The UK Financial Conduct Authority has flagged that autonomous AI agents operating within financial systems could trigger substantial regulatory upheaval, particularly as tokenized assets and programmable money gain traction.
The FCA's recent guidance signals growing concern about the intersection of agentic AI—software that operates independently to execute financial transactions—and the emerging ecosystem of on-chain assets. This combination creates novel risks the regulator views as systemically important.
The core issue centers on autonomy and control. When AI agents manage transactions involving tokenized money or assets, traditional regulatory frameworks struggle to assign accountability. Current rules assume human decision-makers at critical junctures. Autonomous agents collapse that assumption. If an AI protocol executes a trade, initiates a transfer, or manages liquidity without direct human instruction per transaction, liability chains become murky.
The FCA emphasizes that tokenized assets amplify these concerns. Unlike traditional digital representations of value, tokenized assets operate natively on blockchain infrastructure, often with built-in programmability. Layer in autonomous agents, and financial flows can happen at machine speed with minimal human oversight.
The regulator's warning arrives as financial institutions explore both AI automation and tokenization independently. Banks pilot AI for trading and settlement. Central banks test digital currencies. The FCA appears concerned that convergence of these trends outpaces regulatory capacity.
Key pressure points include market manipulation risks, liquidity mismanagement, and contagion. An autonomous AI agent operating across multiple tokenized asset pools could amplify volatility across interconnected protocols without adequate safeguards. One agent's miscalibration could cascade through linked systems.
The FCA also flags governance questions. Who controls the AI? What happens when it malfunctions? Can regulators effectively supervise software behavior in real time? Traditional licensing and authorization regimes assume identifiable institutions making decisions. Decentralized protocols running autonomous agents challenge these assumptions fundamentally.