The year is 2030, and a customer’s personal finance agent detects that their energy tariff is about to renew at an unfavourable rate. It contacts the supplier’s sales agent, negotiates a new deal, and executes the switch, all without the customer even being aware. No form was filled in, no one reviewed the terms and conditions and no human decision was made at the point of transaction.
This is the soon-to-be reality of agentic AI. And while it’s going to be a game-changer for efficiency, it risks causing a compliance nightmare. But only if we’re unprepared.
The identity problem
The entire architecture of financial compliance rests on the assumption that there is a human being at the source of every transaction. Know Your Customer (KYC) exists because of that assumption. Accountability frameworks such as the Senior Managers and Certification Regime (SMCR), liability chains, and consent obligations exist because of it. If we remove the human from the equation, a significant amount of what we have built starts to unravel at an alarming rate.
The issue is not that AI systems make mistakes. Humans are equally fallible. The challenge is that the mechanisms we use to establish intent, verify authority, and assign accountability were designed specifically around human actors. When an autonomous model negotiates a financial agreement on instructions set three months ago, there is no meaningful way to establish whether those instructions still reflect what the customer would want today. KYC cannot answer that question because it was never asked to.
Faster than compliance can handle
Part of what makes agent-to-agent finance so difficult to govern isn’t the complexity of any single transaction, but the speed at which sequences take place. Agent-to-agent exchanges can execute in milliseconds.
Let’s say a lending agent agrees repayment terms with a creditor’s agent, but the terms fall marginally outside the customer’s pre-authorised range. The model, however, deems that margin acceptable. No human has reviewed it, and by the time a compliance officer flags it, the agreement has settled and cascaded into downstream positions. The accountability gap is not just about who is responsible, but also that the gap widens faster than any human oversight process can close it.
KYC was designed for a world where humans set parameters and machines executed within them – not one where machines interpret, negotiate, and act autonomously.
KYA: the framework the agentic economy needs
What is needed is not a wholesale reinvention of compliance, but a deliberate extension of its core principles to a new category of actor. Enter Know Your Agent or KYA.
The logic for KYA mirrors KYC closely by design. Just as a financial institution must verify a customer’s identity, confirm their authority to transact, and maintain an auditable record of that relationship, the same obligations should apply to any AI agent acting within the financial system. Every agent should carry a credential issued by the deploying institution, scoped to specific transaction types and value thresholds, and revocable in real time at the customer’s instruction. The agent’s mandate should be as legible to a compliance team as a customer’s account profile.
Where KYA diverges from KYC is in the technical layer that must sit beneath it. Verification alone is not enough when the actor in question can reason probabilistically and act in milliseconds. The architecture must do some of the compliance work.
Separating reasoning from execution
The most important structural principle for agentic finance is the separation of the reasoning layer from the execution layer.
An AI agent can negotiate freely, assess options, model outcomes and agree terms, but the movement of money, data, liability, credit, consent, regulatory standing, and reputational exposure should pass through a system that enforces hard limits independently of whatever the reasoning model has concluded. If an agent agrees to a payment that exceeds a pre-set threshold, the execution layer rejects it. Not because the agent’s reasoning was necessarily wrong, but because human-set limits should be non-negotiable by the system.
This should not be viewed as a restriction of what AI agents can do, but rather as the architectural condition under which autonomous operation becomes trustworthy at all. Without it, every AI agent deployed in a financial context is effectively operating on an honour system, and compliance built on an honour system is not viable.
Being on the loop is not the same as being in it. For transactions above a defined value or complexity threshold, a human sits at the execution boundary not to replicate the agent’s reasoning, but to retain the accountability chain that regulators will require firms to demonstrate. The agent handles the cognitive load: the negotiation, the modelling, and the decision. The human retains only the right of refusal, and only where the stakes warrant it.
The firms that move first will set the terms
Regulation on agentic AI in financial services is coming. The real question is whether the industry helps shape it or inherits whatever framework gets imposed in the absence of better options.
The firms that build KYA infrastructure now, including verification credentials, execution controls, and human oversight at the boundary, will not be scrambling when the rules arrive. They will have already solved the problem regulators are trying to address, which is the most reliable way to influence how those rules get written.
