Zahava Dalin-Kaptzan on AI-Driven Fraud Prevention at Riskified

1. How do you balance fraud prevention with customer experience to reduce false declines?

False declines are a costly problem for merchants. They not only lose the order value, merchants incur sunk acquisition costs and risk damaging their reputation with each false decline. It's also a negative experience for customers: 40% do not return to a merchant after a false decline, resulting in lost future business. The challenge lies in identifying orders that are statistically risky but don't fit clear fraud patterns. Merchants want to prevent fraud, but some of these orders are legitimate but flagged as false positives.

The key is balancing a seamless checkout experience for legitimate customers with robust fraud prevention. Requesting additional verification for every order might prevent fraud but increase cart abandonment. A balanced approach uses verification only when necessary and employs sophisticated fraud detection.

Adaptive Checkout addresses this challenge by tailoring each checkout flow to the order's risk profile. This minimizes friction for low-risk orders and requests additional verification only when needed.

The process begins by filtering out blatant fraud before authorization and enriching orders with additional data, making it easier for banks to identify and approve legitimate customers. By surgically analyzing each order's risk, even riskier orders have a better chance of approval, turning potential false declines into approved transactions. This selective verification approach is crucial for improving conversion rates without sacrificing fraud protection or positive checkout experiences for legitimate customers.

2. How do you integrate machine learning and behavioral analytics to identify fraud patterns? 

The power of machine learning is the ability for it to ‘learn’ independently as it runs. There are various ways to implement this, but we believe the best approach is a layered one. 

It starts with having vast amounts of quality data. Riskified trains its machine learning models on hundreds of millions of data touch points from across our global merchant network. Our algorithms learn to distinguish safe behavior from suspicious patterns and stop existing fraudulent behaviors. But to truly stay ahead of emerging fraud tactics, we also apply real-time anomaly detection using unsupervised machine learning, which flags unusual behavior based on combinations of various order characteristics — such as many orders suddenly exhibiting copy-pasting of credit card details alongside proxy use. Our engine can detect it and stop fraud  MOs before they spread. 

Machine learning also adapts the checkout flow for every transaction, surgically applying additional security measures for select higher-risk orders and enabling merchants to confidently approve more genuine orders while blocking fraud at various stages of the process.

3. What impact does AI-powered fraud prevention have on cart abandonment rates and conversion optimization? 

Cart abandonment has many causes, but a long or overly complicated checkout process is a major one. Consider an order that initially appears suspicious, such as one placed from a new device or with a recently issued credit card. Using AI and Riskified’s merchant network data, we can compare an order against millions of data touchpoints to draw a clear picture of a shopper’s true identity. This allows us to identify legitimate customers and expedite their checkout. Conversely, imagine a returning customer with a stored credit card and no unusual activity in their order data.Why ask them for a CVV if there is no need? Some customers may not have this information readily available, potentially leading to cart abandonment. AI enables us to precisely request verification, such as a CVV or a one-time passcode, only when needed. This reduces checkout friction and increases successful conversions of legitimate orders.

4. What best practices should businesses follow when implementing AI-driven payment security solutions? 

It’s important to think of fraud prevention and conversion as two sides of the same coin. Make sure that whatever solution you’re assessing for payment security also helps to optimize conversions and prevent the false declines.

Make sure to connect with other stakeholders to understand the full scope of security issues related to payments. For example, does the payment security solution address post-purchase concerns like returns and policy abuse? Is there a need to protect against abusive behaviors like serial returns or false claims of INR? To empower merchant fraud and customer service teams with real-time ability to address these challenges, ideally, the fewer integrations you have to deal with, the better. 

Lastly, look for more than a solution - look for a partner that future-proofs your business. AI and ML solutions should never be a black box - merchant teams need technology that provide them with the visibility, flexibility, and control they need to tailor solutions aligned with their business strategy and success.  

For example, when you have declines, do you understand why they were declined? Can you add your own logic into the decisioning? How can you get a clearer image of the customer’s identity? Mapping this out can help you choose the right solution and make decisions that will improve revenue and security in the long run. 

5. How does AI-powered checkout impact payment authorization rates across different industries? 

Machine learning-based solutions can detect far more patterns than a rules-based solution – and unlike static rules, they can adapt and learn in real time. 

Using a sophisticated AI-powered solution that detects and screens out fraud prior to issuer authorization ensures that fewer fraudulent orders reach the issuer. Elite fraud prevention solutions are able to analyze high- quality data and share, that data with issuers at scale. This helps them filter out fraud while understanding the context behind orders better. For example, there can be a significant difference between the risk of buying a $1000 fridge online versus buying a $100 gift card at the same merchant. AI solutions provide the context that distinguishes between safe and risky orders. In the long run, this will lead to more trust with the issuer, higher authorization rates, and less risk of falling into a monitoring program.

6. How can AI-powered fraud prevention solutions be customized to meet specific business needs? 

The ideal AI-powered solution should be very customizable. No two businesses are alike, no one can decide your risk tolerance, and no one knows exactly how your business operates – so generic products won't cut it. 

Start by making sure fraud is viewed properly. Alongside the standard ML models, Riskified has developedgeo-specific, vertical-specific and sometimes even merchant-specific models, all of which is crucial for accuracy. 

Also, when you review transactions for fraud, do you want to review them after authorization, where you can manually overturn declines and leverage additional data like AVS match in the US? Do you prefer to review orders pre-authorizatio? Or maybe leverage both pre- and post-auth review?

For example, policy abuse is extremely specific to the merchant and their defined policies, so having ML to determine risk is an elemental part of the equation. And each merchant will want to deal with policy abusers differently – warning them, blocking them at checkout, denying claims, etc. 

Even in CNP fraud, how strict do you want to be? If you want to send a one-time password to some risky orders with verifiable phone numbers, how long would you give your customers to verify the code? High-end fashion offering limited stock would have a different view from fast fashion. And some forms of verification, like 3DS, will work better with consumers in one region, and less well in others. 

There is no one-size-fits-all solution, and it will look different from merchant to merchant, as it should.