The Support Hub is one of the primary ways merchants and partners can reach us when something goes wrong, whether they need to change a bank account or troubleshoot a terminal that stops working. The Support Hub already has the information users need. What we wanted to change was how users access that content and how they experience the support journey.
We wanted the Support Hub to offer the same AI-first experience already available in our Developer Portal. Developers get an AI-first experience while they're building against our APIs, and when they go live, they transition to the Support Hub. So why would they not receive the same AI-first experience there too?
The information didn't need to change, just how users access it
An AI-first interface can bring existing support content together in ways a static guide can't. Instead of searching through articles to find an answer, users can type out a question or use voice search to describe what they're trying to do or what's gone wrong. The Support Hub can then ask for more details, such as the product or device they're using, and use that context to guide them to the right information. As they work through an issue, they also get a checklist of steps to help them track their progress.
When users need more help, they can now raise a support ticket directly from the Support Hub instead of signing into one of our portals. When an issue escalates, our Operations Team sees the conversation history attached to the ticket, so they have the background information they need to take over.
The questions users ask also give us a record of where the existing content falls short. We can use those gaps to identify the content we need to improve or create next.
How long did it take to rebuild the Support Hub with AI
The Support Hub rebuild took about three days, and AI changed how we approached the work.
We started with a research task in Claude Code, looking at competitors and other industries to understand what we could bring into the Support Hub. As part of that research, we also looked at retrieval-augmented generation. We had a large collection of support articles, and we couldn't load all of that content into the model's context all at once. We needed a way for the AI to retrieve the right information based on what a user was asking, rather than giving it everything upfront. Retrieval-augmented generation became the answer.
From there, Claude worked through the look and feel, planning, and implementation, while we checked in periodically and guided the work. Because the content was already there, we could focus the rebuild on how users interact with it and create a better support experience.
Why testing was harder than building
The hardest part wasn't building the Support Hub. It was testing how someone would actually search for an answer.
Claude already knew the support content it was working with, so when we used it to test its own work, it naturally approached a question with knowledge of where the answer was. A real user doesn't have the same knowledge. We had to keep iterating on how Claude searched for and responded to questions so that the experience better matched how someone would actually use the Support Hub.
Once we had the technical direction in place, the implementation itself was surprisingly straightforward. Claude handled the work while we reviewed the results and guided it when needed.
The cost of moving faster
Speed to market is what matters most here. Getting a new experience to market faster means capturing that market sooner.
A lot of what shaped this rebuild came from the Developer Portal project before it. With AI, a full rebuild can now take days instead of months, making the cost of experimenting much lower. Getting the direction right on the first try is no longer the point. The goal is to try something, learn from it, and change direction quickly when it doesn't work.
The Support Hub also went through a couple of iterations, where we tried different designs, conversational styles, and features. Across those iterations, token spending came to roughly $1,000. The last full iteration, run entirely on a competing model to see how it compared with Claude, cost $253.19. That is a relatively small cost compared with the potential value of better customer satisfaction and lower attrition.
How much you spend also depends on how clearly you define the requirements. Clear, detailed requirements up front give the AI more to work with and cut the back and forth as it works out what's needed. That keeps the cost down and helps you reach the desired outcome with less wasted effort.
Build, learn, and iterate
Our advice for anyone taking on a similar rebuild is to start, try something, and iterate. With AI, the first attempt doesn't have to be the right one because the cost of getting it wrong is low. You can experiment, learn what works and what doesn't, and use each iteration to refine the result.