Who Owns the Customer in the Age of AI Agents?

On September 8, 2026, Meta launched Muse, a personal AI agent designed to act on a user’s behalf: filling out forms, negotiating, booking travel, and making purchases. Twelve days later, Amazon blocked Muse from shopping on its site, arguing that third-party agents should identify themselves and obtain a retailer’s permission before transacting on behalf of its customers. The dispute might look like a technical disagreement over how AI agents access a website. But it marks an important shift. For a little over three decades, Amazon has largely determined how customers search, compare, and purchase products on its interface. Muse introduces a different possibility: The customer can show up with an agent that already understands what they want, can evaluate options on their behalf, and can potentially transact without handing Amazon control of the decision process. You might dismiss this as simply another shift in online search, with consumers now discovering products through an AI interface rather than a traditional search engine. But something more consequential is happening: AI is moving from helping consumers find products to representing them in the market and is increasingly making decisions and acting on their behalf. Companies are already responding to that shift. Amazon is restricting outside agents; Shopify has opened its merchant catalog and checkout infrastructure to Meta’s AI channel; Walmart is building its own shopping agent while also integrating with external agents such as ChatGPT and Gemini. This shift highlights a new strategic problem that online stores and marketplaces are starting to face in the age of AI: You can own the transaction, the supply, and even the customer’s engagement with your service while nonetheless losing control over the decisions that bring the customer there. Increasingly, that control may belong to the agent that represents the customer. Ten years ago, writing in HBR, two co-authors and I explained how platforms and marketplaces build power through network effects, data, and engagement—and how doing so could bind customer engagement and strategic control together. In the years that followed, building on that idea, I’ve had the opportunity to advise leaders at many of the world’s leading platform companies on how to design their platforms for durable success. The principles we articulated in that article still matter, but thanks to AI agents and their integration capabilities, that relationship between customer engagement and strategic control is beginning to come apart. You might not think so. Traffic from AI sources to U.S. retail sites grew 393% year over year in the first quarter of 2026, according to Adobe. And these visitors were more engaged once they arrived: They stayed 32% longer, viewed 10% more pages, and were 27% less likely to leave immediately. But a retailer or marketplace can continue to attract highly engaged customers even as more of the decisions that determine what those customers consider, compare, and ultimately choose are made elsewhere. As more of the decision shifts to the agent, the retailer risks losing influence over which products enter the customer’s consideration set and which trade-offs shape the choice even while headline metrics such as traffic and sales remain healthy. The retailer also risks losing control over higher-value parts of the relationship: expanding the basket through cross-sell, learning from what the customer rejected, and bringing the customer back for the next purchase. Over time, the retailer can be relegated to an invisible yet operationally intensive position, fulfilling the transaction while the agent increasingly shapes the decision and owns the relationship. So when it comes to controlling the customer relationship, how are the rules of engagement changing? In this article, I’ll discuss six important ways. 1. Owning discovery no longer guarantees that you own the customer. Conversational search has become a major focus of AI investment among online marketplaces, and with good reason. Users can describe what they want in natural language instead of navigating filters and menus, which improves product discovery. But better discovery is not the same as greater control of the customer relationship. Increasingly, the consumer’s own AI agent can perform that discovery across multiple marketplaces at once. Once the same agent can search across competing marketplaces, each marketplace becomes just a source the agent can query rather than the place where the consumer begins the decision. In October 2025, Zillow became the first—and at the time only—real estate app available inside ChatGPT. Five months later, though, this didn’t seem like much of an advantage anymore. Realtor.com had launched its own ChatGPT app, putting a direct competitor into the same conversational interface. In travel, Booking.com and Expedia launched inside ChatGPT on the same day, with Tripadvisor subsequently joining them. In retail, Walmart announced an integration with Gemini using Google’s Universal Commerce Protocol; Target subsequently used the same open protocol to bring its products into Gemini and Google AI Mode. These AI initiatives, driving higher user engagement, would have offered much cause for celebration in the platform era. In the agent era, however, what matters is whether the marketplace can help the consumer’s agent make a better decision than it could make elsewhere. Doing that requires learning from what happened after discovery: which conversations became offers, which offers became contracts, which contracts became completed sales, and why. 2. Your network effects are no longer the guaranteed path to owning the customer relationship. In our 2016 article, my co-authors and I described network effects as the defining source of power in the platform era. Network effects still matter. But AI can reduce how much of the network a competitor needs to recreate. An entrant can instead start with privileged access to one side of a market and use AI to organize the other. The incumbent’s network remains valuable, but the entrant no longer needs to own an equivalent network to compete. The freelancing marketplace Upwork, for example, built powerful network effects by aggregating freelancers and clients and matching them to one another. Its ability to match the two sides improved as its network grew. Mercor approached the same matching problem without recreating comparable scale on both sides. It used AI-based assessments to extract data and create richer candidate profiles, then matched those profiles against employer requirements. A conventional marketplace entrant would normally need to build substantial participation on both sides before it could perform good matches, but Mercor could generate part of that intelligence directly through AI-based assessment. In effect, it substituted depth of information for breadth of network: By extracting more useful intelligence from one side of the market, it could improve matching for the other without first waiting to build Upwork-scale network effects. The AI-powered expense-management firm Ramp is taking a similar approach in corporate procurement. Ramp’s expense-management and accounts-payable tools gives it visibility into what companies are buying, from whom, and at what price. It uses that purchase data to then help those same companies evaluate vendors, negotiate contracts, and make purchasing decisions, all without having to assemble a large supplier network from scratch In the platform era, the two sides of a marketplace protected each other: taking buyers required access to sellers, and attracting sellers required buyers. But in both of the above cases, AI allowed a company with a strong position on one side of the market to navigate the other without recreating the incumbent marketplace in its entirety. 3. Your customer data may no longer give you the deepest customer understanding in your category. Marketplaces that understand customer intent deliver better recommendations, thereby improving discovery. But agentic commerce raises the bar: Instead of recommending what a customer might want, it represents the customer in the decision. That distinction comes down to scope. A marketplace may understand how someone searches and behaves within the categories it serves. A personal agent can potentially know the person across categories and contexts. Muse can recall details across conversations, negotiate, and make purchases on consumers’ behalf, for example. Instinct, another personal AI assistant, connects across a user’s email and devices, remembers the context of their everyday life, and can follow up on tasks, arrange rides, book services, contact businesses, and make calls on their behalf. Access to broader context and the ability to act on the customer’s behalf are two important sources of advantage. However, in order to improve over time, the agent must also learn how users make decisions—what they choose and reject, and why. Tripadvisor illustrates what happens when that feedback loop is missing. The company built more than a billion travel reviews—a formidable strategic asset. For more than two decades, that proprietary corpus helped Tripadvisor understand what travelers valued, and it attracted search traffic to the site. AI now weakens both advantages. Its interfaces can synthesize review sentiment directly for the traveler, reducing the need to visit Tripadvisor itself. More importantly, Tripadvisor.com often loses visibility into the downstream transaction. By contrast, Tripadvisor’s transactional businesses, including Viator and TheFork, grew at double-digit rates in 2025 and now account for nearly 60% of company revenue. Those businesses sit closer to the eventual transaction and observe more of the outcomes from which an agent can learn. 4. Locking in the supply is no longer sufficient to control the demand relationship. Many marketplaces create switching costs by embedding themselves in participants’ daily workflow. Airbnb hosts use the platform to manage reservations, pricing, calendars, messaging, and payments. Amazon sellers can rely on Amazon for inventory management, fulfillment, advertising, pricing, and payments. Upwork freelancers manage communication, milestones, timesheets, billing, and payment inside the platform. These workflows create real supplier-side lock-in. In the platform era, controlling supply in this manner helped protect demand: A rival marketplace had to persuade suppliers to move or multi-home before it could offer customers a comparable choice. Agents can weaken this control on both sides. Consider the Airbnb example again. As already discussed, a traveler’s AI agent can search across Airbnb, Booking.com, Vrbo, hotels, and direct-booking sites, comparing options without requiring the traveler to begin inside any one marketplace. Additionally, on the supply side, a property-management AI agent can help a host coordinate calendars, pricing, guest communication, cleaning, and payments across Airbnb, Booking.com, Vrbo, and direct bookings. This results in a two-sided weakening of lock-in: The customer’s agent makes demand portable across marketplaces, while the supplier’s agent makes supply portable across them. Airbnb can remain an important transaction venue without necessarily owning either side’s relationship. 5. Launching your own AI agent may not help you own the customer either. Walmart’s own experience suggests that an AI agent can deepen customer engagement: By early 2026, engagement with its Sparky shopping assistant was rising, and customers who purchased through Sparky had baskets about 35% larger than other customers. But stronger engagement with the agent does not by itself determine who controls the customer relationship. What matters is where the transaction is executed and who captures the learning that follows. In October 2025, Walmart started allowing customers to purchase Walmart products directly inside ChatGPT using OpenAI’s checkout. In January 2026, Walmart also adopted Google’s purchasing protocol, which meant transactions could also complete inside Gemini. Two months later, however, after discovering that in-chat purchases converted at only about one-third the rate of purchases on its own site, Walmart moved away from the generic in-chat checkout model and toward bringing its own shopping experience into ChatGPT—a move that seems designed to allow the company to retain control over the customer account, cart, loyalty, and transaction experience. Walmart’s AI assistant, Sparky, remained essentially unchanged across all three implementations. What changed was where the transaction was executed. By reclaiming checkout, Walmart reclaimed visibility into actual customer purchases, as well as insights on which options they rejected. 6: Owning customer data may no longer help you own the customer relationship. In the platform era, data generated from user activity helped personalize the consumer experience. Most such data, though, was optimized for understanding browsing behavior and improving recommendations to users, not for enabling agentic decision-making. A marketplace may know what a consumer searched for, clicked, and considered. But a personal agent connected across applications may see what happened beyond those initial considerations. Google’s Personal Intelligence, for example, can draw on Gmail, Photos, and other Google services, allowing its AI to use information such as purchase receipts, hotel confirmations, and past travel experiences when helping users make future decisions. That doesn’t mean marketplace data is unimportant. Its value depends on how much of the customer’s decision and outcome the marketplace can directly observe and the frequency with which those observations can be made. Grocery gives Instacart an unusually strong position because purchases are frequent, repetitive, and completed within the platform. Over time, it can learn not only what a household searches for but also what it repeatedly buys, substitutes, rejects, and prefers across brands, budgets, and dietary needs. That closed high-frequency feedback loop gives its AI useful signals for the next decision in a way that marketplaces where the final transaction happens elsewhere cannot easily replicate. Customer data that helps personalize the experience is no longer enough. The new advantage lies in data that lets you observe decisions through to their outcomes and then use that feedback to help AI agents make better decisions. . . . All six of these changing rules point to the same shift. Even if marketplaces still have strong engagement, network effects, supplier lock-in, proprietary data, and even their own AI agents, they no longer necessarily control the customer relationship. As consumers delegate more of discovery, comparison, negotiation, and purchasing to AI agents, those traditional advantages become more portable. The stronger position belongs to the system that can connect a customer’s intent to the decision, observe the outcome, and use that feedback to make the next decision better. In the agent era, owning the customer will increasingly mean earning the right to represent them.

Who Owns the Customer in the Age of AI Agents?

Summary.

Partner Center