Yellowcard Alia

Yellowcard had two million customers and a customer support team drowning in tickets. Most of the volume was the same handful of questions: how to deposit, how to trade, why something was pending. Alia was built to answer those automatically, in the app, before a customer ever needed to open a ticket.

I designed the full Alia experience: conversational flows, edge cases, the feedback loop, and the escalation path to human support. Every conversation either resolves on its own or hands off cleanly to a human, by design, so no customer ever hits a dead end talking to a bot about their own money. This is the story of what it took to make an AI assistant feel trustworthy inside a crypto product.

Support volume was high and growing. The team was spending most of their time on low-complexity, high-frequency queries that had known answers. That left less bandwidth for the complicated cases that actually needed a human. Something had to handle the first line.

The answer was in-app. Customers shouldn’t have to leave to find help. If we could surface accurate answers at the exact moment someone was confused, we’d deflect tickets and actually make the product better at the same time.

The hardest part of designing an AI chat experience isn’t the happy path. It’s making sure someone who doesn’t know what to type isn’t just staring at a blank input. We started every conversation with four suggested questions pulled from the top support queries, enough to orient someone without overwhelming them.

The copy had to feel human, not robotic. Yellowcard’s brand voice is intentionally warm and the AI couldn’t be the cold, corporate outlier. Every response was written to be concise, scannable, and in a tone a real person would actually use. We also built in clear escalation paths so a customer who couldn’t get what they needed from Alia could always reach a human without feeling trapped.

Crypto customers are already skeptical by default. Putting an AI in between them and their money needed to be handled carefully. We added a disclaimer upfront so customers understood what Alia was and wasn’t, and we never tried to hide that they were talking to an AI. That transparency was non-negotiable.

We also added thumbs up/down feedback on every response so customers could flag when something wasn’t useful. And when they did, Alia didn’t just shrug. It offered to open a support ticket and hand things off to a human. The escape hatch was always one tap away, and that made the whole thing feel safer to use. The number I watched most closely wasn’t how many people used Alia, it was how many bailed straight to a human on their first try. Escalation rate settled around 9% after the first two weeks, mostly people testing what Alia could actually handle. If that number had stayed near the 20%+ it opened at, it would have told me the suggested questions were getting people in the door without actually helping them.

AI breaks in weird ways. People type gibberish. People type in other languages. People send incomplete sentences or total nonsense to test what happens. We had to design for all of it. Alia needed to respond gracefully when it couldn’t understand a query instead of returning something confusing or silent.

We also added a 160-character limit on inputs. Shorter prompts meant more precise queries, fewer resources to process them, and faster responses. The character counter turned red at the limit so customers knew to trim before sending. Small thing, but it changed behavior in the right direction.

Designing for AI means designing for uncertainty. You can’t fully predict what a user will type or how the model will respond, so the interface has to handle both gracefully. That edge case work was the real product. Everything else was polish on top of it.

The other thing that stuck: the feedback loop matters as much as the launch. Thumbs up/down sounds small, but it’s the mechanism that makes the system learn. Analytics would tell us what Alia got asked. The feedback told us how well it actually helped.