How Coinbase Built a Two-Tier Customer Support Model With Human Specialists and AI Scale

Customer support is reorganizing into a two-layer system where automation and human expertise work in tandem. As AI absorbs the steady stream of routine requests, the issues reaching human teams are the ones with real stakes: confusion, urgency, fraud concerns, or moments where trust is on the line. That shift is changing the job itself, pushing traditional agents into more specialized roles that require product fluency, judgment, and the ability to guide customers through complex, high-context situations that automation alone cannot resolve.

Wes Griffith, Senior Director of Global Consumer Support Experience at Coinbase and a former Director at Amazon, is already managing that transition in practice. Drawing on years of experience scaling large support organizations and automation systems, he sees the leadership challenge shifting from simply deploying AI to designing a cohesive model where automation and human specialists work together to protect customer trust.

"AI in CX is not humans versus AI. It’s AI handling scale and humans handling significance. But that means it’s incumbent on the company to know the difference," says Griffith. The philosophy has clear structural implications for how support teams are organized. At Coinbase, where automation now handles close to 80 percent of customer contacts, the company formally reclassified its frontline staff from "agents" to "specialists," reflecting a deeper change in the role itself. Instead of focusing on high-volume ticket resolution, specialists increasingly step in when situations involve ambiguity, urgency, or emotional stakes. As automation takes on routine work, human teams are freed to focus on the moments where judgment, explanation, and trust matter most.

• What's the why?: But the new title reflects more than a branding change. Instead of closing tickets, specialists are expected to guide customers through complex or sensitive situations, often by explaining the reasoning behind the systems designed to protect them. “We’ve really focused on sharing the ‘why’ behind what the customer is experiencing,” Griffith explains. In doing so, specialists also act as a feedback channel for the company, capturing signals about shifting customer behavior or friction points that can inform product decisions and refine policies.

• Who, when, how: Griffith notes that the handoff from AI to human is one of the most trust-sensitive moments in the system, and Coinbase treats it as a deliberate operational decision rather than a default escalation. “Determining when to execute the handoff is typically some combination of defining the issue, and then gauging the customer segment,” he explains. In practice, that means some customers or situations may bypass automation entirely, particularly when vulnerability or the need for reassurance is detected.

• Escaping the labyrinth: To prevent customers from getting trapped in frustrating automated loops, Griffith’s team also enforces a system-wide safeguard: every interaction has a finite number of turns. If progress stalls, the system escalates the case to a human specialist who can reset the interaction and rebuild trust. The threshold varies depending on the situation. Routine tasks like retrieving tax documents may remain in automated flows longer, while higher-stakes scenarios escalate immediately. “If a customer is worried about fraud or thinks they’ve been scammed, that’s one turn with automation and then straight to a human,” Griffith says. “Those are moments where trust matters too much to keep them in a system.”

Perfecting the handoff, Griffith says, depends on ensuring that context moves seamlessly from automation to the human specialist. “We’re pushing to get a tighter coupling between the work the AI agents are doing and the context they capture,” he explains. Specialists are already supported by internal AI tools that surface key signals from prior interactions and account activity, helping them diagnose issues faster without forcing customers to repeat themselves.

• Learn from churn: His advice to other leaders is pointed: Identify the interactions that are direct inputs to customer churn. Then, reverse engineer your support experience to handle those moments with the right resource, every time. “If upholding a policy increases churn risk, a human can meet that moment with empathy and reduce that risk, even if the outcome for the customer doesn't change. The company is going to be better for it in the long term.”

The two-tier model raises the bar on both sides of the system. AI will need stronger context retention, better emotional detection, and enough policy flexibility to handle nuanced situations within guardrails. Human specialists, meanwhile, will need deeper product fluency, stronger risk literacy, and the judgment required to navigate complex, high-stakes interactions. As automation continues to absorb routine support work, the remaining human role becomes more strategic, focused less on resolving tickets and more on protecting trust when it matters most.