How to layer fraud checks on top of Anthropic's KYC Screener agent
Anthropic released a pre-built KYC Screener agent last month. It runs a four-step workflow on onboarding records to extract structured data from KYC documents, evaluate that data against a firm's KYC rules, screen named parties, and escalate exceptions to a compliance file for human review.
The Anthropic template is purpose-built for meeting basic KYC compliance requirements during onboarding, and it lowers the cost of getting it right. For teams with strong in-house engineering and an existing data stack, it's a foundation worth building on.
But meeting a regulatory requirement doesn't mean your business is protected from fraud. Large organizations know this, and smaller ones will quickly find out if they don’t put fraud prevention measures in place.
What the KYC Screener isn’t built for
The Anthropic KYC Screener agent is designed to be a client-onboarding analyst that assembles and screens a KYC file based on structured data. While it might handle document parsing and rules evaluation well, the following actions are out of its scope:
Telling you if the document is risky. Extracting data from a government ID won’t tell you if the ID is legitimate. Proper risk detection involves analyzing the image for tampering or forgery.
Flagging fraudulent selfies. Catching fraudulent selfies, including AI face spoofs and deepfakes, requires visual detection models and injection attack detection. The latter can look for signs of app tampering, device integrity, and virtual camera usage.
Uncovering connected accounts. Focusing on onboarding one user at a time could lead you to miss signs of risk across users. For example, a fraudster might use the same device to submit multiple applications or create multiple accounts using the same selfie.
Six fraud checks worth adding to your onboarding flow
Some organizations separate fraud detection into a completely different flow. Others adopt a fraud anti-money laundering (FRAML) approach that combines the two. Either way, running the following six types of fraud checks can help you answer important questions during onboarding.
Document: Is the barcode structure consistent with an issued document? Does the portrait match the visual zone? Are there signs of pixel manipulation or GenAI?
Selfie: Did a person present the selfie, or was a recording injected? Is the person wearing a mask or submitting a picture of another screen? Does the face match the ID portrait? Is the selfie similar to prior submissions?
Device: Is the user on a real mobile device or an emulator? Does the camera metadata match the device’s hardware?
Behavioral: How long did the user spend per step? Did they use a suspicious number of keyboard shortcuts? Did they hesitate on specific questions?
Network: Is the user on a VPN or using Tor? Is their location consistent with the address on the ID?
Link analysis: Is the user connected to other accounts through a shared device, network, identifying information, or other data points?
Where are you on the business maturity curve?
KYC compliance is a baseline requirement for every organization covered by the related rules under the Bank Secrecy Act (BSA), and the Anthropic KYC Screener is a meaningful starting point. However, as your business grows, your fraud attack surface expands, and it becomes more important to layer in dedicated fraud checks. Here’s a quick look at how to think about layering KYC with fraud prevention based on a company’s stage.
Early stage: Speed and simplicity are a necessity
In their earliest days, pre-product market fit and while volume is still low, companies often optimize for speed and simplicity. They may need to work with a vendor that offers APIs for KYC and anti-money laundering (AML) screening, and they might want to implement some fraud prevention checks. But they don’t necessarily want or need overly complex tools and setups.
Anthropic’s KYC Screener could be a good starting point for a basic KYC onboarding workflow if you have the data pipelines set up. But keep these pointers in mind:
Have a human review every decline. An incorrectly declined good customer is a lost customer. You might automate decisions in the future, but you may want to manually review each declined user until you validate the agent's accuracy.
Log the full reasoning behind the outcome. When something goes wrong (and it will), you'll need to reconstruct why the agent decided what it did.
Additionally, remember that you still need data pipelines for the KYC Screener agent to work. Persona can send verification results and watchlist outcomes straight to the agent via an API. For teams that prefer an MCP interface, Persona's MCP server is in early access.
Growth stage: Fraud prevention becomes a larger priority
As your business scales across customers, markets, and products, your fraud attack surface expands. You're likely focused on growth, but that focus can create blind spots that allow fraudsters to slip in. Having automated and configurable protections in place allows you to fight off most basic and many advanced fraud attacks without impacting genuine growth.
A common approach is to start by collecting and analyzing passive signals, such as device, network, and behavioral signals. If you’re collecting a user’s information to run KYC screening anyway, you can use some of the same checks to gather many of these passive signals.
Stacking multiple signals can help increase assurance and allow you to automate decisions with confidence. You also may be able to use dynamic risk verification flows to add or remove friction based on the initial risk signals.
Mature stage: Consolidating your stack is more important
Mature companies often have a collection of separate tools for KYC/AML, fraud detection, case reviews, and orchestration. As a result, they often have a new challenge in trying to manage their compliance and fraud stacks.
Consolidation can be a big undertaking, but an all-in-one platform for KYC, fraud prevention, and reviews can remove data silos that hinder fraud detection and manual review efficiency. You can also still benefit from agentic reviews.
For example, Persona’s identity platform offers Case Review Agents that can review each case using data that you collect from Persona’s KYC/AML checks. If you’re also layering in fraud detection, the risk signals you collect are available to the same agent.
A KYC onboarding agent is the starting point
It's exciting that agentic workflows are coming to KYC, and the Anthropic KYC Screener is a solid starting point for some teams. But what you build around it should match where your business is at.
If you're an early-stage builder, connecting Persona's data via API gives the agent clean, structured inputs from day one. As your fraud attack surface grows, a purpose-built platform that handles KYC and fraud detection starts to make more sense.
If you're ready to connect your setup to real identity data, the API is the place to start. If you want to explore what a platform-native agent looks like, Case Review Agents is in early access. Reach out to our team if you have questions or want to get started.