All Insights

    AI Governance

    Know Your Agent in 2027: if your bank can't tell which agent is acting, the loss lands on you

    US banks now treat an AI agent using your login as you. What Know Your Agent (KYA) means, how Singapore, US banks and India differ, and five checks for 2027.

    Prashant Bhardwaj20 November 20268 min read
    Share

    On 19 November 2026, two answers to the same question arrived from opposite sides of the world. The question: when an AI agent moves a customer's money, who answers for it?

    In Singapore, Chia Der Jiun, Managing Director of the Monetary Authority of Singapore, told the Singapore FinTech Festival that some users are already giving agents their login credentials to financial accounts. "I would not advise giving personal agents such access," he said, "as there is insufficient assurance of safety." MAS will instead facilitate a Know Your Agent protocol, so that banks can verify an agent before it acts.

    In the US, the answer came in the small print. Citi's revised online banking terms treat actions taken with credentials a customer hands to an outside agent as authorised by the customer, even when the agent ignores instructions or exceeds its permissions.

    Two answers, one message for 2027: if your bank can't tell which agent is acting, the loss lands on you.

    Questions this article answers

    What does Know Your Agent mean for a bank?

    It is a way to verify an AI agent before it acts: who the agent is, whom it represents, what authority it holds and what mandate it is carrying out right now. See the four fields.

    Who carries the loss when an AI agent makes a payment?

    Increasingly, the customer. US banks are rewriting their terms, Singapore is building a protocol, and India still makes a human approve. See the three answers.

    Can an AI agent be tricked into paying a fake invoice?

    Yes, even with Know Your Agent in place, if a fraudster gets a fake supplier onto the approved list. See where KYA still fails.

    1. Why a shared login breaks accountability

    When an agent logs in with your password, the bank sees you. It can't tell whether software is acting, whether it is staying inside your instructions, or who is responsible if it transacts, in Chia's words, "incorrectly or beyond their instruction".

    That was manageable when agents only read statements. It isn't now that they pay invoices, book travel and move deposits. A clerk who makes a mistake sits inside a chain of accountability. An agent on a borrowed password breaks that chain.

    Banks have noticed. Citi says it does not permit third-party agents to log in to customer accounts or transact on its platforms, and JPMorgan and Bank of America announced similar policies in October. So the 2027 question is not whether agents will touch money. It is whether they do it with an identity of their own, or by pretending to be you.

    2. Three answers on who carries the loss

    United StatesSingaporeIndia
    ApproachLiability moves to the customer through banking termsA common protocol identifies the agentRails first; the human still approves
    What happenedCiti: agent actions on shared credentials count as customer-authorised. Wells Fargo: from 24 November, customers must revoke agents themselvesMAS will facilitate a KYA protocol and a dedicated, secure channel for agent accessAgent checkouts on UPI; the user approves with a UPI PIN or a pre-set mandate
    Who carries the lossThe customer, in most casesWhoever the protocol identifies as responsibleMostly the user who approved
    Open questionDo protections against unauthorised transfers still apply?Will banks adopt the protocol at scale?Who answers when a mandate lets the agent decide?

    United States. Citi's terms go further than most. The bank says it has no obligation to investigate whether an agent followed the customer's prompts, subject to applicable law. Wells Fargo's updated agreement adds another step: customers must revoke an agent's authorisation themselves, and stay liable unless they can show they did. PNC draws a better line. Bank-approved connections get limited access, customers may not give agents their passwords or one-time codes, and agents generally can't start payments unless PNC permits it. Regulators haven't settled the question. The OCC's examiner guidance on outside agents is non-binding. The CFPB says agreements can't waive protections against unauthorised transfers, but how that applies to an agent the customer authorised is unclear.

    Singapore. MAS and industry published SAFR, the Safeguards for Agentic Finance at Runtime white paper, on 3 July 2026 and updated it on 9 October. It is a reference model, not regulatory guidance. Its next phase is the KYA protocol, which The Business Times reports will cover five things: the agent's identity, whom it represents, its authority, its mandate, and its provenance and reliability. Banks then decide the terms each agent is allowed.

    India. Payment firms are building agent-ready rails first. On Cashfree's agent-ready checkout, the user still approves every payment with a UPI PIN. Navi's MD and CEO Rajiv Naresh has called Know Your Agent the more complicated piece of agentic payments. There is no shared KYA standard yet.

    3. The four fields a bank checks

    Take an illustrative case. Ravi runs ExportCo, a mid-sized exporter in Chennai with a three-person finance team. He deploys an AI agent called Ledger to pay suppliers. Ledger does not use Ravi's login. It has its own cryptographic identity, so the bank knows it is talking to Ledger, not to Ravi.

    Before the bank executes a payment, it checks four fields.

    FieldWhat it provesExportCo exampleThe bank says no when
    1. IdentityThe agent is a known, registered entityLedger's own credential, separate from Ravi'sThe agent is unknown, spoofed or using a human login
    2. SponsorA real, accountable principal stands behind itExportCo, with Ravi as the named personThe agent can't prove on whose behalf it acts
    3. AuthorityThe standing limits on what it may doSupplier payments up to $50,000, approved suppliers only, business hours onlyThe amount, payee or time falls outside those limits
    4. MandateThe live instruction behind this action"Pay the fabric supplier by Friday"There is no current instruction, or the payment doesn't match it

    Keep fields 3 and 4 apart. Authority is set once and reviewed rarely; the mandate is today's job. Mixing them is how an agent ends up with permanent power for a one-time task. We covered the design side in our enterprise blueprint for agentic AI.

    MAS adds a fifth item, provenance and reliability: where the agent came from and how far it can be trusted. For most enterprises, that is a vendor due-diligence question to settle before the agent is ever registered.

    4. Where Know Your Agent still fails

    Ledger's authority cap limits the amount. It says nothing about whether the bill is real. A fraudster doesn't need to break KYA. A convincing invoice under $50,000, from a supplier name that looks right, is enough.

    So a serious setup stacks three gates. A "no" at any one of them blocks the payment and sends it to a person.

    1. Grounding, run by the agent: the invoice must match the purchase order, the delivery record and the supplier's bank details on file.

    2. Know Your Agent, run by the bank: the agent, its sponsor, its authority and its mandate.

    3. Fraud screening, run by the bank: the same device, behaviour and anomaly checks that watch human payments.

    Now the gap. All three gates trust the approved supplier list. If a fraudster talks someone into adding a fake supplier, or changing a real supplier's bank account, every gate passes. The payment is authorised, within mandate and normal-looking.

    Put the human check where trust is created. A human approval on every change to the supplier or beneficiary list matters more than one on every payment. Approve every payment and your team burns out. Approve every list change and you close the door that fraud actually uses. This is the runtime control problem behind the 2027 AI agent control layer, and it is why spend limits alone are not governance.

    5. Five questions to answer before 2027

    Each question maps to a check in our CLEAR framework.

    QuestionWhat a good answer looks likeCLEAR check
    Identity: does each agent have its own credential?No agent uses a human login or a shared service accountLineage
    Sponsor: who answers for each agent?A named person and legal entity per agent, on recordAccountability
    Authority and mandate: are they stored separately?Standing limits in one place; today's instruction logged with each actionConsistency
    Upstream trust: who approves list changes?A person signs off every new supplier, beneficiary and limitEvaluation
    Revocation and evidence: can you switch an agent off and prove it?Any agent revoked within minutes, and any payment traceable from instruction to settlementResilience

    If you can't answer three of the five today, start there. Our 2027 enterprise AI readiness checklist already lists shared credentials as a red flag. This is the same problem, now with money attached.

    This is Part 1 of The 2027 Agent Economy series. Next: who pays when an AI agent gets it wrong, and how merchants can know the human behind an agent.

    For an outside view of your agent plans against these five questions, see our AI Production Readiness & Governance Review, or request a strategy conversation.

    FAQ

    What is Know Your Agent (KYA)?

    Know Your Agent is due diligence on the AI agent acting for a customer, not only on the customer. Before the agent acts, a bank checks its identity, whom it represents, its authority and its current mandate. MAS's planned protocol also covers provenance and reliability.

    How is Know Your Agent different from KYC?

    KYC verifies the human customer. KYA verifies the software acting on that customer's behalf. It complements KYC rather than replacing it.

    Who is liable if an AI agent makes a wrong payment?

    It depends on the bank and the jurisdiction. Several large US banks, including Citi and Wells Fargo, have changed their terms so that customers carry most of the risk from agents they authorise. Check your bank's terms and take legal advice for your own case.

    What is MAS SAFR?

    SAFR, or Safeguards for Agentic Finance at Runtime, is a white paper MAS and industry published on 3 July 2026. It sets out runtime checks that verify and record an AI agent's proposed actions before they execute. It is a reference model, not regulatory guidance, and its next phase adds the Know Your Agent protocol.

    Does India have a Know Your Agent standard?

    Not yet. India's agentic payments run on UPI, where the user approves with a PIN or a pre-set mandate. Industry leaders have called KYA the harder piece still to solve.

    5–7 day review

    Find out which of your AI use cases you could not defend to a regulator today.

    Filed under

    2027 readinessKnow Your AgentKYAAI agent identityAgentic paymentsBFSIAI Governance

    Continue reading

    More on AI Governance