AI fraud rules: Build rules using natural language (new feature on the Ravelin Dashboard)

Artificial intelligence has been touted as one of the biggest tech innovations of our time.

Granted, some of the AI hype might be questionable, but in the world of fraud prevention, machine learning is already recognized as near-essential, especially for bigger organizations as well as those that need to scale operations on a regular basis.

But there is also plenty of untapped potential in other applications of AI and generative AI technology for and within fraud prevention solutions and strategies.

Today, we’re going to present the latest way AI tech is improving the Ravelin product: AI fraud rules.

What are Ravelin’s AI fraud rules?

Released in June of 2024 and just out of Beta and available to all users, Ravelin’s AI rule generator is a new feature on the Dashboard. Instead of scrolling through the dozens of rule-generating options and building rules manually, fraud managers and analysts can use Ravelin’s AI rule generator to speed up their work.

Just type out what you want the rule to do and Ravelin’s system and the large language model (LLM) will generate the rule for you.

Additionally, AI-generated rules can be edited, added to, and tested – just like any rule on the Ravelin platform.

"Which types of AI are or can prove useful in the fight against fraud?" Responses from Ravelin's Fraud Survey 2024.

How can LLM technology help stop fraud?

Because they understand (and use) natural language, LLMs are thought of as a valuable tool in the battle against fraud.

In the case of the new AI rules generator, they are a quick and easy way to add new fraud rules to the Ravelin platform, allowing you to take swift action against real-time fraud attacks.

In fact, in Ravelin’s Fraud & Payments Survey 2024, 46% of respondents around the world said that large language models (LLMs) could prove useful in the fight against fraud. The figure jumps to 54.2% for respondents in senior leadership. It’s not a stretch to gather that LLM features in a fraud platform can prove very useful in getting C-suite buy-in.

How does AI work for fraud rule generation?

In addition to generating text of various types, an LLM has the benefit of understanding natural language. Therefore, on fraud prevention platforms such as Ravelin, it can speed up the creation and editing of complex fraud rules, saving time and effort and allowing for faster response times to real-time challenges.

Here's how AI fraud rules work on Ravelin:

1. Navigate to Rules > Create rule.

2. Select a checkpoint for the rule to apply to. You will also want to select a rule category.

3. At this point, you would normally need to build the rule using various conditions from drop-down and search menus. Instead, select Build with AI.

4. In the prompt that shows up, phrase your rule using natural language. This means that you should ask the AI to do this in the same way you might ask a colleague.

   • The AI understands synonyms, such as block for prevent.

   • You can combine several conditions in one request, for example checking users’ IP address plus their device, but excluding customers who have already bought more than $500 in products in the past year.

5. Select Generate.

6. Carefully read the generated conditions, and ensure they all align with your intentions.

7. You can click on Estimate rule impact to see how this rule would affect your customers if it had been active in the past week.

8. Choose to Run in Test Mode or Run in Live Mode and name your rule.

   • We recommend publishing in Test Mode first, where possible. This makes the rule passive and allows you to evaluate the performance and impact of the rule without affecting any customers.

9. Click Publish rule.

Tips for AI fraud rule creation

• You can phrase your rule in any language, which allows team members to use the rules generator in the language they are most confident in.

• We recommend running new rules in Test Mode first, unless responding in real time to an attack. This allows you to evaluate the rule, and see whether it is a good idea to implement it in earnest.

• On Ravelin, a rule can contain up to sixteen conditions, and the same is true for AI rules.

• Always read the outcome of the AI fraud rule generator to confirm it is as desired – you can easily make changes and adjustments, if needed, as well as add conditions.

Note: Although Ravelin is a firm believer in the power of machine learning in fraud prevention, we know that fraud rules can help in certain situations, such as a temporary measure in a new fraud attack or to ensure that a certain category of loyal customer always enjoys friction-free checkouts.

In some cases, rules block more requests than the ML models, but the ML models do a much better job at blocking more subtle cases of fraud, which is too difficult to stop with rules. The Ravelin team is always at hand to support with this and provide expert advice for your particular circumstances.

Examples of AI rules that can be built with Ravelin

Here are some examples of AI rule prompts that will generate a fraud rule on Ravelin:

1. “Always allow customers who are using a gmail email address, are in the market LatAm and have placed at least 200 orders”

2. “send the customers to manual review if they're placing an order over ¥2990 and have placed more than 5 orders in the last hour and are using a PS4.”

3. “block all customers from France using AMEX and values over 200 euro and who have failed 3DS more than 4 times, had more than 7 chargebacks and are using a risky IP, and are using the currency SEK”

4. “Senden Sie Transaktionen an 3DS, wenn Sie das Gateway „Supergate“ verwenden, eine Acquiring-Bank in Irland haben oder eine kostenlose E-Mail-Adresse verwenden”

5. “フランスの銀行が発行したカードを使用し、特定の電子メール アドレスを使用し、過去 1 日に 10 回以上注文に失敗した顧客を確認する”

6. "Block all payments! We're under attack!" (we wouldn’t recommend this – but on the other hand, we can guarantee that it will result in 0% payment fraud...)

AI-native fraud prevention

Ravelin prides itself in being an AI-native fraud and payments expert, putting cutting-edge tech to good use to prevent fraud and abuse.

In addition to AI fraud rules on our platform, we use sophisticated machine-learning fraud prevention models which, as well as other features, leverage natural language processing (NLP) to better understand text when making recommendations.

This means, for example, that the model will be able to take into account the name or the popularity of an item being added to a shopping basket, or even the type of discount applied and whether it might be promo abuse.

Parting thoughts

Remember that although rules have their place in fraud prevention, at Ravelin we believe that the best way to prevent fraud at scale is machine learning.

Consider letting your machine learning fraud model pick up the bulk of fraud, and instead use rules sparingly to test out new tactics, block new and different attacks, and ensure good customers enjoy the frictionless shopping journeys they deserve.

And when it’s time to write those rules, consider getting the help of our new AI rule creator.

Further reading

• Ravelin Help Center: Rules

• Rules are still relevant to fraud detection – here’s why

• Podcast: Leader, follower or laggard when it comes to AI?

• Ebook: Machine learning practices at Ravelin

• Machine learning and fraud prediction: What data is required to make it work?

• 3DS rule experimentation and optimization

• Ravelin insights: Machine learning for fraud detection

Could Ravelin's AI-native fraud prevention fuel your secure growth?
Book a call with Ravelin's fraud experts today.