When OpenAI and Stripe announced the Agentic Commerce Protocol in late 2024, the question for retailers shifted from "will this happen" to "how do we participate." Within weeks, major brands launched integrations. Within months, distinct strategic patterns emerged.

McKinsey projects agentic commerce could generate up to $1 trillion in US B2C retail revenue by 2030, with global projections reaching $3 trillion to $5 trillion. But the path to capturing that value isn't singular. Different retailers are making fundamentally different bets on where to position themselves in the agentic ecosystem.

Three approaches stand out: Instacart embedding itself as the fulfillment layer for AI platforms, Walmart building unified "super agents" that connect their entire retail ecosystem, and Michael Kors using AI to own the high-touch consideration phase. Each approach carries different implications for measurement, customer relationships, and infrastructure requirements.

Instacart and the fulfillment layer strategy

Instacart made a strategic bet that the storefront is becoming less important than the infrastructure behind it. Rather than driving traffic to their app, they embedded their entire shopping experience inside ChatGPT.

Instacart became the first partner to launch an app on ChatGPT that offers a complete shopping cycle from query to payment without requiring the user to leave the conversation interface. A user can ask ChatGPT for meal planning help, and the agent builds a cart based on local retailer inventory. The transaction processes directly within the chat using the Agentic Commerce Protocol.

This fixes what Instacart calls the "handoff" problem. Previously, AI models could suggest products or generate meal plans, but execution required deep-linking to a separate application, often resulting in cart abandonment. Now the entire flow happens in one place.

The technical foundation matters here. Instacart's CTO Anirban Kundu noted that powering shopping inside an AI agent requires technology capable of interpreting highly local and constantly fluctuating inventory. Instacart grounds the AI's responses in a dataset covering more than 1.8 billion product instances across 100,000 stores. This reduces the "hallucination" risk where an agent might sell out-of-stock items.

Instacart contributed to OpenAI's Operator research preview, helping define parameters for how AI agents interact with external fulfillment logistics. Their complex data environment served as a testing ground for agentic capabilities.

The strategic implication is clear: Instacart accepts that consumer entry points are fragmenting. Rather than forcing all traffic through a proprietary app, they position their infrastructure as the backend fulfillment layer for third-party AI platforms. They've stated their intention to act as a primary partner for major AI players including OpenAI, Google, and Microsoft.

Measurement implications

This approach creates significant measurement challenges. When a customer completes a purchase inside ChatGPT, Instacart receives the transaction but loses visibility into the consideration phase. They don't know what alternatives the customer considered, what factors drove the decision to use Instacart over competitors, or whether any marketing influenced the purchase.

The customer journey that matters most, the moment of choosing Instacart as the fulfillment provider, happens in an environment Instacart doesn't control. Their measurement capabilities are limited to transaction data and post-purchase behavior.

For retailers considering this path, the trade-off is clear: broader reach in exchange for reduced visibility. You become infrastructure, which scales efficiently but commoditizes your relationship with customers.

Walmart and the unified brain strategy

Walmart is taking a different approach. Rather than embedding into external AI platforms, they're building a unified "super agent" architecture that connects their entire retail ecosystem.

Walmart CEO Doug McMillon described this as a pivotal moment where agentic AI allows them to "rewire how we work, serve and grow." Over the past several months, Walmart has developed AI agents built to make work easier across customer service, associate tasks, and supplier management. Now they're connecting these agents through a unified agentic framework so they can work together.

The architecture uses Anthropic's Model Context Protocol (MCP), which standardizes how AI agents connect to tools and data sources. This prevents "bot sprawl" where separate chatbots handle separate tasks. Instead, Walmart has a single interface that routes complex requests to the right backend system.

Their agents include Sparky for customer-facing interactions, which can handle requests like "plan a unicorn-themed birthday party for 8 kids" by identifying intent across categories and building a cross-category cart. Associate-facing agents help floor staff locate items or answer policy questions. Supplier-facing agents help vendors manage inventory and ads.

The MCP advantage is interoperability. If Walmart updates their inventory database, they don't have to retrain every agent. The protocol ensures agents can read from the new data source automatically.

Walmart reported that ecommerce sales grew 28% in Q3 2026, marking their 14th straight quarter of more than 10% year-over-year growth. They're also partnering with OpenAI to let customers shop Walmart products directly through ChatGPT, but as one channel among many rather than the primary strategy.

McMillon noted that "agentic commerce will be a very important growth aspect or channel" for Walmart, but emphasized their omnichannel approach. They're also testing capabilities in different markets, including a feature in Chile where Walmart creates suggested orders for consumers based on shopping data and sends WhatsApp prompts asking if they're interested. That feature has become about 20% of their ecommerce business in Chile.

Measurement implications

Walmart's unified architecture creates measurement advantages that the fulfillment-layer approach doesn't. Because they control the agent infrastructure, they can capture signals across the entire ecosystem.

When Sparky handles a customer request, Walmart can track the intent expressed, the categories considered, the products viewed, and the final purchase. When the same customer interacts with associate-facing or supplier-facing agents, those signals can connect to build a unified view.

The MCP architecture also means data flows through infrastructure Walmart controls. They can instrument that infrastructure for measurement, enrichment, and compliance verification at the point of collection.

The trade-off is complexity. Building and maintaining a unified agent ecosystem requires significant investment. Walmart has the scale to justify this investment. Smaller retailers may not.

Michael Kors and the consideration phase strategy

Michael Kors represents a third approach: using AI to own the high-touch consideration phase rather than optimizing for transaction speed.

Michael Kors deployed Mastercard's Shopping Muse, an AI-powered recommendation engine developed by Dynamic Yield. The tool, labeled "Your Style Assistant," acts as a digital stylist that translates colloquial language into product recommendations.

Unlike mass-market agents that prioritize efficiency, Shopping Muse handles aesthetic and contextual queries. A user can search for "minimalist aesthetic for a beach wedding" or "cottagecore outfit" and receive curated product matches. The agent translates the language of style into the language of inventory.

The results suggest this approach drives conversion. Early tests showed Shopping Muse generated approximately a 15% to 20% higher conversion rate compared to traditional keyword search.

This strategy acknowledges that for luxury and fashion, the protocol layer matters less than semantic understanding. The value isn't in transaction speed but in the quality of consideration. Customers shopping for premium products want a curated experience that understands context, not just keywords.

Michael Kors activated the tool from their home page, keeping the experience within their controlled environment. The agent appears as a popup with a search box and suggested terms. Results display products categorized by aesthetic themes.

Measurement implications

This approach preserves measurement capabilities that other strategies sacrifice. Because the agent operates within Michael Kors' owned environment, they maintain visibility into the entire consideration phase.

They can see what queries customers enter, which aesthetic terms resonate, how customers navigate recommendations, and which products ultimately convert. The behavioral data that powers attribution, personalization, and customer understanding remains accessible.

The limitation is reach. Customers have to come to Michael Kors first. The agent enhances conversion for existing traffic but doesn't capture demand that originates elsewhere. For a luxury brand where discovery often happens through brand-building rather than search, this may be acceptable. For mass-market retailers competing on convenience, it isn't.

What these strategies mean for your infrastructure

Each approach implies different infrastructure requirements and measurement capabilities.

The fulfillment layer strategy trades measurement visibility for reach. You become the backend for AI platforms with massive user bases, but you receive transactions stripped of the context that informs optimization.

The unified brain strategy requires significant investment but preserves measurement capabilities. You control the infrastructure, which means you can instrument it for the signals you need.

The consideration phase strategy works within existing measurement paradigms. You enhance the experience on surfaces you already control, maintaining visibility while adding AI-powered personalization.

The first-mile question

Regardless of which strategy you pursue, one infrastructure decision shapes everything downstream: where and how you capture data at the point it enters your systems.

For Instacart, this means capturing as much signal as possible from the limited transaction data they receive, then enriching it with customer context from their existing systems.

For Walmart, this means instrumenting their MCP infrastructure to capture intent signals as agents route requests across their ecosystem.

For Michael Kors, this means ensuring their owned-environment data flows cleanly to downstream systems with identity resolution and consent verification intact.

In each case, the architectural principle is the same: capture data at the trust boundary where it first enters your control, enrich it while signals are fresh, and route complete context to downstream systems.

Companies like MetaRouter focus specifically on this first-mile infrastructure layer, operating at the boundary where controller-side capture, identity resolution, and consent verification happen before data reaches CDPs, analytics platforms, or retail media systems. Whether transactions arrive from AI agents, web browsers, mobile apps, or in-store systems, the same infrastructure handles capture and enrichment.

The agentic commerce landscape is still forming. Protocols are evolving. Consumer adoption patterns remain uncertain. But the infrastructure you build at the first mile serves you regardless of which strategies prove dominant. That foundation starts where your data enters your systems.

‍