When Pandora rolled out its first AI-powered service agent earlier this year, the goal wasn’t to showcase cutting-edge technology. It was to survive the peak holiday shopping season without sacrificing the customer experience or margins.

The agent, dubbed Clara, launched as a chatbot on the brand’s website in mid-February and now resolves around 60% of customer service inquiries without escalation to a human agent. Pandora’s previous chatbot resolved around 40%. More unexpectedly, a Pandora spokesperson said Clara has contributed to a 10% increase in the company’s net promoter score, measuring how likely customers are to recommend a brand to others.

“You don’t want to have to staff call centers [to be ready for] periods of peak sales volume,” David Walmsley, Pandora’s CTO, told Glossy. “It’s easier to train an AI service agent than it is to train 600 new service agents for peak; the efficiency we’re getting with the service agent is really paying dividends on Black Friday and this whole period.”

Those results matter for Pandora, which generates roughly 40% of its annual revenue in the final three months of the year. On the company’s third-quarter earnings call, executives repeatedly pointed to efficiency and scalability with AI as key offsets to macro pressure, as Pandora held gross margins near 80% and reaffirmed full-year EBIT guidance of around 24%.

Clara is one of two AI “agents” Pandora has been developing as part of a wider push into agentic AI. The second, Gemma, is a sales agent, and a more ambitious experiment.

Launched quietly in Australia in June and currently visible to about 20% of the region’s site traffic, Gemma is designed to replicate Pandora’s in-store selling experience online. “In five or ten years, more transactions will happen online than today,” CEO Alexander Lacik said on the Q3 earnings call. “And therefore, we need to be offering a more interesting experience than we do today.”

That means guiding customers through emotionally driven conversations about gifting, relationships and memory. “We don’t sell pots and pans,” Walmsley said. “Half of what you buy is the product, and half is the experience in the store.” And selling jewelry requires nuance.

In a physical store, a Pandora associate may ask who a gift is for, why the occasion matters and what memories the piece should represent. Online, that richness has typically collapsed into dropdown menus and search filters. Gemma is Pandora’s attempt to close that gap using conversational AI.

Currently, there’s a “specificity problem” online, Walmsley said. In-store, a customer may say they’re buying for someone who loves windsurfing, for example. Pandora doesn’t sell a windsurfing charm, but a skilled associate would keep asking questions until a meaningful connection emerges, such as a honeymoon trip or a shared place memory, which an existing charm could represent.

“That’s what we’re trying to teach the agentic model to do,” Walmsley said. “And that’s really quite difficult.” Internally, Pandora refers to significant reliance on Gemma as a “moonshot,” considering the challenge of scaling emotional selling.

As the vision goes: If a shopper opts in, the agent opens a conversational chat, asking who the gift is for, what the occasion is and what the piece should represent. As the customer shares more context — such as a birthday or a shared travel memory — Gemma narrows its recommendations based on its learnings, suggesting relevant jewelry and explaining how each option connects to the story, mirroring the guidance of an in-store associate.

Pandora managed to hold gross margins near 80% in the third quarter and reaffirmed full-year EBIT guidance of around 24%. On the earnings call, executives repeatedly pointed to operational discipline and productivity gains as key offsets to macroeconomic challenges.

For its part, Clara offers Pandora more visibility than traditional call centers. Every interaction can be analyzed for resolution quality, sentiment and recurring pain points, to inform forecasting, planning and customer experience improvements.

That data advantage ties into Pandora’s longer-term ambition of building durable customer “memory” across digital channels. While many shoppers interact anonymously, AI-driven conversations capture intent and sentiment in ways clickstream data alone cannot.

“We sell memories,” Walmsley said. “So collecting memories feels like a pretty good thing we can do, from a tech point of view, to help propel the business forward.”

The timing is deliberate. Around 23% of Pandora’s sales now come from digital channels, a figure management expects to rise.

On the earnings call, CEO Alexander Lacik acknowledged that AI has been embedded across the business for years — from media buying and customer segmentation to design development — but positioned conversational AI as the next frontier.

“In five or 10 years, more transactions will happen online than today,” Lacik said. “We need to offer a more interesting experience than we do today.”

Gemma is Pandora’s answer to that challenge. If it works, the agent could allow the brand to scale one of its strongest competitive advantages with its human-style selling without relying solely on physical retail expansion or linear increases in staffing.

For now, Pandora is careful not to overpromise. Gemma’s impact on conversion is still early, and executives are clear that correlation does not yet equal causation. But the direction is set: Clara stabilizes the business during peak periods. Gemma, if it delivers, could reshape how Pandora sells year-round.

As Walmsley put it, “Selling more jewelry to happier customers is the point.”