How Marshmallow is using AI to build a fairer future for newcomers to the UK.

When Marshmallow launched in 2017, it set out to solve a problem most insurance providers weren’t even looking at: the unfair premiums charged to people who move to the UK. “Migrants get penalized because their overseas driving experience doesn’t count,” explains Paul Elliott, Director of Data Science & AI at Marshmallow. “Even if you’ve been driving for 20 years, you’re treated like a new driver. That leads to unnecessarily high prices, and we knew we could do better.”

Today, Marshmallow serves over 200,000 customers, nearly 80% of whom are new to the UK. The company builds its own technology and pricing models to offer fairer car insurance, tailored specifically to people who have relocated from abroad. It’s a mission grounded in purpose: to back people who step outside the norm and help them access the same financial opportunities as everyone else.

But for Marshmallow, affordability alone isn’t enough. As a highly regulated insurer serving multilingual, globally diverse customers, the company also needs to deliver clear, responsive, and human-centered support. That’s where AI came in.

From rigid flows to real self-service

Before Sierra, Marshmallow relied on traditional support flows that had grown organically but inefficiently over time. “They worked to a degree,” Paul says, “but there was very little true self-service, and customers couldn’t take meaningful actions like updating their policies.”

That smarter solution needed to go beyond templated responses or static flows. Marshmallow wanted an AI agent that could understand nuance, take real actions, and handle sensitive pricing decisions—particularly around cancellations, updates to a policy (like a change of address or vehicle), and renewals. That search led them to Sierra.

Building Marsha together

Marshmallow’s AI agent, affectionately named Marsha, was developed in close partnership with Sierra using its Agent SDK. While Sierra’s team handled much of the implementation, Marshmallow’s engineers and data scientists played an active role in shaping Marsha’s capabilities and behavior.

“It wasn’t just about delivery; it was a real collaboration,” Paul explains. “We didn’t want an agent handed to us as a black box. We wanted to understand it, shape it, and own it going forward.”

That autonomy mattered. With help from Sierra, Marshmallow’s team incorporated their own APIs, and built a full feedback loop for daily iteration and deployment. According to Paul, the two teams met in daily standups and pushed updates every day or two, even across time zones. The turnaround was fast, focused, and impressively coordinated.

Quantifiable impact

Marsha isn’t just helpful; she’s high-performing. Marshmallow uses an Internal Quality Score, or IQS, that takes into account many factors, including regulatory compliance. On this, “Marsha often scores higher than our average human agent,” Paul notes. “That’s huge in a business as tightly regulated as insurance.”

Customer satisfaction scores for AI-handled conversations exceed 82%, with consistently strong qualitative feedback. Customers often describe Marsha as fast, professional, polite, and easy to use. While human agents still score higher in overall CSAT, Marsha has proven herself highly effective in managing the routine but essential interactions that make up the bulk of Marshmallow’s service volume.

“Some conversations really benefit from a human in the loop,” Paul acknowledges. “But others, like renewing a policy or checking for a better price, are perfect for an AI agent. And if that agent can help a customer save money in a way that’s fast and friendly, that’s an amazing outcome.”

Lessons for AI leaders

For Paul, building Marsha wasn’t just about launching a tool—it was about creating a new kind of capability within the company. One of his biggest recommendations for other leaders is to involve internal teams early—especially those closest to day-to-day customer interactions.

At Marshmallow, the customer service and operations teams played a critical role in mapping out high-impact journeys and identifying the scenarios Marsha needed to support. Their deep expertise in how to talk to customers directly shaped the agent’s design and tone, ensuring it reflected the company’s values and struck the right balance of clarity and empathy.

Paul also urges leaders to involve QA teams early, noting, “They’re essential to ensuring the agent communicates with clarity, accuracy, and empathy, especially given the compliance and pricing implications of AI-powered conversations."

Finally, he highlights the importance of autonomy. By encouraging its engineers to gain fluency with the Agent SDK, Marshmallow has positioned itself to iterate quickly, test new journeys, and expand capabilities without always relying on external support.