Earlier this year, Stripe asked 100 employees to do something highly unusual: stop their day jobs and instead, dream up features and functionality for the payment platform using the newest generation of OpenAI’s language learning model, GPT-4. Engineers from teams spanning support, onboarding, risk, and docs considered where Stripe could use artificial intelligence that understands free-form text and images and builds human-like responses, to improve—or alter—a feature or workflow.
“Our mission was to identify products and workflows across Stripe that could be accelerated by large language models and to really understand where LLMs work well today and where they still struggle,” says Eugene Mann, product lead for Stripe’s Applied Machine Learning team. “But just having access to GPT-4 enabled us to realize, ‘Oh, there are all these problems that could be solved with GPT surprisingly well.’ ”
Stripe powers the payments of small and large businesses across the internet. While Stripe builds their ecosystem to support all aspects of the payment process, developers—those building with and integrating Stripe’s software—are their primary users. The more adept Stripe developers get at deploying Stripe, the farther Stripe will spread through the digital payments universe and grow the GDP of the internet.
Stripe had previously been using GPT-3 to help their support team better serve users through tasks like routing issue tickets and summarizing a user’s question.
Stripe’s team put together a list of 50 potential applications to test GPT-4; and after vetting and testing, 15 of the prototypes were considered strong candidates to be integrated into the platform, including support customization, answering questions about support, and fraud detection.
Better understanding users’ businesses
To better serve users and provide the right type of support, Stripe tries to understand exactly how each business uses the platform and customizes support accordingly. It’s an obvious and important step, but one that requires a lot of human hours.
“Many businesses, such as nightclubs, keep their websites sparse and mysterious so it can take a great deal of searching and clicking to understand what’s going on,” says Mann.
Now, Stripe uses GPT-4 to scan these sites and deliver a summary, which outperforms those written by people.
“When we started hand-checking the results, we realized, ‘Wait a minute, the humans were wrong and the model was right.’” Mann says. “GPT-4 was basically better than human reviewers.”
Answering support questions about documentation
Another critical way Stripe supports developers is through extensive technical documentation and a robust developer support team to answer technical questions or troubleshoot issues. GPT-4 is able to digest, understand and become that virtual assistant—almost instantly.
“GPT just works out of the box,” he says. “That’s not how you’d expect software for large language models to work.”
GPT is able to understand the user’s question, read detailed documentation on their behalf, identify the relevant section and summarize the solution in one place.
Fraud detection on community platforms
There's also the need to manage malicious or bad actors. Stripe maintains a robust community on forums like Discord, which not only crowdsources help for niche technical questions but raises the visibility of developers for future work. However—because it is on the internet—malicious actors find their way into these forums, often trying to gain critical information from community members or regain credibility with Stripe's community team after being kicked off the platform.
Just by analyzing the syntax of posts in Discord, GPT-4 has been flagging accounts where Stripe's fraud team should follow up and be sure it isn't, in fact, a fraudster playing nice. GPT-4 can help scan inbound communications, identifying coordinated activity from malicious actors.
Now the Stripe team is thinking about the next round of features. GPT could be deployed as a business coach that can understand revenue models or advise businesses on strategies. With the potential applications only broadening as GPT gets smarter over time.
Mann says he and his team are essentially painting with a new canvas every day.
“Some of these features,” he says after listing all the ways Stripe has had success so far, “feel like magic.”