What Call Centers Teach Us About AI and the Future of Learning at Work – Ryan Wang
“AI is stripping away the rote parts of the job—so people can focus on the parts that make them human.”
Call center automation is the canary in the coal mine for how AI will reshape work and learning.
Before AI rewrites every industry, it’s already reinventing one of the largest labor markets in the world: customer service.
I wanted to understand what that shift actually looks like on the ground. How work is changing, what new skills are emerging, and whether AI is really making jobs better or just fewer.
So I sat down with Ryan Wang, CEO of Assembled, which builds AI agents that automate customer service work for some of the world’s largest brands. His view of automation is refreshingly different from the headlines: AI isn’t replacing workers wholesale. It’s changing what they do, how they learn, and what makes their work meaningful.
We talked about:
• Why full automation is still a myth—and what the Jevons paradox reveals about AI’s limits
• How AI is already making frontline jobs more creative, purposeful, and judgment-based
• What the call center can teach educators and employers about reskilling at scale
This transcript has been edited for clarity and length.
On Call Centers As the Canary in the Coal Mine
ALLISON: What’s the most common misconception you hear about call center automation, from the media, investors, or customers?
RYAN: Media and investors often assume call center jobs will be automated away. Yet every customer I speak with believes that’s completely untrue. Just three weeks ago, Sam Altman suggested call center work would be the first to vanish under AI. But Klarna’s experience tells a different story: after automating its entire customer service arm, it had to roll the move back. Rehiring proved difficult, and even executives ended up taking customer calls, learning firsthand why this work resists full automation.
On the surface, call center tasks look rote and repetitive, and some – like canceling a subscription – can be automated. But in practice, the systems handling sales, onboarding, tickets, and compliance are fragmented and unreliable. Agents juggle multiple internal and external tools across screens while making real-time judgments to meet customer needs and uphold policy. For all the usual reasons—change management, knowledge gaps, integration—it’s never as simple as waving a wand.
ALLISON: Technology will likely help us solve issues like information fragmentation and integration. Do you think we’ll eventually be able to fully automate customer service, just on a longer timeline than people expect? Or are you skeptical of that argument?
RYAN: I doubt we’ll ever reach full automation. In fact, we’re seeing a Jevons paradox in AI for customer service: even with AI agents and chatbots available, customers are contacting support more, asking more troubleshooting questions, and growing businesses are expanding their support teams.
To understand why, remember why customer support exists. It’s rarely about simple tasks that should already be available in the app. People reach out because they need help, access to information they can’t find, or human judgment in a moment of uncertainty.
No product is perfect, and products are always changing – which means customer questions are always changing. No automation will perfect customer service; if it could, the product itself would already be flawless. But that’s not how good products evolve or stay valuable.
ALLISON: You’ve said we’ll likely never reach full automation. When will we see fifty to eighty percent of call center tasks automated?
RYAN: It’s a moving target. The tasks we handle today won’t be the same in five years because products and their support needs keep evolving.
That said, about half of today’s tasks could be automated within three to five years, roughly one technology cycle. But many customers are still mid–digital transformation—they need to finish migrating to the cloud before stacking AI. To get there, they’ll need a clean, connected knowledge base to train chatbots, and agentic capabilities that integrate with billing and account systems. Every company runs on a patchwork of legacy tools, some custom, some out of the box. Connecting AI to all of them is complex. Transformation takes both time and sustained investment.
ALLISON: This response maps so well to a framework about why the future of skills and work is unpredictable. First, we have to explore the affordances of the technology. Then you have to think through how organizations will choose to use it, and their rate of adoption. Only then can you explore what new tasks or jobs will be created as a result of this transformation. Reasonable experts can disagree on the timeline and slope of any of these vectors. You’ve touched on all three so far in this conversation.
RYAN: I like that framework a lot.
On How AI is Already Making Call Center Work More Purposeful
ALLISON: Humanist is interested in this idea of purposeful work – where workers have agency, accountability, and are doing things they enjoy with people they enjoy. How can we use AI to make call centers not just more efficient, but purposeful?
RYAN: Interestingly enough, I think we already are.
First, we’re automating the rote work, using AI tools to remove manual, repetitive tasks from the workflow. For instance, after closing a ticket, agents used to “wrap up” by writing notes or creating follow-up tickets for product engineering in a specific format. They don’t have to do that anymore.
Most call center employees would gladly trade the tedious, mechanical parts of the job for more time solving complex problems and connecting with customers. They do this work because they enjoy helping people and exercising empathy. AI lets them focus on that.
Second, we’re deploying AI to accelerate training. Products change constantly, and AI can deliver updates the moment they happen, in an interactive way that helps agents quickly understand what’s new. That’s a huge win.
Third, we’re using AI to better organize people and their time. Customer support tends to swing between chaos and idle time—everyone rushes in during crises, then waits when things quiet down. AI helps align staffing with demand, putting people where they’re needed and reducing wasted effort. It makes the work itself more balanced and sustainable.
ALLISON: So even in a world where it is possible to have fewer of these jobs – although you have some provocations that perhaps we will have more of them – they will on the whole be better jobs when it comes to purpose, meaning, creativity, engagement, etc.
RYAN: Yes, I absolutely believe that.
On The Opportunities and Challenges Facing the Customer Service Workforce
ALLISON: I’m curious about the new tasks and jobs emerging as agents take on lower-level, routine work. What’s changing in call center workers’ day-to-day after adopting your agents?
RYAN: Our AI agents automate two main types of support: “knowledge answers” and “agentic answers.” When someone asks a question answerable by help center content—or by reasoning across multiple articles—that’s a knowledge answer. If they need help executing a task, like accessing another system to pull data, that’s an agentic answer.
As we automate more, quality assurance becomes critical. Knowledge must be accurate, workflows seamless, and compliant. So we’re enlisting call center managers as QA reviewers. Their subject matter expertise is key: they guide tone, style, and resolve policy questions that aren’t well-documented enough for AI to handle alone.
Customer service work is also becoming more process-driven as regulations and products evolve. It’s a cross-functional role now, very product-oriented and judgment-based.
ALLISON: That’s so interesting, especially since process work seems quite different from traditional customer service. There’s a big leap in building and validating processes, including accounting for edge cases.
RYAN: Actually, there’s more skill overlap than people think. Top support agents can intuit answers to new questions because they’re deeply familiar with customer pain points. I once heard someone from Dropbox say these reps understand the product as well as the people who built it. So if there is a skill shift or gap, it lies in being able to make that deep, intuitive knowledge repeatable and scalable, as opposed to simply applying it to one-off cases.
ALLISON: Are there other new tasks or roles emerging from agent adoption?
RYAN: Yes, Product Operations is one. These roles turn customer feedback into product insights. Great support agents are perfect for this; they’re analytical and naturally become the customer’s voice on product teams. We’ll see more of this work and ideally less siloing between product and support teams too.
ALLISON: Are your customers, these large call centers, investing in training or upskilling to help employees adapt as AI automates routine work?
RYAN: People are definitely concerned about the loss of entry-level work, but aren’t sure how to respond.
“Tier One” support, customer service parlance for basic issues, is being automated. But Tier One is where people build core skills: understanding the product, solving problems, handling upset customers. That learning space is shrinking. So how do we prepare people for Tier Two?
Some customers are testing AI tools for training and coaching, but something’s still missing; reps need realistic scenarios and the nuance of handling unexpected edge cases to prepare for more complicated work. AI can help with the expected, but what about curveballs? That’s the unsolved part.
ALLISON: What does automating customer support teach us about the future of work more broadly? What lessons should be amplified to support learners and workers through this disruption?
RYAN: To get real value from AI, the right people need to collaborate on deployment strategy. Too often, leadership or investors focus on cost savings and efficiency, while operations leaders want to dig into specific use cases and their impact on people. But both perspectives need to come together early to set clear, actionable goals for AI strategy. If we’re reducing headcount, let’s be specific and intentional about how we’re doing that. If we’re improving quality, great - let’s define how AI supports that.
One Small Signal
ALLISON: What’s one small signal in the world today you think will become amplified over time?
RYAN: It’s still unclear how AI will make economic sense in call centers. Right now, the focus is “Can we automate this?” But few are asking, “Is it worth it economically?”
Here’s the math: Many first-gen LLM-powered agents cost $1 per chat. U.S.-based reps cost $3–5 per chat, so AI seems like a win. But outsourced support, like in Colombia, can cost just 50 cents per case. For AI to be clearly worthwhile, delivering 3–4x ROI, it would need to cost around 10–15 cents per chat.
And call center workers are actually quite efficient. This isn’t about automating low-effort work—they’re skilled, fast, and effective. So it’s not just can AI do it, but can it do it at the right cost?
ALLISON: What needs to change for AI to hit that 10-cent-per-chat target? What’s standing in the way?
RYAN: When companies like Assembled started building AI support agents, we assumed LLM costs would drop. For a while, they did—but that trend has stalled. GPT-5 wasn’t cheaper. If anything, our per-unit LLM bills are rising.
Yes, the models are getting smarter, but they are also more expensive. And if you switch to smaller, cheaper models, you may not get the performance you need. So the assumption that faster, better AI would also get cheaper hasn’t played out – at least not yet.
Explore more on AI and the future of work:
The Most Durable Jobs in the AI Era - And No One’s Training for Them - Matt Sigelman
AI at Work: What Zapier’s Chief People Officer Thinks Comes Next – Brandon Sammut