Goldman Sachs analyzes the era of agency-based AI consumption: Who will be the biggest winner in the e-commerce sector?

In the AI era, the shopping chain is being rewritten.

According to a recent research report from Goldman Sachs Global Investment Research, AI agents will reshape the entry point for consumer intent—evolving from the past "users actively searching, comparing prices, and placing orders" to "AI agents understanding needs, filtering products, and triggering payments." The competition in the value chain has also shifted from "who has page traffic" to "who can capture and execute purchase intent."

The research report condenses its core judgment into one sentence: "Value should increasingly follow consumer intent." The implication is straightforward: if commercial intent shifts from search boxes, retail websites, and apps to AI-powered agent interfaces, advertising budgets, payment traffic, identity verification, anti-fraud measures, and merchant distribution fees will all follow suit. This won't happen overnight; the timeframe is closer to the evolutionary cycle of e-commerce itself, potentially lasting three to five years or more.

This migration has already begun to impact investment logic. Goldman Sachs' framework posits that winners are determined by five things: distribution capabilities, consumer trust, merchant engagement, transaction infrastructure, and the ability to capture business intent. The affected industries span e-commerce, digital advertising, payments, identity verification, fraud prevention, and cybersecurity, but the debate consistently centers on a core question: who owns the consumers?

The first items to be taken over were paper towels and cleaning products.

The implementation order of proxy shopping depends on two things: how complex the transaction is and how serious the consequences of buying the wrong thing are.

Low-risk, low-complexity product categories will be the first to be taken over by agents. For items like paper products, cleaning supplies, packaged foods, and groceries, consumer brand loyalty is relatively low, allowing agents to drive repeat purchases based on historical buying records, price, inventory, and delivery speed. When out of stock, agents can also more easily replace SKUs or brands. Private label brands have an advantage in these scenarios because "value" and "availability" will be the priorities for agents.

Complex scenarios such as ride-hailing, restaurant reservations, and delivery, but with low costs for error, also meet the criteria for early adoption. Consumers may still have preferences, but if the results don't perfectly match, the consequences are relatively manageable.

The most difficult to automate are highly personalized consumer goods with high error costs . Shoes, jewelry, some clothing and luxury goods are heavily influenced by personal preferences, resulting in high return risks. Travel, hotel and airline scenarios can bring significant value to merchants, but if a mistake is made, consumers suffer even greater losses, thus requiring more mature authorization, confirmation and liability rules.

Goldman Sachs estimates that approximately $2.6 trillion in US spending falls into high-probability categories, i.e., low-risk, repeat purchase types. Another key figure is that for every 2% increase in offline bank card spending through agency services, e-commerce growth could contribute approximately 1 percentage point. This explains why agency-based commerce is framed within the context of "continued e-commerce penetration," rather than simply as an AI application.

The advertisements won't disappear; they'll just be collected through a different channel.

The core of digital advertising has never been placement, but rather commercial intent. In the desktop internet era, advertising followed web page traffic; in the mobile internet era, advertising followed user relationships within apps; in the AI agent era, advertising may follow "default recommendations."

In the short term, AI will primarily improve advertising efficiency: reducing creative costs, optimizing ad placement, and increasing ROI. The bigger variables lie in the medium to long term: as more commercial intent occurs within AI interfaces, the value of traditional search pages, retail pages, and feed ads will be revalued.

The key issue isn't whether advertising budgets disappear, but rather who becomes the new recommendation gateway. AI-native sponsored recommendations, ranking, and default selections could all become new forms of advertising monetization. After search is disrupted, it may not necessarily move towards a free, open market; new players may emerge who charge for it.

The positions of these three types of companies are the clearest. Meta combines personal AI with an AI-native advertising interface, providing both a consumer-side personal agent entry point and an advertising monetization path. Alphabet still has a discovery entry point, with its AI interface extending discovery beyond traditional search, and agents and payment agreements simultaneously reaching the infrastructure for "intent-to-transaction." Amazon is closest to the transaction; retail media is inherently built on purchase intent; price, inventory, and delivery reliability are precisely the metrics that both consumers and agents optimize.

In the past, advertising bought "exposure," then it bought "conversion," and the next step may be buying "being selected by an agency."

Major retailers: While reaping the benefits, they also need to guard against having their entry points intercepted.

Agency-based business has two effects on large retailers.

On the positive side, AI agents naturally optimize pricing, inventory, delivery speed, and transaction reliability, which will favor large-scale platforms. Amazon and Walmart have advantages in fulfillment networks, inventory depth, and pricing power. Small merchants may also benefit, as agents can lower the discovery barrier and help users find more suitable long-tail supply; however, as ranking functions increasingly value low prices, speed, and stability, scale will remain a key advantage.

The downside is equally real: large retailers are reluctant to entrust discovery portals and customer relationships to horizontal AI agents. Once users no longer directly access a retailer's app or website, first-party data, membership systems, cross-selling, and retail media revenue will all be weakened. Agents bring in orders, but they may also steal customer mindshare.

Walmart's Sparky provides a prime example. Launched in June 2025, this AI assistant offers features including customer service, creating shopping carts based on frequently purchased items, summarizing reviews and providing recommendations, and helping users plan and compare products. Since its launch, weekly active customers have doubled year-over-year and increased by 60% month-over-month, with average order value 40% higher. Users have already adopted it for complex multi-item shopping scenarios such as party planning.

Walmart's strategic boundaries are also quite clear: collaborating with AI players to share "perishable data" such as real-time pricing and inventory to gain higher-level discoveries and new customers; while keeping long-term customer data in its own hands. For the advertising business, Sparky is changing more than just ad placements—for high-frequency users, ad load is lower but value is higher, and the logic of search queries triggering ads is migrating to the logic of contextual signals triggering "solutions".

Brand differentiation: Standardized products fear being replaced; high-participation categories win by leveraging their attributes.

Consumer goods companies will differentiate themselves based on product category when facing AI agents.

In categories with low engagement and product homogeneity, the risks are more direct. For paper products, cleaning supplies, shelf-sold food, and basic apparel, AI agents can manage the entire purchase cycle, automatically switching brands when out of stock. Private label brands with low prices and high availability are more likely to be selected. Brands that previously relied on shelf placement, search page exposure, and habitual repeat purchases will see their advantages diminished. CLX and KHC were listed as higher-risk by Goldman Sachs due to their private label brand exposure.

High-engagement categories are relatively more resilient. In high-end beauty and skincare scenarios, agents can make recommendations, but consumers still make the final decision. Higher average order value, efficacy claims, brand preference, DTC relationships, and member loyalty will all retain the returns on brand building. EL is in a better position due to factors including its DTC experience partnership with Shopify and its participation in AI-native brand discovery through Profound.

Proxy recommendations also bring a new constraint: the number of options is far fewer than on the first page of search results. In the future, brands may no longer be competing for "first-page rankings," but rather for "whether they can be understood and recommended by AI." Product attributes must be machine-readable. Goldman Sachs noted that MDLZ's Oreos only appeared in 10% of AI cookie recommendations; PG is pushing forward with optimizations for generative engines (GEO). Marketing budgets will also shift accordingly—the more consumers participate in the final decision-making process, the more valuable brand advertising will be; the return on investment for substitutable standardized products will be more easily swallowed up by proxy ranking rules.

Apparel Soft Thread: Brand Story is More Important in the Era of Agency

The logic is similar for apparel, footwear, bags, and yarn products, but with a greater emphasis on brand differentiation.

Basic, price-driven retail faces greater controversy. If AI agents only compare price, size, delivery, and return convenience, weaker brands are easily replaced. Categories driven by innovation, with significant functional differences, and requiring explanation and recommendation are better suited to succeed in agent-based discovery.

Winning brands need to enable AI to "understand" why they are different. Product innovation, differentiated features, brand narrative, and AI engine optimization will all become more important. Offline experiences and marketing campaigns will not lose their value; on the contrary, they may serve to "drive consumer adoption and referrals."

Payment networks: New shopping methods are still running on old tracks

Agency-based commerce does not bypass payment infrastructure. Agreements between Google, Stripe, Visa, and Mastercard are integrating the "agent-to-merchant" authorization process into existing card organization systems. AI agents can change the order entry point, but transactions may still rely on card networks, tokenization, authorization, risk control, and chargeback processing.

The biggest unresolved issue is liability attribution. In traditional cash-on-delivery transactions, fraud losses are typically borne by the issuing bank; in cardless transactions, merchants bear more responsibility; the liability structure for non-card payments is more decentralized; and the boundaries of liability for agency-based businesses are still unclear. Misguided purchases, intent verification, fraud losses, chargebacks, and returns will all impact widespread adoption.

Visa's tokenization rate for cardless transactions illustrates the adaptability of its existing network: 4% in 2019, rising to 54% by 2025. This path supports the conclusion that when new behaviors emerge, mature networks are not necessarily replaced; instead, they may increase stickiness through security, authentication, and value-added services.

Goldman Sachs estimates that a 5% increase in the growth rate of Visa/Mastercard's cybersecurity-related value-added services would accelerate value-added service growth by 90 and 200 basis points respectively, and boost total net revenue growth by 20 and 80 basis points respectively. Modern consumer finance and BNPLs also benefit—agency-based shopping may bring financing options to an earlier stage of the purchase process.

Identity and Anti-fraud: Risk Control Logic Forced to Rewrite

The more agent-based commerce develops, the more easily traditional anti-fraud signals become ineffective. In the past, risk control relied on devices, IP addresses, browsing behavior, mouse movements, typing rhythm, and cookies; when AI agents execute transactions, these signals become unreliable—requests may come from shared cloud infrastructure, IP addresses may belong to model providers, transaction speeds exceed normal human levels, and there is no typing rhythm or mouse movement.

The risk control logic has therefore shifted from "judging whether a person resembles a real person" to "judging whether the agent is registered, authorized, who the underlying consumer is, and whether the transaction is within the scope of authorization." The new key control points include agent identity credentials, the binding relationship between the consumer and the agent, the relationship between persistent identity and account device, transaction context, and additional confirmation for high-risk actions.

At the identity and decision-making levels, Equifax, TransUnion, and FICO occupy different positions. Equifax has the broadest coverage, spanning consumer identities, synthetic identity risks, enterprises, tax IDs, payment accounts, credit, income, and employment data, covering multiple verification nodes in the transaction lifecycle. TransUnion excels in persistent identity resolution and cyber intelligence, linking consumers, accounts, devices, and digital activities; the value of this type of identity graph increases significantly when legitimate agents and malicious traffic share the same cloud infrastructure. FICO primarily operates at the decision-making level; when agents initiate funding comparisons or applications, FICO is closer to determining "whether to approve, how to price, and whether there are any irregularities."

Cybersecurity: The Gateway to Proxy Commerce

As legitimate automation increasingly resembles malicious automation, cybersecurity investment will be driven ahead of schedule.

Cloudflare and Okta complement each other on this chain. Cloudflare stands at the merchant edge: Web Bot Auth encrypts and authenticates automated traffic; Bot Management distinguishes legitimate shopping agents from malicious bots; API Shield protects catalogs, inventory, accounts, and checkout APIs; and AI Crawl Control manages AI crawlers' access to website content. The real questions merchants care about are: who can access product data, who can place orders, who is simply scraping, and who is hoarding inventory or engaging in credential stuffing attacks.

Okta addresses the issues at the identity and authorization level, resolving how agents and consumers are linked, which accounts and data agents can access, what actions require additional approval, and when agent permissions are revoked. Its Auth0, Token Vault, Async Authorization, Fine-Grained Authorization, and Okta for AI Agents correspond to the authorization chain rather than website protection.

For agency-based commerce to take off on a large scale, five questions must be answered first: Is the agent genuine? Has the consumer authorized them? What can they access? Which actions require confirmation? Is the current activity malicious? If this process fails, the smoother the shopping experience, the greater the risk.

Ticketing Market: Opportunities and New Tolls Coexist

Live entertainment ticketing clearly demonstrates the opportunities and risks of agency-based businesses.

The opportunity lies in discovery and personalization. AI agents can help consumers find events by price, location, artist, venue, and time window, eliminating the need to switch between multiple websites. The integration of Ticketmaster and Meta Muse demonstrates how agent-based discovery can be layered onto existing infrastructure; StubHub sees AI as an incremental customer acquisition channel. Agents can also increase attendance by proactively pushing events with remaining availability to interested users.

However, distribution channels will not be free in the long term. In the short term, if the customer acquisition cost of AI channels is lower than that of paid search, the secondary ticketing market will see a decrease in customer acquisition costs and an improvement in profit margins. As the scale expands, agency platforms may gain pricing power, and the ticketing market will need to pay referral fees, display fees, or commissions to ensure appearance in AI-generated recommendations.

The greatest risk lies with platforms that offer homogeneous inventory. In secondary ticketing, the same seat type may be listed on multiple platforms. Once AI compares all-inclusive prices, competition will shift from brand and marketing to price and fulfillment reliability, putting pressure on service fees. Agents fulfilling explicit requests and reducing user browsing activity will also lower the value of ancillary sales and paid in-site display. Therefore, primary ticketing platforms with differentiated inventory are in a better position, while secondary ticketing platforms face the dual pressures of commoditization and new distribution fees.

The real dividing line: who controls the "default selected" position?

The essence of agency-based commerce is to delegate more steps in the shopping chain. The earliest changes are in discovery, comparison, recommendation, and replenishment; the most difficult to change are the highly personalized, high-risk, and high-responsibility-cost decision-making processes.

This main theme has a direct impact on the industry: advertising follows intent, payments run on existing networks, brands must become machine-readable, retailers want to access AI traffic but don't want to lose customer relationships, and identity and security companies are responsible for putting brakes on the entire system.

Goldman Sachs' research report ultimately concluded that the most valuable position is not necessarily the model itself, but rather the "default selected" position. Business competition in the era of AI agents may ultimately return to an old question: when consumers no longer manually flip through pages, search, and compare prices, who can determine the first recommendation?

~~~~~~~~~~~~~~~~~~~~~~~

The above content is from Zhuifeng Trading Platform .

For more detailed analysis, including real-time updates and firsthand research, please join the [ Trading Channel Annual Membership ].

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.