You're watching support volume double every quarter while your CFO pushes for AI that "deflects 80% of tickets." But in fintech or healthtech, one wrong automated answer could trigger regulatory action, fund loss, or compliance violations.

Bob van Winden lived this tension at Stripe. Over nine years leading global support operations through 100x growth, he discovered that AI works completely differently in regulated, high-stakes environments. Now as COO at Bridge (acquired by Stripe in 2024), he's building AI-first stablecoin operations, applying those hard-won lessons from day one.

Why Deflection-First AI Breaks Under Regulatory Complexity

When van Winden joined Stripe from Google in 2015, he brought a clear strategy: reduce conversations, automate deflection, scale efficiently. His first presentation laid out exactly how to avoid hiring a thousand support people.

Patrick Collison's response changed everything: "It's a really good thing that we're talking to our users. Why would you want to stop?"

That insight proved critical as Stripe scaled from 200 to 8,000+ employees. The challenge was multiplicative complexity:

• Product complexity. Stripe evolved from a single payments API to Atlas, Billing, Radar, stablecoins, and multiple acquisitions. Each added integration points and compliance requirements that standard AI training never captures.

• Technical complexity. "The financial stack depends on global banking partners. You don't even know if the issue is on Stripe or a third party," van Winden explains. Support needed to reason across logs, APIs, partner systems, and international regulations.

• Emotional stakes. Van Winden recalls a restaurant chain escalation: "If this money doesn't land by 3 p.m., we're not making payroll for 120 people."

This is why deflection-first AI fails in regulated industries. The inquiries reaching humans aren't the ones that should have been deflected, they required human judgment from the start.

The One Metric Change That Transforms AI Implementation

Stripe discovered that First Contact Resolution (FCR) was an unhelpful metric in isolation. Agents wrote exhaustive "conditional essays" when answering customers, trying to cover every scenario in one response. Customers got overwhelmed. Satisfaction dropped.

Van Winden's team developed First Agent Resolution (FAR): the percentage of issues resolved by one person, regardless of message count.

"If the second or third touch is with the same person, customers don't really mind," van Winden explains. "But if they're passed around to different people, they get extremely frustrated."

FAR incentivizes ownership over speed. It encourages routing complex issues to the right expert immediately rather than bouncing customers through automated deflection first.

Why Stripe Reversed Course on Outsourcing

Initially, Stripe outsourced support to avoid having 50-70% of company headcount in support roles. However, some problems emerged. "Attrition is almost always higher," van Winden explains, and compared to internal staff, "the level of care and skin in the game is just really hard to replicate."

Stripe made a strategic shift: they asked what they'd do with a blank sheet of paper. The answer was clear - employees should work directly for Stripe, with equity and skin in the game.

They built in-house support centers in different locations, hired directly, and let attrition gradually shift the balance. "It's improved quality and made it much easier for all different parts of Stripe to build into that operational muscle," van Winden says.

Building AI Into Operations From Day One

At Bridge, van Winden finally had the opportunity most CX leaders don't: designing AI into the foundation rather than retrofitting it. Van Winden observed, "early automation just couldn't deliver. It's really only in the last two or three years that AI is adding real business value."

The Bridge approach:

• AI handles the repetitive layer. Password resets, transaction status, account navigation. Anything with deterministic patterns gets automated with clear escalation paths.

• Humans own the relationship layer. Complex investigations, multi-party technical issues, compliance-sensitive escalations flow immediately to specialists with full context.

What Customer Obsession Actually Looks Like

Even at 8,000+ employees, Stripe's leadership regularly reviewed specific customer cases. Patrick Collison kept his email public and personally forwarded escalations.

Van Winden learned to prioritize anecdotes: "Going into individual cases is where real value comes from. You start to deeply understand a space in ways the data won't show you."

This shaped how Stripe approached AI: not as a tool to avoid customer conversations, but to scale the conversations that matter most.

For CX leaders in fintech, healthtech, and regulated industries, that's the blueprint: use AI to handle what scales easily, preserve humans for what drives retention, and measure what actually predicts customer trust.