining user consent and custody control.