Sendbird’s founder built messaging infrastructure for 300 million users. Now he’s betting the future of customer service is memory, not more chatbots.
John Kim spent the past decade building Sendbird into the messaging infrastructure behind more than 300 million users. Now he is betting against his own creation. Kim’s newest venture, Delight.ai, starts from a simple critique: most AI customer service is still transactional, offering an answer and little else before the relationship resets the next time a customer reaches out. His alternative is an AI concierge that remembers who a customer is and acts on that memory, a proposition that raises as many questions about surveillance as it answers about convenience.
In this exclusive interview with us, Kim discusses the guardrails built into Delight.ai’s Trust OS, the instinct from his first startup, Paprika Lab, that he’s had to unlearn, and why he believes AI will redirect customer experience roles rather than eliminate them.
Full interview;
Sendbird built a chat infrastructure for 300 million users, then launched Delight.ai, claiming AI customer service is broken. What did you see in your own product that convinced you to disrupt your own core business?
Building Sendbird gave us a front-row seat to hundreds of millions of customer interactions, and what became clear to us was that the next wave of customer experience would be less about making support faster and more about creating an experience that felt deeply personal. Most AI customer service today is largely transactional. You ask a question, it finds an answer, and the interaction ends there. But that’s not how great human service works.
Think about walking into a store where the salesperson already knows you. They remember what you bought last time, what you liked and didn’t like, and can make a recommendation based on what they know about you, so you don’t have to start from scratch. That’s what a great AI concierge does: it understands who you are and has the context to anticipate what you might need next.
That’s the opportunity we saw with Delight.ai. The next generation of customer experience will be built around AI concierges that develop an ongoing understanding of each customer and use that context to take action on their behalf and proactively anticipate their needs. The shift is from AI that just answers questions to AI that actually knows and serves the customer.
Also Read: Cookies Are Dying. Advertisers Are Fighting Over Scraps
Delight.ai relies on “long-term memory” to remember past interactions. Where is the line between an agent that genuinely delights and one that feels like quiet surveillance?
The line is whether memory is being used to serve the customer or to know more about them. Our 2026 Delight AI Index found that 70 percent of customers actually want AI to remember them. People don’t want to constantly repeat their problems or start from scratch every time they interact with a brand. However, wanting AI to remember you doesn’t mean giving it unlimited permission. Only 16 percent of consumers are comfortable with AI acting completely autonomously, and 47 percent want to approve actions before they happen. That tells us customers want personalization, but they also want control. A great concierge works the exact same way. The goal should be to use memory to remove friction and make the experience more personal, while giving customers control and transparency over how AI acts on their behalf.
You built a “Trust OS” to govern enterprise AI agents. What is one action you’ve deliberately forbidden an agent from taking — no matter how much revenue it could unlock?
We’ve drawn a hard line around AI overriding human-defined guardrails. If a company determines that an action requires approval or sets specific parameters for what an agent can do, the agent must operate within those boundaries. That philosophy is built into Trust OS. We give companies the ability to gradually roll out changes, pause them, or roll them back. The more capable AI agents become, the more important these boundaries become. If you can’t explain why an AI agent acted the way it did and intervene when necessary, you shouldn’t give it authority in the first place.
Also Read: When AI Agents Do the Shopping, Who Wins Loyalty?
You turned Sendbird into an AI-native company, complete with non-tech teams building their own tools and an internal system of quests. What was the most unexpected failure or success of that culture experiment?
We expected our engineers to embrace AI quickly, but what we didn’t anticipate was how quickly people across marketing, recruiting, and other non-technical teams would start identifying problems and building their own solutions.
A big part of that was getting the infrastructure right first. We developed an AI Transformation Model that moves from one-off AI to connected AI to autonomous AI and, ultimately, to AI-native, where agents run critical workflows across teams. We learned that you have to standardize before you scale: where apps live, how permissions work, what services are allowed, and how things get built, so employees have a path that is already safe and structured.
For example, we created Automators as an internal marketplace where anyone can post a “quest” for something they want automated or built, and AI takes the first pass at scoping and building it. From there, anyone in the company can help take it across the finish line. That also opened the door to do things we never would have put on an engineering roadmap. Our marketing team created a fully functional e-commerce swag store in under two days, complete with Stripe integration, custom designs, and even a Konami Code Easter egg that unlocked secret details about our annual conference, Spark.
What I’ve learned is that AI-native companies fundamentally change who gets to be the builder. When employees can turn an idea or problem into a working solution themselves, you remove many of the bottlenecks around innovation.
Paprika Lab was acquired before Sendbird existed. What’s one gut instinct that served you as a first-time founder that you’ve had to consciously unlearn this time around?
With my first company, my instinct was to build something people wanted to use. We experimented constantly, followed where we saw engagement, and ultimately built Paprika Lab to millions of users. With Sendbird and Delight.ai, I think much more about the depth of the problems we’re solving. Are we actually changing the way a customer experiences a brand? Can we build an AI concierge that understands a customer and anticipates their needs before they ask? In this stage, I am focused on whether we’re solving a problem deeply enough to fundamentally change customer experience forever.
Also Read: The Loyalty Assumptions Holding Marketers Back
If autonomous agents start handling transactions and memory at scale, what is the first specific job title in a traditional CX team that stops existing?
I don’t think this is necessarily about a specific job role or title disappearing. What will change most quickly is the type of work CX teams spend their time doing, particularly the repetitive, transactional tasks that still consume a large part of their day. With agentic AI, I think CX teams will spend more time designing the customer experience, defining how agents should behave, and stepping in when judgment or empathy matters. AI won’t eliminate the need for people in customer experience, but it will change where human expertise and talent are most valuable.
