Why We Built Berbix: Lessons Learned from Airbnb

We like to only half-jokingly say that the starting state of the proactive Trust & Safety team at Airbnb was a product manager, an engineer, half of a data scientist and an already-signed contract with one of the now-legacy photo ID verification providers. It had previously been decided that photo ID verification would be part of the platform from early on, and it turned out to be an incredibly prescient decision.

On the early Trust & Safety team, our mandate was simple but broad: stop all bad things from happening on the platform. While simply stated, it was an incredible challenge. Many companies have anti-fraud efforts that can follow a fairly standard playbook that was established in the early days of the internet in order to stop credit card fraud, account takeovers, phishing, etc. And we certainly were able to follow the basic playbook to much success.

However, there were few platforms like Airbnb whose core product spans not only the typical online interactions, but very sensitive offline interactions as well. There are few activities more sensitive than staying in the home of someone you’ve never met before, and we had to not only protect those interactions, but build trust in a platform that was blazing new trails in offline interactions.

We used every technique under the sun in our efforts to stop all bad things from happening on Airbnb. And while perfection is an impossible standard to meet when it comes to human-to-human interactions, we were largely successful. But in our vast experimentation to try to address the unique challenges we faced, we found that one tool was particularly effective at stopping intentionally fraudulent users: government-issued photo IDs. Fraud is often a numbers game, and when you increase the cost of fraud, you drive those bad users to other platforms. Unlike common digital assets that are used for basic verification like emails and phone numbers, acquiring a legitimate-enough photo ID to fool a human is expensive.

While this technique turned out to be quite effective at stopping these bad actors, it wasn’t without major challenges.

First, it was a terrible user experience. Uploading your photo ID is an understandably sensitive interaction. But the legacy providers are primarily human driven, even if their marketing suggests otherwise. This means that you typically have minutes-long waits from the time your ID is uploaded until you get a result. And in the event that it’s inconclusive or rejected, you would have to come back and try again. Getting someone to come back and try again is extremely challenging, and was the source of terrible customer experience tickets as reservations could be canceled due to someone missing an email.

Next, it was incredibly expensive on the technical side. The amount of engineering hours required to not only build the basic functionality, but to handle the complex asynchronous responses was immense. We needed to build an entirely new reservation status in order to support reservations in the ID check purgatory. Furthermore, in addition to the asynchronous mechanisms, we also had to build polling as sometimes we’d never get the webhook back from the provider.

Additionally, it was incredibly expensive due to the inaccuracy of the results we’d receive. We ended up staffing an entire team of operations folks to double check the results coming back from our ID verification provider due to our very high level of distrust in the outcome. This meant that there were often two humans in the loop for every ID check that could occur.

Lastly, it was expensive from a dollar cost perspective. When it’s a person who is checking the ID and each check takes potentially a few minutes, you’re looking at costs that understandably balloon considerably, even if they’re only checking some fraction of the IDs.

It was for these reasons that my cofounder and I decided to build Berbix. Given the incredible value that can be derived from checking a photo ID in terms of the ability to deter bad actors, we knew there was a way to avoid the expensive challenges above. And we’ve solved each of these problems by leveraging the major technological advances that have occurred over the last several years.

The two primary advances that Berbix has achieved over and above the competition are:

  1. Our product is truly fully automated, enabling us to provide instant processing of every step of the verification process
  2. We’ve built novel ways to detect fake IDs that enable us to catch in some cases 4x more fraud than the legacy providers

A key insight that we had early on at Berbix is that the ability to successfully provide a conclusive photo ID verification result is highly dependent on the quality of the images that are collected during the user experience. If the images are blurry, have glare, or are otherwise hard to read, it’s going to be extremely challenging for an automated (or human driven) process to provide conclusive results. It was this key insight that enabled us to build a product that wins in head-to-head live user tests against every vendor we’ve competed against.

On the user experience, we’ve built a delightful process that coaches users every step of the way. Because our system is truly fully automated, we are able to process images in a fraction of a second. This enables us to let the user know that an image that they’ve provided won’t be sufficient for a conclusive result and to ask them to try again. Because the user is already in that part of the verification process, asking them to try again is a low lift vs. asking them to come back once an inconclusive result has been returned. We call this Live User Coaching, and it is a key element to our high conversion rates.

On the technical complexity side, because we’re truly fully automated, we’re processing each image as it’s received by our infrastructure. This means that by the time the user has completed the collection process, we are already done processing all of their information. We call this technique Progressive Verification. This means that Berbix is able to give synchronous results each and every time. This is a game changer for technical teams that otherwise need to account for minutes-to-hours long delays.

On the accuracy front, because of the importance of image quality to provide conclusive results, our ability to prompt a user to try again instantly enables us to collect images of sufficient quality to provide accurate, instant results. Furthermore, our focus on machine readable components enables us to avoid common pitfalls that occur when you only derive information from human readable components via optical character recognition.

Lastly, from a dollar cost perspective, we’re able to provide greater value than the legacy providers due to the innovations outlined above. And the fully automated nature of our product enables us to be very competitive when it comes to our price. We work with our customers to ensure that the price they pay is commensurate with the value that they’re able to derive from our solution.

We’ve built a much better mouse trap, and we’ve proven that time and time again in head-to-head tests against our competitors, both legacy and newcomers. We have prevailed in every live A/B test any of our customers have run due to our emphasis on user experience, customer experience, and accuracy.

If any of the above resonates with you and you’d like to find out if Berbix can help you solve your identity verification challenges, reach out to us at berbix.com/contact.