Deliveroo is growing fast. From a single bike that its founder Will Shu rode in 2013 to tens of thousands of orders per day and a $1bn valuation in four short years. It is now in over 140 cities across the globe and is a very familiar sight in many of Europe’s cities. In the words of Will Shu, Deliveroo is now part of the lexicon.
In 2016, as the company’s visibility and popularity grew, fraudsters began to target the company in earnest. By the Spring, it was clear that steps were going to have to be taken to secure the business against chargebacks that were becoming extremely costly.
“When we began to realise that we really needed to tackle fraud, we initially assumed we would have to hire a fraud manager and then a team to build capability internally. We really didn’t want to do that for cost, speed and efficiency reasons. We didn’t and don’t see fraud detection as a core function for us. The team at Ravelin told us about their approach and we were immediately convinced that Ravelin was the way to go.” - Emma Whibley, Finance Manager at Deliveroo
Ravelin’s approach is to use machine learning as the principal technique to predict whether any order is likely to be fraudulent or not. Based on Deliveroo’s own historical chargeback data, Ravelin applies its proprietary algorithms to score all activity for fraud in real time and feed the result back to Deliveroo’s ordering system. This means that fraudulent orders are blocked before they are accepted by the restaurant.
Deliveroo sets the risk threshold they are willing to accept and can vary it for different markets. Once set, prevention is entirely automatic. There is no manual review of orders and no escalation process. Deliveroo and Ravelin work together to assess if the threshold is correct per market based on the impact on chargebacks and the incidence of false positives (where a good customer is blocked by mistake).
Deliveroo went live with Ravelin in July 2016 and the impact was immediate. Within four months fraud had been slashed by half, saving the company significant bottom line revenue - the most immediate target of their fraud prevention strategy.
It is a feature of machine learning models that they improve with time - exposure to live data and feedback from the the customer service team allow the machines to learn more of the specifics of a client’s data.
So it was with Deliveroo who continued to see a significant drop in fraud to where it is today - at a rate that would stand comparison with any peer in any industry. Absolute chargeback fees today are a quarter of what they were a year ago. Impressive on its own but when considered against a background of continued stellar revenue growth for the company, it is even more impressive.
A 90% reduction in chargebacks overall was achieved with a very close watch across the company on the acceptance rate for orders and an absolute commitment to make the signup process for new customers as simple as possible.
Emma Whibley commented: “the feeling within Deliveroo is that Ravelin is doing a great job.”
Fraud is largely managed within the customer service team at Deliveroo and not in a specialist fraud team. Ravelin scores all users and fraud decisions are taken automatically. The customer service team has a vital role in managing customers who contact them to query blocked orders.
Deliveroo uses Ravelin’s dashboard and graph network tool - Ravelin Connect - to manage these queries and assess whether customers calling in because of blocked orders are genuine customers or potential fraudsters. It is well known that fraudsters are significantly more likely to call to complain if an order is blocked - hoping to convince the customer service executive there has been an error.
Andreia Silva, Senior Customer Account Executive, explained: “I could not do my job without Ravelin. It’s the #1 tool we use to see if a customer is committing fraud or not.”
Ravelin Connect shows Andreia and the team instantly whether a customer is connected to a fraudster or a chargeback by a phone number, a device, a physical location, or a credit card or some combination of the above. In months of working this way she has made the wrong call only a handful of times in thousands of reviews.
The result is a super-fast assessment of the merits of a claim. Genuine customers can be quickly approved and the feedback sent instantly back to the system. Fraudsters can be declined quickly and accurately ensuring no loss of revenue. More importantly that account can also be flagged and related accounts discovered stopping the fraudster from simply trying again with a new account.
Matt Barker, team leader for Customer Accounts, concluded: “we are now building our processes around Ravelin rather than the other way around. It is central not just to fraud but to customer success.”
Max de Grunwald is the product manager for Fraud within Deliveroo. He meets regularly with Ravelin to discuss future developments:
“The Ravelin team are really responsive to our suggestions. For instance we have had made a lot of requests for the functionality of the rules within Ravelin to be extended and for the creation of some rules to match Deliveroo policies. Ravelin has been quick to understand and implement these changes. It’s a real pleasure to know that we are being listened to and our ideas are implemented.”
Rules sit alongside the fraud detection models and help keep chargebacks and false positives at a very low rate for Deliveroo. What’s more, Deliveroo continues to add cities and markets to its portfolio and can expand with the confidence that a new market does not mean a greater fraud risk. Ravelin is flexible enough to vary the risk threshold per territory and works with whichever PSP Deliveroo is using in each country.
Martin Sweeney, CEO of Ravelin concluded: “Ravelin’s impact at Deliveroo shows how using machine learning techniques to extract the right signals from a merchant's data is a fast track to great results. More important than the technology though is the co-operation between the teams at Deliveroo and Ravelin. An open and honest approach to tackling this fraud problem has been the real key to getting fraud low and keeping it low.”