Corrected on 1 November 2026. The version published on 7 July 2026 got several things wrong.
- EU AI Act. We said its high-risk rules take full effect in August 2026. Annex III rules apply from 2 December 2027 and Annex I rules from 2 August 2028. We also overstated their reach.
- Accountability. We said a named senior manager is personally accountable for an agent's decision. The test is reasonable steps, and SM&CR doesn't cover firms authorised only as payment or e-money institutions. We also called agent identity and authority existing regulatory expectations. They aren't yet.
- Attribution. We overstated what Sheldon Mills and the Bank of England had said, including on human-in-the-loop, sampling and kill switches.
- Figures and sources. We misdescribed the 2% figure from the Bank and FCA's 2024 survey. Weak secondary sources are replaced or cut, and vendor names are gone.
We've also added what has happened since. The argument and the five questions stand.
Compliance has spent thirty years building an identity architecture around two kinds of actor: natural persons and legal entities. KYC for the humans. KYB for the companies. Every control, every audit trail, every regulatory framework assumes that behind every action in a financial system there is one of those two things — a person or an entity — that can be identified, verified, and held accountable.
That assumption is now breaking.
AI agents — systems that can act, not just recommend — are entering financial workflows at speed. Firms are building them to review KYC cases, triage AML alerts, unwrap UBO structures, initiate payments, and interact with customers. Risk.net reported in April 2026 that JP Morgan was building an agentic onboarding system that would cut a process taking up to five days to less than a minute. (Source: Risk.net, April 2026) ING said in 2025 that it was redoing its client due diligence workflows with AI, drawing on public sources and behavioural data rather than asking clients for every data point. (Source: Computer Weekly, July 2025) Gartner expects 40% of enterprise applications to embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. (Source: Gartner, August 2025)
Which raises a question that most compliance frameworks cannot currently answer: when an agent takes an action in a regulated workflow, who took that action?
This is the identity problem that "Know Your Agent" — KYA — exists to solve. And for compliance teams, it's arriving faster than the frameworks are.
A third category of actor
The reason KYA matters isn't that AI agents are new technology. Compliance has absorbed new technology before. It's that agents break the identity model itself.
A traditional automated system — a screening engine, a rules-based monitoring tool — is a deterministic instrument. It does exactly what its configuration says, every time. When it acts, the "who" is straightforward: the institution that configured it. The system is a tool, and tools don't have identity problems.
An agent is different in kind. It plans. It chooses between approaches. It calls other tools, queries data sources, and takes multi-step actions toward a goal — often in ways that weren't explicitly scripted. Two runs of the same agent on the same case may take different paths. That autonomy is exactly what makes agents useful, and exactly what makes the old identity model insufficient.
When an agent clears an AML alert, the question "who cleared this alert?" now has several possible answers: the institution that deployed the agent, the vendor that built it, the model provider underneath it, the employee whose instruction triggered it, or the agent itself. In an examination, "the system did it" has never been an acceptable answer — and with agents, working out what the acceptable answer even looks like requires infrastructure most institutions haven't built.
KYA is that infrastructure. The term is a market label, not a legal one, but in its emerging form it combines four things: verified agent identity (what is this agent and who operates it), authority binding (what is it permitted to do, and on whose behalf), runtime controls (what stops it exceeding that authority), and tamper-evident audit (a complete, provable record of what it did and why). The FCA-commissioned Mills Review reached for a similar list: identity linkage, verifiable mandates covering scope, limits and revocation, and standards for evidencing control and audit. (Source: The Mills Review, pp 107–108, July 2026)
If that structure sounds familiar, it should. It's the same conceptual shape as KYB: verify the entity, establish who controls it, understand what it's authorised to do, monitor it continuously. The phrase "Know Your Agent" fits because the problem rhymes with the ones compliance already knows how to solve.
The market is moving before the regulation
What makes this moment unusual is that the commercial infrastructure is being built ahead of any regulatory mandate — a reversal of how compliance frameworks normally develop.
Across the identity and payments ecosystem, KYA frameworks are already taking shape. Some are built around the concept of a verifiable agent passport — a tamper-resistant identity token that ties an agent to the verified business behind it, extending KYB principles to a new category of actor. The card schemes have built their own programmes. Mastercard's Agent Pay, launched in April 2025, requires agents to be registered and verified before they can pay. (Source: Mastercard, April 2025) Visa's Trusted Agent Protocol, launched that October, helps merchants tell legitimate agents from malicious bots. (Source: Visa, October 2025) In September 2026 the two schemes and Ant International said they would explore common principles for KYA, with each network keeping its own verification and decisioning. (Source: joint release, September 2026)
A further school of thought centres on human binding — the principle that an agent's activity should be explicitly authorised by a verified real person, with authorisation and responsibility always traceable to that human, verifiable dynamically when risk arises.
These are different architectures answering the same underlying question: how do you extend identity, authority, and accountability to a category of actor that is neither a person nor a company? No single standard has been established yet, and multiple protocols coexist. The FIDO Alliance only began work on agentic standards in April 2026, with Mastercard and Visa chairing its payments group. (Source: FIDO Alliance, April 2026)
For compliance leaders, the takeaway isn't to pick a winner. It's to recognise that agent identity is becoming a real infrastructure layer, and the institutions that understand it early will be the ones writing their own requirements rather than inheriting someone else's.
The regulatory position: no new rules, but no escape either
The UK regulatory position on AI is often summarised as "no AI rulebook" — and that's accurate, but dangerously incomplete.
The FCA has been consistent that AI in financial services sits inside existing frameworks: the Consumer Duty, the Senior Managers and Certification Regime (SM&CR), operational resilience. It still doesn't plan AI-specific rules. (Source: FCA, AI approach) What's less appreciated is how demanding that position actually is. The FCA's October 2026 good practice paper says firms are responsible for the AI systems they develop, procure and deploy, including their outputs and outcomes, and that they can outsource AI services but remain responsible for them. The paper creates no new requirements, but it shows what the FCA thinks good looks like. (Source: FCA, AI in financial services: good practice and areas for improvement, October 2026) In banks and other firms under the enhanced regime, any use of AI in an activity or business area falls within a senior manager's responsibilities. (Source: FCA AI Update, paras 3.40–3.41, April 2024) David Geale, the FCA's Executive Director of Payments and Digital Finance, told the Treasury Committee that somebody is "on the hook" through the senior managers regime, and that not understanding what you've deployed is no defence. (Source: oral evidence, Q140 and Q160, 15 October 2025) Delegating a decision to an algorithm doesn't transfer regulatory responsibility to the algorithm or the vendor.
Translate that into the compliance context. Suppose an AI agent approves a corporate onboarding case that later turns out to involve a sanctioned UBO. The firm answers first: for a breach of financial sanctions, the Office of Financial Sanctions Implementation can impose a civil penalty without showing that the firm knew or suspected anything. (Source: Policing and Crime Act 2017, s.146) The FCA can look at its systems and controls. And at a firm within SM&CR, it can look at the senior manager responsible for that area. That isn't liability for every decision. The test is whether they took the steps a person in their position could reasonably be expected to take to avoid the breach, and the FCA says it doesn't expect senior managers to be technical specialists. (Sources: FSMA 2000, s.66A(5); FCA good practice paper, section 3.3) But consider what they'd have to show, about a decision they didn't make, didn't review, and may not be able to fully reconstruct. The Mills Review sees the same risk: without guidance, more autonomous AI could open a gap between outcomes and anyone's ability to identify who was responsible or to exercise meaningful control. (Source: The Mills Review, p 81) That is the accountability gap KYA exists to close.
SM&CR doesn't cover everyone this post is written for. It applies to firms authorised under FSMA, so a firm authorised only as a payment or e-money institution sits outside it, and HM Treasury asked in July 2026 whether stronger senior manager accountability would help in that sector. (Sources: HM Treasury, August 2023, para 2.91; HM Treasury consultation, July 2026, question 19) That isn't a free pass. An institution that outsources an important operational function, including an IT system, must not delegate senior management's responsibility (Payment Services Regulations 2017, reg 25; Electronic Money Regulations 2011, reg 26), and the Money Laundering Regulations call for a senior officer responsible for compliance, where appropriate to the firm's size and nature, and a nominated officer (reg 21).
And the ground is shifting further. Sarah Breeden, the Bank of England's Deputy Governor for Financial Stability, said in June 2026 that regulatory frameworks weren't built with autonomous agents in mind, that a human in the loop for all agent actions is unlikely to be realistic, and that more sophisticated governance and accountability frameworks may be needed. (Source: Bank of England, Agents of change, 30 June 2026)
The Mills Review, published on 6 July 2026, recommends that the FCA consider leading a trusted framework for how agents are authorised, identified and held accountable. (Source: The Mills Review, p 107) On 19 October Jessica Rusu, the FCA's chief data, information and intelligence officer, said it is embedding six of the review's seven recommendations into live work, and that it will work with industry, other regulators and frontier labs on a trusted framework for agentic finance, covering agent identity, delegated authority, accountability and execution. That was a speech, not a rule, and it gave no date. (Source: FCA, Building the future of finance in the age of AI, 19 October 2026)
Payments are moving too. The FCA's March 2026 priorities said it would work with the Treasury on modernising payments regulation, including considering whether regulation needs to change to support agentic AI payments. (Source: FCA, Regulatory Priorities: Payments, p 7) HM Treasury then consulted on how payment rules should adapt, including authentication, consent and liability for unauthorised transactions. That consultation closed on 6 October 2026, and no rule has changed yet. (Source: HM Treasury, Modernising Payment Services Regulation, question 15)
In the EU, the AI Act's high-risk rules don't apply yet. Annex III rules apply from 2 December 2027 and Annex I rules from 2 August 2028. (Source: Regulation (EU) 2026/1744) They will catch a UK firm that places a high-risk system on the EU market, or whose high-risk system's output is used in the EU, but they don't reach every system that touches an EU customer. (Source: AI Act, Article 2) In financial services that chiefly means credit scoring of individuals and life and health insurance pricing, and the Commission's draft guidelines put AML/CFT systems outside the credit scoring category unless they're also intended to assess creditworthiness. (Sources: AI Act, Annex III, point 5; draft Commission guidelines, para 309) For onboarding, the nearer date is 10 July 2027, when the EU's AML Regulation will require meaningful human intervention in any automated or AI-driven decision to enter, refuse or maintain a business relationship, and let the customer ask for an explanation. (Source: Regulation (EU) 2024/1624, Article 76(5)) And part of the AI Act already applies to agents. Article 50's transparency duties have applied since 2 August 2026, and the Commission's guidelines say an agent that interacts with people must be designed to disclose that it's artificial and on whose behalf it's acting. (Source: Commission guidelines, para 31, July 2026)
The direction of travel is unmistakable. Not an AI rulebook — but a rising evidentiary bar for exactly the questions KYA answers: what did the agent do, under whose authority, and how do you prove it?
What this means operationally
Here is where the gap between AI adoption and AI governance becomes concrete. In the Bank of England and FCA's 2026 survey, 86% of the 196 firms that responded said they use AI. Among the 105 that completed the full survey, 80% reported AI agents deployed, but autonomy was limited: 80% had agents giving decision support, where the agent recommends and a human decides, 46% had agents executing within predefined limits, and only 2% had agents with broad autonomy. Even counting plans for the next 12 months, no more than 9% expected agents to approve or decline customer outcomes. These are participating firms, not the whole sector. (Source: Bank of England and FCA, Artificial intelligence in UK financial services – 2026, October 2026)
Part of what's holding autonomy back is practical: integration and legacy systems were the most-cited non-regulatory constraints. But the harder limit, in my experience, is that most institutions cannot yet answer the governance questions that come with it. Having operated compliance workflows through several generations of automation, these are the five questions I'd expect any examiner — or any second line worth its salt — to ask about an agent in a regulated workflow:
1. What is this agent's identity and provenance? Not "we use an AI tool" — specifically: which agent, which version, built by whom, running on which model, changed when. A compliance decision made by version 2.3 of an agent needs to be attributable to version 2.3, not to a vague notion of "the system." Model updates change behaviour; if you can't tie decisions to versions, you can't explain historic decisions at all. The Financial Stability Board's October 2026 sound practices, which set no international standard, point the same way: unique identifiers, so a firm can see which actions each agent took and which humans configured and oversaw it. (Source: FSB, p 26)
2. What is it authorised to do — and where is that written down? An agent that can gather registry data, assemble a case file, and draft a recommendation is one risk profile. An agent that can approve onboarding, clear alerts, or file a SAR is a fundamentally different one. The boundary between "prepares the decision" and "makes the decision" needs to be explicit, documented, and enforced at the system level — not implied by how the team happens to use the tool. The FCA's October paper describes firms registering agents in a central inventory showing who owns them, what they can access and what controls apply. (Source: FCA good practice paper, section 3.5)
3. Who is the accountable human, per decision? SM&CR doesn't answer this for you: it names someone for each area of the business, asks whether they took reasonable steps, and doesn't reach every firm. I'd still want every agent action in a regulated workflow to map to a named individual with genuine oversight capability, because that mapping is how you prove those steps. The FCA's October paper describes one firm that mapped every material AI use case to a named senior manager, and another that recorded whether staff needed to review AI outputs, whether they did, and whether they approved, overrode or escalated. (Source: FCA good practice paper, section 3.3) "The head of compliance owns AI" is a statement of intent, not a control. The mapping has to survive contact with a specific case: this decision, this agent, this accountable person, this evidence of oversight.
4. What does the audit trail actually capture? An activity log ("agent accessed Companies House at 14:32") is not a decision trail. A decision trail shows what the agent was asked to do, what information it gathered, what it concluded, and why. No rule I've found sets that standard for agents yet, but the direction is clear. The Mills Review says that in regulated activity a useful answer isn't enough if its basis can't be reconstructed, and asks the FCA to consider standards for evidencing how agents behave. (Source: The Mills Review, pp 17 and 108) The FSB describes firms requiring agents to summarise their actions and reasoning, and says the documentation may include the reasoning at each autonomous decision point. (Source: FSB, p 53) Sampling still has a place — the FCA lists spot checks and sample reviews among the monitoring it has seen work — but a sample is only as good as the record behind it. For systems making thousands of decisions a day, I'd expect the question to move from "what did you spot-check?" to "show me this one."
5. What are the boundaries and the kill switch? What happens when the agent encounters a case outside its competence? Does it escalate, or does it guess? Can you halt it immediately if behaviour drifts? I've found no UK rule that requires a kill switch for agents, but the idea keeps recurring. The FSB's sound practices include implementing and testing risk-based kill switches that can pause, limit, override or remove an agent's access. (Source: FSB, p 53) The Monetary Authority of Singapore's guidelines, which take effect in October 2027, expect tested activation protocols for AI with kill switches. (Source: MAS, para 5.3) At the Bank of England, Sarah Breeden has said simulation work could explore whether trading needs market-wide guardrails analogous to circuit breakers or kill switches. (Source: Breeden, June 2026) An agent without a tested stop mechanism is an operational resilience finding waiting to happen.
The uncomfortable parallel
There's a pattern here that anyone who has worked in compliance operations will recognise, because we've lived through it once already.
When KYC automation arrived, institutions bought tools faster than they built governance. The result was a decade of remediation: lookback exercises, s166 reviews, fines for controls that weren't doing what firms assumed they were. Starling's automated sanctions screening had been checking customers against only a fraction of the sanctions list since 2017, which the bank discovered in January 2023. It was fined £29 million in 2024. Monzo was fined £21 million in 2025. In both cases the FCA said financial crime controls had failed to keep pace with growth. At root, those were governance failures. (Sources: FCA, Starling, October 2024; FCA, Monzo, July 2025)
Agentic AI is arriving with the same adoption curve and higher stakes, because this time the technology can act, not just process. The institutions that treat agent governance as a launch prerequisite will avoid the remediation cycle. The ones that bolt it on afterwards will relive it, with personal accountability attached wherever SM&CR applies. The FCA already lists retrospective governance among the things it wants firms to improve.
The good news is that compliance teams are better positioned for this than they think. The conceptual toolkit already exists — it's the same one used for KYB. Verify the actor. Establish who controls it. Define what it's authorised to do. Monitor continuously. Keep evidence that survives examination. The actor is new. The discipline isn't.
Where this leaves compliance teams
Know Your Agent is not yet a regulatory requirement, and nobody should pretend otherwise. No UK or EU rule requires it by name. The Mills Review notes that the UK doesn't currently define agent identity or require agent verification (Source: The Mills Review, p 143), and I've found nothing from a UK regulator on whether software counts as a person purporting to act for a customer under regulation 28(10) of the Money Laundering Regulations.
But parts of it already sit inside existing duties the moment an agent touches a regulated decision. The firm stays responsible for what its AI does. Someone has to be accountable for it. Due diligence records have to be kept (regulation 40). Agent identity and delegated authority are different: they're what the FCA's planned framework is for, so they're a direction of travel, not an existing rule. The FCA hasn't written a new rulebook, and says it doesn't plan to. But the existing one doesn't answer every question about agents, which is why that framework is being built.
For compliance leaders evaluating or deploying AI agents this year, the practical agenda is clear: maintain a register of every agent in your workflows and what it's authorised to do; map each one to a named accountable person, who should be a senior manager where SM&CR applies; demand decision-level audit trails from any vendor, not activity logs; define and test escalation boundaries and stop mechanisms; and treat agent identity as part of your identity architecture, alongside KYC and KYB — not as an IT procurement detail.
The institutions that get this right won't just be defensible. They'll be the ones able to deploy agents aggressively, because their governance can keep pace with their ambition. In compliance, the ability to prove control has always been the real enabler of speed.
KYC told us who the customer is. KYB told us who the company is. KYA will tell us who — and what — is acting on their behalf. The third pillar is being built now, and compliance teams have a rare opportunity to be early to a framework instead of remediating their way into it.
This post is research, not legal advice. It reflects what we had read by 1 November 2026, and the position is moving, so check the sources, and ask your own lawyers before relying on it.
First Mile Labs builds KYB/KYC orchestration and configurable risk decisioning infrastructure for banks and fintechs — with decision-level audit trails designed for the way compliance is examined, not just the way it's automated. If you're thinking through AI governance in your onboarding workflows, talk to us.
See automated KYB in practice
Book a demo and walk through a live KYB case from application to decision.
Request a demo →