Know Your Human (KYH): AI agents need a policy, not a blanket pass Skip to content
AI-assisted fraud & geo-compliance investigations with GeoComply MCP
Learn How >>

Know Your Human (KYH): AI agents need a policy, not a blanket pass

Read time:
0 minutes

When an agent pays a bill for your customer, it can look like a fraudster’s bot. Block it and you may stop a real customer. Let it through without checking who sent it and you may admit someone using stolen credentials.

For a bank or payment platform, that decision affects whether a legitimate payment goes through and what evidence your team has if it’s disputed. The challenge is to support customers’ agents while keeping control over who can act.

When an agent acts for a customer, Know Your Human (KYH) focuses on connecting its activity to that customer’s approval. The goal is to let legitimate tasks move ahead with fewer unnecessary interruptions, while giving your team evidence of who approved which agent and what it can do.

Different agent journeys need different evidence

On October 20, 2026, Visa, BILL, and Billtrust announced a live pilot allowing BILL’s authorized payment agent to complete approved invoice payments through supplier portals. The agent uses tokenized payment credentials and controls tied to the buyer’s approved instructions, rather than shared supplier-portal logins. Visa’s Trusted Agent Protocol supports agent recognition and action-specific authorization through digital signatures.

That’s a purpose-built payment flow. A general-purpose assistant may instead open your website and sign in with the customer’s credentials. The request can reach your platform without that same protocol-enabled path.

Meta’s Muse uses a dedicated virtual machine with its own browser, where it can compare prices or make an approved purchase. ChatGPT’s cloud browser also runs remotely through OpenAI’s ChatGPT Work service, navigating supported websites and pausing when sign-in or confirmation is needed.

Claude in Chrome provides a different example: it reads, clicks, and navigates inside the customer’s Chrome browser. Anthropic’s permissions guide prohibits it from making purchases or financial transactions.

For your team, the distinction is whose environment you’re seeing. A hosted agent’s remote browser is expected infrastructure, and its location does not establish the customer’s location. A familiar customer browser, meanwhile, doesn’t prove a person clicked the button.

Recognize the agent, then establish genuine approval

Once you can distinguish agents, your team can decide which to observe, which need customer approval, and which your policy doesn’t permit.

Knowing which assistant is acting identifies the service, not who sent it. A customer may have signed in successfully and asked it to compare plans; that still doesn’t authorize a purchase. Agent identity, customer authentication, and permission for the action each need their own evidence.

Where approval is needed, it must reflect the account holder’s decision. An approval link alone doesn’t establish that: automation or remote control can also be present in the approval environment. Those signals need assessment alongside customer authentication, rather than treating a completed page as proof of human approval.

The agent’s browser provides evidence about the automation making the request. A check on the customer’s own device adds evidence of its location at approval, device integrity, and network risk. For location, that means checking whether the signals have been spoofed or manipulated. Linking those checks to the authenticated customer, the specific agent instance, and the customer’s explicit approval creates a record of what was allowed, when, and for what purpose. That record needs to hold up too: the signed verification evidence must match the request and be recent enough for the decision, and the approval must still be valid.

Keep valid approvals useful, with clear limits

Where platform policy permits, a valid approval can cover routine access for an agreed period. But your customer’s assistant and someone else’s copy of the same service may share a name. Permission belongs to the agent and purpose the customer approved. Identifying that particular instance takes more than recognizing the service.

A sign-in approval is different from permission to make a payment. If a customer asks an agent to compare subscription plans, that covers research. Buying an upgrade or accepting a recurring charge needs permission covering that action. Customers also need a way to withdraw permission.

A customer can move after approving an agent. The permission may still be valid, but the earlier location check no longer establishes where their device is now. If location matters to the action, the evidence needs to be recent enough for that decision.

GeoComply’s role in agent verification

A general-purpose assistant can use your existing sign-in page with the customer’s credentials. GeoComply’s browser-based checks use browser, device, and network signals to detect agent activity and recognize supported agents. The agent doesn’t need to register with us or declare its identity.

That recognition can help your team apply different policies instead of treating all automation alike. Your team decides which agents to allow, which to block, and which actions need customer approval.

Our approach connects agent recognition with account-holder approval and a check on the customer’s own device. Where location matters, the evidence comes from that device at approval, rather than the agent’s remote environment. GeoComply checks location signals for spoofing and manipulation, drawing on more than a decade of geolocation experience in regulated markets. Linking those checks to the customer’s explicit approval, the agent, and the purpose gives your team a record of what was allowed and why. If the activity is disputed, your team can trace the decision back to the evidence behind it.

Start with the decision you need to defend

Start with one task, such as an agent-initiated payment or subscription change. Map the customer, agent, permitted action, and evidence available at each step. Check what happens if the amount changes, permission expires, or device signals conflict. Could an investigator reconstruct why the action was allowed?

Use that review to identify where the evidence already supports a decision and where something is missing. Talk to GeoComply about agent recognition and account-holder approval for your customer journey.

Frequently asked questions

Know Your Human (KYH) means linking an AI agent’s activity to the real account holder behind it and establishing what that person approved. It connects the customer, the agent, and the purpose of the approval, helping platforms determine whether the agent has permission to act. Verification on the customer’s own device adds evidence about the approval environment, including location, device integrity, and network risk. A successful sign-in or device check alone doesn’t establish consent.

No, opening an approval link alone does not prove the account holder approved. A reliable approval record needs to establish the customer’s identity and connect their decision to the agent’s request.

No, a hosted agent’s IP address does not establish the customer’s physical location. A location check on the customer’s device records where it was at approval, not throughout later agent activity. If location matters to the decision, decide how recent the check must be and when to check again.

No, KYH doesn’t replace applicable Know Your Customer requirements or authorization for a specific action. Knowing who is behind an agent, assessing the approval environment, and confirming what the customer allowed remain separate checks.

Related Posts

Black October: how fintechs can protect traders from AI-era cyberattacks

Balancing compliance and privacy: Essential strategies for today’s crypto firms

Enhancing Financial Security: Advanced Geolocation for Comprehensive Sanctions Compliance