80% of respondents to the new Bank of England and FCA survey on AI in UK financial services run AI agents. No more than 9% expect to let them pay. The survey came out on 26 October, the regulators' fourth on the subject. 196 firms responded, 105 completed the full questionnaire. It describes the participating firms, not the whole sector. 80% of full-survey respondents have agents deployed, 14% more plan to within a year. Then the autonomy split. 80% use agents for decision support, 46% for constrained execution, 22% for conditional autonomy, 2% for broad autonomy. The payments row reads 7%, 5%, 2%, 0%. Treasury and cash management, 6%, 6%, 0%, 0%. And the line that matters: "even including planned deployment over the next 12 months, no more than 9% of respondents expected agents to" approve or decline customer outcomes, execute trades, "initiate, execute or block payments", or change limits and controls. That is a stated intention, not a ban. Nothing stops an agent paying. The firms are simply not planning to let one. A second finding explains part of why. For high-materiality AI built in-house, 77% rely on deep knowledge of the system. For third-party systems, 24%. The gap is filled by the vendor's word: "81% instead relied on vendor assurance and contractual arrangements." A contract tells you what was promised. It does not tell you what happened. Same pattern in the controls. Pre-deployment review, human approval and vendor due diligence are each above 90%. Ongoing monitoring, around half. A check that runs once binds at the moment it is made. An agent that pays acts after that moment, every single time. So the bottleneck is not the model. It is the mandate. The Bundesbank's September Monthly Report gets there from the payments side. Its authorisation layer decides by whom and under what conditions a payment mandate is granted to which AI agent, and "mandates can only be issued to uniquely verifiable AI agents". What would move that payments row off zero is a delegation someone other than the agent and its vendor can check. Who the agent acts for, a limit it cannot change itself, the condition that releases the money, and a record a third party can read afterwards. That is what we build on @arc , @circle 's Layer 1. @FirmataProtocol composes ERC-8004 identity, ERC-8183 conditional escrow and x402 settlement, so an agent's payment releases on a condition, not on an instruction. Meridian Pay takes USDC and EURC at checkout, Payment Streaming pays by the second, YieldVault holds waiting balances in USYC, not interest on USDC, no rate guaranteed. Firmata makes no one legally accountable, it changes what can be checked. Arc testnet, independent audit before mainnet, soon. Full piece in the Arc House 👇 https://t.co/gO3ZmYoxcC #AIagents #AgenticCommerce