Insight

|

Share

Agentic AI and the New Face of Fraud: Why Banks Must Know the Human Behind the Agent

YEO Mule Accounts: Inside Track briefing on account handovers and continuous identity assurance.

Get the

Document

AI assistants answer questions. AI agents act.

An agent can receive a goal, plan the required steps, and carry out actions via applications or APIs. In banking, this could include managing bills, moving money, creating payment instructions or supporting treasury activity.

This creates significant opportunities for faster, more personalised financial services. It also changes the identity question banks need to answer.

Banks and payment networks are developing ways to identify an AI agent, define its permissions and verify that its requests fall within an approved mandate. These controls are essential.

But knowing which agent is acting does not always establish whether the right human created the mandate, approved an exception or remained present during a supervised action.

Agentic banking therefore requires two distinct forms of assurance:

  • Know Your Agent: Which agent is making the request, and what is it permitted to do?

  • Know Your Human: Is the enrolled customer present at the control point where human involvement is required?

Banks will need both.

From assistance to delegated action

The difference between an AI assistant and an AI agent is important.

An assistant responds to a request with information, analysis or a draft. The human remains responsible for carrying out the action.

An agent can operate within an agreed mandate. It may access accounts, communicate with other systems, and complete permitted actions without requiring the customer to approve each step.

That changes the risk model in three ways.

Delegation

The customer grants software authority to act within defined permissions and limits.

Autonomy

Once authorised, the agent may complete routine actions without the customer having to return for each step.

Machine speed

A compromised or misused mandate could allow actions to be attempted across accounts and systems much faster than a human could perform them manually.

The problem is not that agentic payments lack security controls. Agent credentials, scoped permissions, tokens and signed instructions can help banks establish which agent is acting and what it has been authorised to do.

The remaining question is where the bank still needs evidence that the enrolled customer is present.

The human-presence gap

A trusted agent could still operate under a mandate created or approved by the wrong person.

Consider the points where identity risk could emerge:

  1. Mandate creation: Was the enrolled customer present when authority was granted to the agent?

  2. High-risk exceptions: Who approved a new beneficiary, increased payment limit or unusual transaction?

  3. Supervised actions: Did the enrolled customer remain present during a particularly sensitive action?

  4. Disputes: What evidence can the bank produce about who was present at the relevant control point?

  5. Agent compromise: Is the agent still operating within its approved permissions?

The final question requires agent security, behavioural analysis, transaction monitoring and other controls. Human-presence technology does not establish whether an agent itself has been compromised.

However, at selected points where the bank requires the customer to return, it can provide an additional identity signal.

Know Your Agent and Know Your Human

Know Your Agent and Know Your Human solve different problems.

Know Your Agent uses credentials, tokens, permissions and signed instructions to identify the software and control what it may do.

Know Your Human provides evidence that the enrolled customer is present when the bank requires human involvement.

One should not be treated as a replacement for the other.

The bank may allow an agent to complete routine, low-risk actions under an existing mandate. When the agent reaches a defined trigger, such as adding a new payee or requesting a payment above an agreed limit, the customer can be asked to return to the banking app.

At that point, the bank needs more than a tap from an authenticated device. It may need additional evidence that the person present matches the enrolled customer.

Where continuous identity assurance fits

YEO’s Continuous Facial Recognition SDK can add identity evidence at three selected points within an agentic banking journey.

1. Mandate creation

Before an agent is granted authority, CFR can check that the enrolled customer is present and passes the configured facial matching and liveness checks.

The mandate, permissions, and transaction limits remain under the bank's control.

2. High-risk exceptions

If an agent attempts an action outside its normal authority, the bank can request another identity check before applying its risk policy.

The result can inform a decision to proceed, pause, step up authentication or refer the action for review.

3. Supervised sessions

For selected actions requiring continued human involvement, CFR can provide signals relating to presence, absence or facial mismatch throughout the relevant part of the journey.

Biometric processing takes place on the user’s device. Structured event outcomes can be shared with the bank’s existing systems, while the bank retains control over when checks occur and how it responds.

An agent requests a high-value payment

Consider an AI agent authorised to manage routine payments.

The customer creates a mandate defining the agent’s permissions and limits. The agent then carries out permitted activity without requiring the customer to approve each action.

Later, the agent attempts to add a new beneficiary and release a payment above the agreed limit.

The bank identifies the exception and asks the customer to return to its application. CFR checks whether the face present matches the enrolled template and passes the configured liveness checks.

The result is sent to the bank’s controls, which decide whether to continue, challenge, pause or review the payment.

CFR provides evidence about identity and presence. It does not establish the customer’s intent or guarantee that the payment is legitimate.

That distinction matters. A genuine customer could still be manipulated or coerced into approving an action. Device intelligence, behavioural analytics, scam detection and transaction monitoring remain essential.

Identity evidence within the wider fraud stack

The strongest agentic banking security model will combine several signals:

  • Agent identity and credentials

  • Mandate scope and permissions

  • Device intelligence

  • Behavioural analysis

  • Transaction monitoring

  • Customer identity and presence

  • Exception and dispute records

Continuous identity assurance contributes one part of that picture: whether the person present at a selected control point matches the enrolled customer and passes the configured checks.

For fraud and financial-crime teams, this provides another signal around mandate creation and exceptions.

For risk and dispute teams, it creates an event record showing whether presence, facial match and liveness checks passed.

For digital banking teams, it offers a way to introduce additional identity assurance within selected customer journeys.

For security teams, on-device processing provides an architecture in which biometric information need not be stored in a central YEO biometric database.

Start with one control point

Banks do not need to introduce identity checks across every agent action.

A credible proof of value should begin with one defined journey and one control point, such as:

  • Creating an agent mandate

  • Adding a new beneficiary

  • Increasing an agent’s transaction limit

  • Approving an unusually high payment

  • Recovering or changing account access

The bank can then measure detection performance, false interventions, completion rates and device impact against agreed scenarios.

Agentic banking will require banks to identify and govern the software acting on behalf of customers. It will also require them to decide when the human must return.

Know Your Agent secures the agent. Know Your Human helps establish whether the enrolled customer is present when it matters.

CTA

Where would your bank require the customer to return during an agentic journey?

Book a session with YEO to identify one control point where continuous identity assurance could be evaluated.

Get the

Document

Solution Mentioned

Ready to Know Who's There?

Share

About us

We stopped asking "who logged in." We started asking "who's still there." YEO began as a secure messaging app. Today we build the patented continuous identity verification infrastructure that regulated industries trust to prove who's really there.

Certifications

© 2026 YEO messaging Ltd is registered in England and Wales 10785061.

🇺🇸 US patent No: 10,607,035
🇬🇧 GB patent No: 1713943.7
🇨🇳 Chinese patent No: 201880071020.4
🇪🇺 EU patent No: 3676738

Certifications

© 2026 YEO messaging Ltd is registered in England and Wales 10785061.

🇺🇸 US patent No: 10,607,035
🇬🇧 GB patent No: 1713943.7
🇨🇳 Chinese patent No: 201880071020.4
🇪🇺 EU patent No: 3676738

Certifications

© 2026 YEO messaging Ltd is registered in England and Wales 10785061.

🇺🇸 US patent No: 10,607,035
🇬🇧 GB patent No: 1713943.7
🇨🇳 Chinese patent No: 201880071020.4
🇪🇺 EU patent No: 3676738

Certifications

© 2026 YEO messaging Ltd is registered in England and Wales 10785061.

🇺🇸 US patent No: 10,607,035
🇬🇧 GB patent No: 1713943.7
🇨🇳 Chinese patent No: 201880071020.4
🇪🇺 EU patent No: 3676738