David Walmsley, Chief Digital & Technology Officer, Pandora
Ron Schmelzer at Dreamforce 2025
Pandora, the world’s largest jewelry brand, is rebuilding the digital sales floor around AI agents that are more than just glorified interactive voice responders. The aim is to recreate the hour-long, story-rich experience that typically happens in Pandora’s retail boutiques and make it work in an interactive online experience. The company’s results from its latest agentic AI move suggest that this approach can work. Customer satisfaction jumped, call deflection doubled, and the team now treats conversational AI as a core sales capability rather than a sidecar experiment.
Build Experiences People Want
David Walmsley, Pandora’s chief digital and technology officer, framed the program with three priorities at the Dreamforce 2025 event. The first priority at Pandora is to design jewelry people actually want. Along with that is the priority to sell that jewelry with intelligence and empathy, and un the company with a tight operating model. AI shows up in all three.
While AI in many forms has been in Pandora’s tech environment for many years, the new addition is an agentic, conversational service agent that replaced a pre-existing chatbot on the Pandora website.
“We deployed it in a market and we saw satisfaction as measured by the net performance score (NPS) jump eight points [over the prior chatbot,” Walmsley said. “The deflection rate doubled. The lift comes not only from more direct and relevant responses, but also from tone, ability to handle a breadth of topics, and fewer dead ends. In other words, It’s just a lot nicer about the subject.”
The company realized momentum from a rapid development process for the conversational agents. Walmsley describes a first-week push in January where his team implemented the technology with a speed-first approach. Out-of-the-box integration with the recently released Salesforce Agentforce product surprised a veteran crew used to long development cycles.
Of course, when it comes to getting accurate results, good quality data matters most. Walmsley points to 270 distinct “definitions of inventory” across Pandora’s stack. That kind of entropy slows everything. The approach the team is taking is to iteratively clean data as the data are put to use. This is an approach Walmsley prefers over waiting for data perfection that never comes.
Tooling choices reflect pragmatism. Pandora leans on SAP, Microsoft, and Salesforce, yet hosts internal agents where people already work, which for Pandora often means Teams. Walmsley asked platform partners to clarify boundaries so buyers can plan around overlapping roadmaps. That clarity is still evolving.
Turning product stories into machine context
Unlike other products that might be easier to purchase online, jewelry shopping is not as easy to select with searches and filters. Pandora sells meaning at charm scale. The agent needs to tease out stories and map them to an existing catalog.
Walmsley offers the contrast in plain terms. A customer says, “My wife loves windsurfing.” The model must translate personal interests into motifs, moods, and relevant offerings. It’s important for AI commerce systems to get this right. Case in point, early on, the agent surfaced a dog charm because it associated windsurfing with leisure activities. In another instance, it suggested an elephant for Thailand, which was a good match, but then confused the flag of Wales with whales.
To address these cue misses, the team now feeds the agent richer design-time materials, not just web copy, to sharpen those semantic frames. That detail matters for commerce. The better the agent grasps themes like “sunset,” “beach,” or “first trip abroad,” the more natural it becomes to assemble a bracelet that feels personal rather than mechanical.
Walmsley’s comparison benchmark is a human associate who asks two or three crisp questions, follows a thread, and builds a small set that fits the recipient’s story. The online agent has to do the same work without losing patience or context.
An Agentic Commerce Roadmap
The first two agents cover service and selling, with more planned on the roadmap. The further plans for composing multiple agents together is also next on the plan, with multi-agent composition to tie together loyalty, promotions, and workflow helpers that trade state among themselves. The real unlock comes when an agent can act. For example, process a refund of a shipment. Amend a promo. Retrieve the wishlist a shopper built at home and pull it up in a store with an email address. Pandora is not there yet, though the intent is clear. Walmsley calls that “giving the agents agency,” a step that moves the system from helper to closer.
For most online commerce operators, omnichannel closure remains the prize. Most journeys begin in digital. For Pandora, only about 22% percent of transactions complete online, yet the online properties are often used for research. Bringing that pre-work into the store in a way that shortens the consult and keeps the magic intact creates both sales and loyalty lift.
Pandora is not alone in pushing conversational commerce into the cart. Walmart just announced a partnership with OpenAI that lets shoppers and Sam’s Club members buy through ChatGPT using Instant Checkout. The retailer intends to let customers plan meals, restock staples, and discover products through a chat flow that completes the transaction. The move expands Walmart’s own suite of generative tools, which already includes AI-powered search, review summaries, and product comparisons.
Amazon rolled out Rufus, a conversational shopping aide that now runs in the Amazon app and on desktop for U.S. customers. The assistant aims to answer open-ended shopping questions, compare options, and compress research time into a single thread.
At Williams-Sonoma, Salesforce’s Agentforce 360 sits behind a portfolio-wide deployment focused on service coverage and efficiency. France’s Carrefour has experimented with Hopla, a ChatGPT-based shopping helper that steers customers to ideas and baskets drawn from preference data. The retailer also rolled internal assistants out at scale, signaling a twin track of shopper-facing and employee-facing AI.
Klarna’s AI assistant initially handled two-thirds of service chats and projected a material profit boost. A year later, leadership shifted back toward more human contact, citing the need for a finer customer touch in certain flows. Klarna’s experience shows both the opportunities and challenges of involving AI in the commerce flow.
Walmsley’s Advice to Tech Leaders: “Just Take the Cellophane Off”
Walmsley offers advice to other tech leaders looking to put AI and agentic commerce into real world action. He suggests tying AI to the strategy you already own. For Pandora that means better product bets, better selling, and a cleaner machine behind the scenes.
He suggests starting with a service agent that can reduce dead ends and measure lift. Then afterwards push into selling conversations that play off the brand’s story. He also suggests that tech teams feed models the same materials that designers use. This provides richer context and improves recommendations that reach beyond literal keywords.
For agents, Walmsley insists that agents act inside guardrails. For data engineering, the approach he offers is to clean the messy data while you ship. Waiting for perfect inventory definitions or canonical IDs will keep the project theoretical forever. And he also tells other leaders to bring the top team along so decisions about partners, platforms, and privacy do not stall in committee.
Walmsley, whose experience with interactive digital media goes back to the CD-ROM days, sums his decades-long lessons learned with a simple phrase, “just take the cellophane off”. That means to just get started. Get your hands dirty and iterate the experience in the real world with real customers and real experiences. Don’t just keep the technology wrapped up and on the shelf.
“Proof by doing,” as Walmsley puts it.