When Rocket Money’s user base exploded and support requests surged, the obvious answer was more headcount. Director of Operations Michelle McGowan chose a harder path, and what she built instead is a blueprint for how AI and humans can divide labour in ways that make both better.
It was, in the language of operations, a solved problem that hadn’t been solved. And Michelle McGowan, Rocket Money’s Director of Operations, knew exactly why. “We realized that no matter how well we tried to design those flows,” she says, “we were never going to be able to predict every situation.” The routing logic was too rigid for the messiness of real customer needs. People don’t contact support in tidy, predictable categories. They come in mid-crisis, mid-confusion, often mid-argument with their own bank statement.
Rocket Money is a personal finance app with a specific and sensitive purpose: helping people see where their money is going and cancel the subscriptions they no longer want. It’s a product that sits at the intersection of financial data, personal habits, and occasional embarrassment. The customers who contact support are often stressed. The questions they ask, about disputed charges, billing cycles, account access, are not trivial. Getting the answer wrong, or routing them to the wrong person, erodes exactly the kind of trust that a fintech company lives or dies by.
By 2024, with millions of users and a support volume that had grown alongside them, the old model had reached its limit. The question McGowan faced was the same one confronting support leaders across the industry: not whether to bring in AI, but how to do it without breaking something fragile.
Ten Percent
McGowan’s team started with a number that is easy to underestimate: 10 percent. When they introduced Fin, Intercom’s AI Agent, they routed just one in ten conversations through it. Not because they lacked ambition, but because they understood something about trust that a lot of AI rollouts miss. In financial services, customer trust is not a soft metric. It is the product. An AI that gives a wrong answer about a billing dispute does not just create a bad support experience. It creates a customer who is now less certain that Rocket Money knows what it’s doing with their money.
“Every step of the rollout was deliberate,” McGowan says. “We tested, we measured, and we made sure the experience was up to our standard before we expanded.” From that initial 10 percent, the team expanded Fin into specific workflows where the scope was clearly defined and the risk of error was manageable: billing management, app troubleshooting, account access requests. These were the categories where a good answer was reliably achievable and where the cost of a mistake, while still real, was contained.
The phasing was not just a risk management strategy. It was also a learning strategy. Each expansion gave the team data on where Fin performed well and where it needed refinement. The knowledge base that powered Fin’s answers got better with every cycle. Edge cases got logged, reviewed, and fed back into the system. What started as a controlled experiment gradually became something the team understood deeply enough to trust.
What the Numbers Don’t Capture
Today, Fin is involved in more than half of all conversations at Rocket Money, resolving 68 percent of them, tens of thousands of interactions each month. Manual triage has been eliminated. Human CSAT scores have risen by six points, a counterintuitive result that reflects something important about how AI changes the nature of human work rather than simply replacing it. And the total ROI of the implementation has reached nearly one million dollars.
But McGowan is more interested in a different metric. “We wanted a system that could support customers reliably, at scale, and still feel personal,” she says, “because we recognize our customers’ financial journeys are deeply personal.” The efficiency gains are real, but they were never the point. The point was to build something that customers could trust, and that meant getting the human-AI balance right.
That balance looks different from what most people imagine when they think about AI replacing support jobs. At Rocket Money, the agents who used to spend their days manually rerouting tickets are now doing something more valuable: they’re managing Fin. Refining workflows. Identifying edge cases. Improving how the AI handles exceptions. The work has moved from repetitive execution to something closer to quality engineering, and the team is actively hiring a full-time Knowledge Manager whose job will be to ensure the underlying knowledge that powers Fin stays accurate and current.
A New Division of Labour
The knowledge management role is worth pausing on, because it represents something that rarely makes it into conversations about AI and employment: the idea that AI doesn’t just redistribute existing work, it creates new categories of work that didn’t exist before.
At Rocket Money, human agents still handle every conversation that requires oversight or escalation. They step in when a situation is genuinely complex, when a customer is distressed, when the answer isn’t in the knowledge base. What’s changed is the quality of those interactions. Because agents are no longer burning hours on routine triage, they arrive at the hard conversations with more time, more focus, and more capacity to actually help. The six-point rise in human CSAT is, in part, a measure of that freed-up attention.
“This is what a modern support team looks like,” McGowan says. “It’s not about removing humans. It’s about redesigning the work so humans are focused where they add the most value.” That framing , redesigning the work rather than automating it, is the distinction that separates the AI rollouts that hold from the ones that create backlash. Rocket Money’s team didn’t feel replaced. They felt redeployed.
The Capability, Not the Project
There is a phrase McGowan uses that captures the most important thing Rocket Money got right: “AI is not a project. It’s a capability.” Most companies treat AI adoption as something to be completed, a rollout with a launch date and a done state. Rocket Money’s approach was built around the opposite assumption: that what they were building would need to keep getting better, indefinitely, and that the systems, roles, and practices to support that improvement were as important as the technology itself.
That is why the Knowledge Manager hire matters as much as the Fin deployment. The AI is only as good as the information it draws on, and that information needs someone responsible for keeping it accurate, complete, and current. Building that role into the team structure signals something about how seriously the organisation is taking the long game.
Twelve months into the transformation, the results at Rocket Money are strong enough to be compelling and honest enough to be instructive. The near-million-dollar ROI is real. So is the six-point CSAT improvement. So is the fact that the rollout was slow, careful, and built on a foundation of trust, with customers and with the team, that made every subsequent step possible.
The hardest part of deploying AI isn’t the technology. It’s having the discipline to go slowly enough to get it right.
Note: This is a spec writing sample based on publicly available information about Intercom and Rocket Money.
Comments
Post a Comment