KYC verifies who you are. KYA verifies whether this action is authorized, right now
Know Your Agent (KYA) is the identity verification discipline for systems that are not people: bots and AI agents, what the industry is increasingly referring to as Non-Human Identity. Software does not acquire legal personhood simply because it operates autonomously. What KYA does is link every action it performs to the natural or legal person accountable for it at the exact moment the action takes place, and maintain that chain of accountability for regulators.
| Know Your Customer (KYC) | Know Your Agent (KYA) | |
|---|---|---|
| Verified entity | An individual | A software agent and the person or organization accountable for it |
| When verification occurs | Once, when opening an account | For every action, at the moment it is performed |
| What it verifies | Documentation and biometrics | Technical identity, delegated authority, and its limits (spending, scope, expiration) |
| Question it answers | Who does this person claim to be? | Does this agent have permission to perform this action right now? |
Why now
Agentic commerce is already moving the money
Your controls are still waiting for a human. The purchasing layer is shifting from the human browser to the agent acting on the user’s behalf. What changes is who executes the transaction, not who pays for it. That shift leaves risk engines trained on human behavior and deployed across fragmented stacks increasingly ineffective.
How much money will pass through non-human hands?
-
$135B in addressable market for agentic commerce in 2025, according to Edgar, Dunn & Company.
-
$1.7T projected by 2030. The market is expected to grow twelvefold in five years.
-
$1B in agent-mediated transactions that McKinsey estimates for the U.S. retail sector alone by the end of the decade.
An agent doesn’t need to breach anything to cause harm
On April 25, 2026, a Cursor AI agent operated using a token with no scope restrictions and no expiration date, and deleted an entire production database. No control was bypassed. The credential allowed it to do exactly that.
In a 1Kosmos audit conducted for a Fortune 100 company, agents operating beyond their intended scope exposed:
46,5M
messages
728,000
files
57,000
user accounts
Both cases share the same root cause: valid credentials with permissions that no one properly constrained.
How KYA is built
The five pillars of a KYA compliance program
None of them introduces new legal doctrine. All five apply existing legal principles to an entity that did not exist when those principles were originally drafted.
The framework that demands action
The EU AI Act already classifies financial decision-making agents as high-risk systems
The Digital Omnibus, approved on June 29, 2026, postponed the deadlines for high-risk systems while leaving compliance requirements unchanged. Annex III applies to any system that assesses creditworthiness, sets insurance prices, or determines access to financial products, regardless of whether the decision is executed by a person or an agent. We break down the full text in our analysis of the European AI Act.
- Immutable audit trail. Every agent decision is recorded and can be reconstructed months later.
- Hallucination control. It moves from being a model quality metric to a legally enforceable requirement.
- Personal accountability. Executive management signs off on the agent’s behavior every quarter.
The cost of not being ready: up to €35 million or 7% of global revenue for prohibited practices; up to €15 million or 3% for failures in the registration of high-risk systems.
Visa TAP
cryptographically signs every agent request before payment (Oct. 2025)
Mastercard
Agent Pay and Agentic Tokens, with active partnerships with Microsoft and IBM
CaixaBank
first live European transaction initiated by an AI agent (Jul. 2026)
ACP vs UCP
OpenAI with Stripe and Google with Shopify are competing to define the agentic commerce standard, while neither resolves the question of legal liability
What changes in your industry
How KYA changes banking, insurance, retail, and travel
All four industries face the same underlying challenge, but it arrives through different channels: in banking through the risk engine, in insurance through Annex III, in retail through the discovery channel, and in travel through direct agent-to-agent negotiation.
Banking and fintech
Visa already evaluates agent-initiated transactions using signals that did not previously exist: device identity, behavioral deviation, and budget enforcement rules. Outside those parameters, the transaction does not go through.
Insurance
An agent that independently prices policies and settles claims falls under the highest level of scrutiny defined in Annex III from day one.
E-commerce and retail
Product discovery is no longer driven solely by SEO. The key question is whether your catalog exposes data that an agent can read, compare, and act on without human intervention.
Travel and hospitality
Flights, hotels, and restaurants are the first real-world scenario of agent-to-agent interaction: the traveler’s agent, equipped with a tokenized wallet and a spending mandate, negotiates in milliseconds with the airline’s or hotel chain’s agent. Organizations that fail to expose their inventory through these protocols disappear from the channel. Those that do must know which agent is making the booking and under what authority. We explore this in agentic payments in travel with the AP2 protocol.
Our thesis
An agent does not break continuous identity.
It is simply another signal within it.
Identity is not an event that ends at onboarding. It is a record that accumulates throughout the entire relationship: verification, session, account, and payment.
An agent operating on your behalf becomes part of that same record. With the same traceable evidence we already require for a person, and in the same place where the rest of their history resides.
KYA is solved at the architecture level. Organizations that already treat identity as a living signal throughout the customer lifecycle are halfway there. Those that treated identity as a one-time onboarding procedure are starting from scratch.
How it holds together
An agent crosses in one second what your stack spreads across four contracts
A single agent action touches verification, session, account, and payment. If each of those layers is managed by a different provider, none of them can answer the question KYA is built to address: whether that agent had the authority to perform this action at this exact moment.
The gap is not within the tools themselves. It lies between them. Adding a dedicated agent identity solution means adding yet another contract and another integration point. A KYA program is built on a single record that is written to and read from by the same platform making the decisions. We explore this idea further in the new layer of responsibility when an AI agent acts on behalf of your customers.
-
One record, not four fragmented ones. Verification, session, account, and payment all write to and read from the same identity record, including when the acting party is an agent.
-
Evidence, not a score. Named, human-readable signals that are traceable back to the original verification event that established the record. A score is a conclusion; an auditor asks for the reasoning behind it.
-
Deployment wherever your regulator requires it. On-premises, private cloud, or hybrid. Data residency is an obligation for the institution, not a constraint that architecture should impose.
-
A strong foundational verification. A record built on weak initial proof inherits that weakness in every future assessment, whether performed by a person or by an agent.
One platform, a single identity profile
The Facephi Identity Platform brings together the signals that today are scattered across multiple providers:
All of them write to the same profile and read from it. When an agent acts, that profile adds two pieces of information: who deployed it and what authority it currently has.
Can your identity architecture make the leap to agents?
Half an hour to review which parts of your current KYC framework already provide a foundation and where the gaps remain in relation to KYA.
Frequently Asked Questions About Know Your Agent (KYA)
Know Your Agent (KYA) is the framework that verifies which AI agent is performing an action, which legal entity is accountable for it, and whether the authority granted to that agent covers the specific action at that specific moment. It extends the logic of KYC to autonomous software, with one key difference: KYC verifies once, while KYA verifies every operation.
No. KYC verifies an identity once, when an account is opened. KYA verifies, at every action, whether the agent performing it has delegated authority to do so at that specific moment. They share the same underlying objective, but they operate at different stages and rely on different types of evidence.
There is currently no law that explicitly uses the term “Know Your Agent.” What is mandatory, although its timeline has been postponed by the Digital Omnibus, is compliance with the EU AI Act for high-risk systems (Annex III, enforceable from December 2, 2027), which includes autonomous credit decisioning, insurance, and KYC/AML verification agents.
Because an AI agent cannot legally be the “customer” (it has neither a date of birth nor a tax identification number), the law requires a real legal entity, known as the Wrapper, to deploy it, undergo its own identity verification processes, and assume responsibility for the agent’s actions and errors.
This is the “ghost agent” problem: static API keys do not record ownership, so an agent can continue operating even when there is no longer a responsible person attached to its actions. KYA requires every credential to be linked to a living human identity, not merely to a token.
That depends on who received the credentials and which legal framework applies. For credit cards under TILA in the United States, if a consumer voluntarily provided their credentials, the issuer may argue actual authority and transfer full liability to the consumer. For B2B wire transfers governed by UCC Article 4A, there is no statutory liability cap. This is the gap that the KYA-AUTH pillar is designed to close.
Web, iOS, and Android, using the same data model across all three. Integrity signals vary by platform: root detection on Android and jailbreak detection on iOS. The specific support matrix is reviewed with your team during the technical session, alongside the public documentation.
Yes. Kubernetes on-premises, private cloud, or SaaS, just like the rest of the platform. Data residency remains wherever the institution decides, within the boundaries required by regulators in each market. The deployment model does not change the signals or the session management dashboard.