A GUIDE
Know Your Agent: a guide for financial institutions
Know Your Agent, or KYA, is a label for an old set of questions about new software: who is acting, on whose authority, and who answers for it. No regulator requires it by name. This guide says what the term means, what regulators have said so far, and which controls keep coming up.
Reviewed on 30 October 2026. This is research, not legal advice. Your MLRO decides what your firm needs.
WHAT KYA MEANS
A label, not a legal term
Know Your Agent means finding out three things about software that acts for someone: who stands behind it, what it’s allowed to do, and what it did.
No law we’ve read defines the term, and no regulator requires it by name. As far as we can tell it came from the market, but we couldn’t verify who used it first. The one official UK text we found that uses it is the AI Adoption Plan for financial services. The Mills Review and the FSB, MAS and IMDA texts below don’t use it.
KYA doesn’t replace KYC or KYB. They tell you who your customer is. KYA asks what that customer has allowed software to do, and whether you can show it. It comes up when a customer wants software to act on their account.
THREE KINDS OF AGENT
The word covers three different things
Each raises a different question, and a vendor’s “KYA” may mean only one of them. The first two are your customers’ agents.
A customer’s own agent
Software your customer runs, or has had built, to act on its own account.
It looks like the customer acting through a tool. The question is what the customer allowed, and who at the customer said so.
A third party’s agent
Software another company runs, acting for your customer.
The rules on a person acting for a customer point here most clearly, because the operator is someone else. Whether software counts is still open, so ask your lawyers.
Your firm’s own agent
Software your firm runs in its own work, such as onboarding or alert handling.
No customer authority is involved. This is your own governance, and most of the guidance so far is written for it.
One limit applies to your customers’ agents. An agent nobody has declared, using a customer’s own credentials, looks like the customer. A list of known agents can’t show you what nobody told you about.
WHAT REGULATORS HAVE SAID
No regulator requires KYA by name
Here is what they have said so far, as at 30 October 2026. A speech isn’t a rule, and a rule written for people may or may not reach software. Each entry links to its source.
United Kingdom
- In forceLaw
Money Laundering Regulations 2017, regulation 28(10)
Where a person purports to act on behalf of a customer, a firm must verify that they’re authorised, identify them and verify their identity. We found nothing from a UK authority on whether software counts. In July 2025 HM Treasury read the rule as reaching third parties that act for an organisation, and staff as acting as the organisation.
- 26 October 2026Survey
Bank of England and FCA: AI in UK financial services
80% of respondents reported AI agents deployed, but the report says “autonomy is limited”: only 2% reported agents with broad autonomy. Even counting plans, no more than 9% expected agents to initiate, execute or block payments. It had 196 responses, 105 of them to the full questionnaire, and reflects participating firms, not the whole sector.
- 19 October 2026Speech
Financial Conduct Authority: speech
Its chief data, information and intelligence officer said it’s embedding six of the Mills Review’s seven recommendations, and “will work with industry, fellow regulators and frontier labs to develop a trusted framework for agentic finance”, including approaches to agent identity, delegated authority, accountability and execution. The speech gives no date.
- 19 October 2026Good and poor practice
Financial Conduct Authority: good practice paper
Its examples “do not create new requirements”. They include firms that “registered their agents in a central inventory so they were easier to oversee”. The paper also says firms “can outsource AI services, but remain responsible for them”.
- Closed 6 October 2026Consultation
HM Treasury: Modernising Payment Services Regulation
HM Treasury has consulted on how payment rules should adapt to agents. Question 15 asks: “How does existing payment services regulation need to adapt to support agentic payments?” The response isn’t out yet.
- 14 July 2026Independent report
AI Adoption Plan: Financial Services
The one official UK text we found that names Know Your Agent. The government’s AI Champions recommend a trust framework for agentic payments, with “Know Your Agent (KYA) Protocols” as one of three pillars. It’s rated High priority, with no date, and the report says industry should lead first.
- 14 July 2026Speech
Bank of England: Andrew Bailey, Governor
He asked whether agentic AI is “more a tool of the principal rather than strictly an agent”, with the principal fully liable. He put it as a question, not a finding.
- July 2026Advice to the FCA
The Mills Review
It says “The UK does not currently define agent identity or require agent verification.” It recommends that the FCA “should consider leading the development of a trusted framework for AI agent participation in financial services”. The review is about retail financial services.
- 30 June 2026Speech
Bank of England: Sarah Breeden, Deputy Governor
She said “relying on 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.
European Union
- From 10 July 2027Law
AML Regulation, Articles 20 and 22
Firms must verify that any person purporting to act on behalf of a customer is authorised, and identify and verify them. Article 22 lists the data to collect for natural persons, legal entities, trustees and other organisations. It has no category for software.
- From 10 July 2027Law
AML Regulation, Article 76(5)
A firm may take automated or AI decisions, provided that a decision to enter, refuse or maintain a relationship “is subject to meaningful human intervention”. This governs the firm’s own decisions, not its customers’ agents.
- From 2 December 2027Law
AI Act, Annex III
The AI Act’s Annex III high-risk rules apply from 2 December 2027, and the rules for regulated products under Annex I from 2 August 2028. They’re about high-risk systems, not agents as such.
- 20 July 2026Non-binding guidelines
Commission guidelines on Article 50 of the AI Act
Article 50 started to apply on 2 August 2026, with a transition to 2 December 2026 for some existing systems that generate synthetic content. The guidelines say AI agents that interact with people “must be designed and developed in such a way that they disclose both their artificial nature and the person on whose behalf they are acting”.
- April 2026 textNot adopted
Payment Services Regulation
The Council’s final compromise text doesn’t mention AI agents. Its one reference to artificial intelligence is about fraud monitoring. PSD2 still governs.
International
- 15 October 2026Non-binding
Financial Stability Board
Its sound practices “do not constitute a binding international standard”, and it says those specific to agentic AI “remain nascent”. It lists registration, cryptographically verifiable credentials and scoped authority among controls firms may use, including for agents acting on behalf of customers.
- 15 October 2026Circular
Hong Kong Monetary Authority
It expects banks to designate an “owner” for each AI use case. It also issued a concept note on agentic AI, which will inform future supervisory requirements, and says a discussion paper will follow.
- 7 October 2026Guidelines
Monetary Authority of Singapore
Its AI risk guidelines take effect on 7 October 2027, and they cover AI agents. For an agent, the attributes a firm records in its AI inventory may include “agent-specific identifiers”.
- 20 May 2026Emerging best practice
Infocomm Media Development Authority, Singapore
Its framework for agentic AI says that, in the interim, an agent should have its own “unique, cryptographically verifiable identity”, tied to a supervising agent, a human user or an organisational department.
THE FIVE QUESTIONS
Five questions to ask of any agent
Your second line is likely to ask them, whichever kind of agent it is. They come from our post on Know Your Agent.
- 1. Who is it?
- Which agent, which version, and who built and runs it. “An AI tool” isn’t an answer, because an update can change what it does.
- 2. What may it do?
- What it’s authorised to do, within what limits, and where that’s written down. Preparing an action and taking it are different permissions.
- 3. Who is accountable?
- A named person or a company. For your customers’ agents, that starts with who at the customer approved it. For your firm’s own, it’s who owns it in your firm.
- 4. What’s recorded?
- What it was asked to do, what it did, and on what authority. An activity log shows that something happened. It doesn’t show that it was allowed.
- 5. How do you stop it?
- What happens when it meets something outside its limits, and how quickly someone can stop it. If nobody has tested the stop, you don’t know it works.
THE CONTROL MAP
15 controls that keep coming up
Regulators haven’t asked for KYA, but the same controls recur across their guidance. Almost all of it is guidance, and few of these controls have a binding rule behind them. Most of it is also written for agents a firm runs itself, so applying it to your customers’ agents is a judgement for your firm.
Use it as a checklist. Take one agent, and work down the list.
- 01
Unique, verifiable agent identity
Can you tell this agent from any other, by something it can prove and not just something it says?
- 02
Agent register
Do you keep a current list of the agents you know about, and what each is for?
- 03
Accountable person or company
Which named person or company answers for this agent?
- 04
Scoped, recorded delegated authority
What may it do, within what limits, and where is that recorded?
- 05
Check before acting
Is an action checked against those limits before it happens?
- 06
Human approval at thresholds
Which actions need a person’s approval first, and who sets the threshold?
- 07
Measuring whether oversight works
Do you measure how often reviewers change the outcome, or could approval have become a habit?
- 08
Audit trail for each action
Can you show each action, on what authority, and that the record hasn’t been altered?
- 09
Continuous monitoring of behaviour
Would you notice if its behaviour changed?
- 10
Revocation and lifecycle
Can you withdraw its authority, and does that happen when someone leaves or a risk changes?
- 11
Kill switch or safe stop
Can you stop it quickly, and have you tested that?
- 12
Testing and red-teaming
Has anyone tried to make it go wrong, before it went live and since?
- 13
Third-party agent management
If someone else runs it, do you know who they are, and what your contract lets you do?
- 14
Multi-agent and sub-agent control
If it hands work to another agent, do its limits follow, and can you trace the chain?
- 15
Risk tiering by autonomy
Do your controls tighten as the agent does more on its own?
Drawn from IMDA, MAS, the FSB, the HKMA, the FCA’s good practice paper, the SAFR industry paper, public comments to NIST’s NCCoE and OWASP. None of these is a UK or EU rule.
READ NEXT
More on agents and governance
DESIGN PARTNERS
Work through this with us
We’re looking for a small number of firms to work through this with. Talk to us, especially if a customer has asked to let software act on their account.
Talk to us about agents →