Ever since the launch of OpenAI’s ChatGPT on November 30, 2022, you’d be hard-pressed to find a company that’s not building something related to AI. Especially in the customer support space, where it seems like there’s a new AI product announcement every other day.
But how many of those companies have a specially dedicated team of four who spend most of their days in a small, windowless room they affectionately call “The Dungeon”? Surely none of them have a “Live, Laugh, Love” sign hanging on the wall.
When I asked the team behind Assembled’s forthcoming AI product to describe their culture in three words, they pointed to the sign and laughed.
“We actually have four words,” said John Wang, an Assembled co-founder who’s been leading the team since its inception earlier this year. “Make something people want.”
Assembled has entered the chat(bot)
Assembled wasn’t the first company to enter the AI arms race, and it certainly won’t be the last. But despite all the competition it’s up against, the team — known internally as Team New Products — is convinced it’s anyone’s game.
“I don’t think anyone’s nailed it yet,” John said. “Some companies are putting forth great efforts, but the opportunity is wide open.”
Kaytlin Louton, Jason Ma, and Cameron Skarritt are the other three members of Team New Products. Add John to the mix and you’re looking at a team with 13 years of collective Assembled experience. That’s a lot of tenure for a company that launched in 2018.
Since April of this year, this small group of veteran Assemblers has been living, breathing, and eating AI — as well as soaking up whatever wisdom they can from successful product leaders by hosting them for educational dinners.
Borrowing a page from the Y Combinator playbook, Team New Products was initially given three months to explore the potential of leveraging AI in a new product. That sprint ultimately proved fruitful, and in July, the team received the green light to go all in.
“This is definitely a bet, but I think it's a strategic bet,” John said. “Support is something that we understand better than a lot of other folks. And we have a lot of existing customers that can benefit from our AI product being really good.”
Selling the dream to the rest of the team
For all the optimism on Team New Products, it took some convincing on their part to get other folks throughout the company on board.
While Team New Products was holed up in The Dungeon, Assembled was in the process of signing and onboarding multiple customers — including some with incredibly complex, hands-on implementations. If ever there was a time when the collective institutional knowledge of Team New Products would come in handy, this would be it.
Other Assemblers wanted to know: Why are we diverting the focus of four of our most tenured employees? What makes us so sure we can win in a clearly crowded market? And how does this side project relate to Assembled’s core focus, which is workforce management?
This line of questioning had Team New Products on double duty. In addition to building a viable product as fast as humanly possible, the team was working overtime to educate the rest of the company about the big-picture opportunity.
“These are all very valid concerns,” John said. “We’re the ones who see the impact our product is having on customers, and it’s our job to communicate that to the wider team.”
“At the end of the day, I think it becomes a decision about prioritizing the next possible frontier,” he continued. “We as a company should continually be pushing the boundary.”
Please hold while I connect your Cal
Team New Products was officially formed on April 17, 2023, but the idea for Cal — the name affectionately given to the AI product that the team has poured itself into over the past seven months — was actually seeded a couple of months earlier in February.
John and his fellow Assembled co-founders, Brian Sze and CEO Ryan Wang, had long pondered potential applications of AI to create efficiencies in customer support operations. In fact, the very idea for Assembled was born out of the founders' first foray into building support tools for Stripe, where they leveraged machine learning to make agents more efficient with a tool called Agent Assister.
But the release of OpenAI’s chatbot, based on large-language model (LLM) technology, changed everything.
“We have been thinking about LLMs for a very long time,” John said. “But when ChatGPT came out, we were ready to say, hey, let's do this really fast.”
“A chatbot felt like an obvious place to start,” Jason said. “So I built a very simple Slackbot prototype that could help by pulling in sources and answering questions. That was our first iteration of Cal.”
Don’t call it a chatbot
If you’re like most people, the word “chatbot” probably calls to mind some of your more frustrating customer service experiences. You know, the ones where it feels like that chatbot is doing everything in its power to prevent you from speaking to a real person — all while failing to answer your question or resolve your issue?
This is a tactic known as deflection, and its goal is to handle customer support queries without the assistance of a living, breathing support agent.
“Most of our competitors are focused on deflection,” Kaytlin said. “They’re working to cut the number of contacts coming in because they believe this will make the most money for the companies that buy their products.”
“Our approach to AI is actually quite different,” she continued. “We’re looking at the holistic lifecycle for support; and what we’re finding is that, yes, deflection does decrease the number of contacts, but cost per case is still being driven by handle time.”
In other words, Cal isn’t intended to replace what agents do — it’s intended to help them do their jobs faster and more efficiently. And it’s leveraging the expertise of existing agents to become more effective over time.
“How do we make agents more efficient? How do we make sure that each interaction is as productive as possible?” John asked. “We start by helping agents quickly find answers and write replies.”
Team New Products is betting that this approach to AI will reduce support costs without sacrificing customer or employee satisfaction. And so far, their hypothesis isn’t just supported by customer feedback — it’s driven by it.
Making something people want
Prior to working on Cal, every member of Team New Products had varying degrees of customer interaction over the years. But they all agree that their roles on this special projects team are more customer-facing than ever before.
“We’ve always interacted with customers in some capacity,” Cameron said. “But this is that dialed up to 11.”
Suffice it to say, this team spends a lot of time talking to customers.
“Nearly everything we build comes from a user problem that we’ve observed by watching agents do their work,” John said. “Either that or we’ve heard it come up in multiple conversations with different users.”
When the team launched a new feature for alpha users that would automatically generate replies, for example, every person they introduced it to in a span of two days had the same question: “How do I interact with it?”
That was all the team needed to hear. In the three days that followed, they re-engineered the feature to be more interactive. Turns out this was a good move, as updating this one feature increased Cal’s usage by 50%.
“We literally built this because customers told us to,” Kaytlin said. “We hear a problem and we build a solution.”
“The customer is the expert on the problem, not the solution,” John added. “It’s our job to figure out the solution.”
Moving fast and breaking things
In the race to fine-tune the solution before its competitors, Team New Products certainly doesn’t have the largest team working to solve this problem. But what it lacks in person-hours, it more than makes up for in speed and agility.
“There are a lot of processes that we don't have to deal with because we don't have an existing enterprise, so we can move quicker,” Cameron said. “I think that's just a byproduct of starting from base level.”
“There’s a famous YC (Y Combinator) tenet,” Jason said. “I actually don't remember the quote anymore. But basically, don't focus on building scalability until you've actually nailed something.”
John jumped in: “Nail it before you scale it.”
“Yeah, nail it before you scale it,” Jason continued. “Assembled’s workforce management platform is at a certain size and product maturity where, as an engineer, you should spend more time thinking through what it is you’re building. But we had to toss that mentality early on, because it’s possible that something we build tomorrow won’t be used by the end of the week.”
“There’s a lot of experimentation still ongoing,” Cameron added. “It’s a different process entirely.”
Of course, it’s hard to experiment quickly without breaking a few things in the process.
“Cam and Jason break a lot of things,” Kaytlin said. “John probably does too, but I'm not sure what he breaks.”
“Just yesterday, Cam broke something that increased our OpenAI bill by like 20×,” she added.
Things actually break so often that Kaytlin once suggested gamifying it with a board that keeps track of who breaks the most things.
“My idea was that, at the end of the week, whoever broke the least things would get $100 or something,” she said. “But John wasn’t into it because it would discourage breaking things too much.”
“The easiest way to not break anything is to not ship anything,” John said.
Cal today, Cal tomorrow
For its part, Team New Products has done a lot of shipping in a short amount of time. But in many ways, they’re just getting started.
“The things that make me excited about Cal’s potential are seeing a graph of our usage go continuously up, and early users like Honeylove centering their OKRs (objectives and key results) around increasing Cal’s usage and improving Cal’s accuracy,” John said.
Cal is currently in beta with a handful of Assembled customers on Zendesk. And it’s proven especially useful in the consumer goods space. Come early December, the team is expected to make Cal available to all Assembled customers on Zendesk.
Looking ahead, the team has even bigger ambitions for Cal. This includes going after non-Assembled customers, more help desks, and more industries with increased functionality.
This desire to continually push the boundary is rooted in the Assembled ethos of tackling difficult but worthy challenges head-on.
“Everyone is trying to use generative AI right now,” Kaytlin said. “It's just that most people are using it for the easiest application, whereas we're using it for the hardest application.”
“Across the company, we shouldn't be afraid to try and seize opportunities that are out there for us,” John said. “We also shouldn’t be afraid to work hard to make those things happen — even if it feels like pretty long odds.
