A reader comment last week pushed on the Meta AI question. The short version: Meta isn’t trying to own agentic commerce (a smart play given merchant control), and what they are doing is more interesting than most people are giving them credit for.

The Architecture Is Already Set

First, a quick recap of where we stand. The winning model in agentic commerce is the retailer-owned specialist agent. Walmart’s Sparky, Amazon’s Rufus (now Alexa Shopping), Michaels’ Ask Mike. The merchant keeps checkout, keeps the customer relationship, and keeps the data. They’ve essentially built an intelligence layer inside themselves. Anthropic’s Claude Commerce Agents blueprint, open-sourced in September 2026, formalizes exactly this structure: shopping agents and merchant agents that sit within the merchant’s perimeter, not as a consumer-facing oracle trying to own the relationship.

OpenAI went the other way and tried to own the transaction end-to-end with its shopping agent, then quietly retreated toward advertising revenue. The lesson being written in real time: whoever controls checkout controls the economics, and no retailer is giving that up willingly (see Agentic Reality: Advertising)

According to Bain & Company’s Agentic AI in Retail report (May 2026), consumers already trust retail-owned agents three times more than third-party agents to complete transactions. That trust gap is structural, not temporary. It reflects decades of accumulated customer relationships, loyalty programs, and transaction history. A third-party agent asking to complete a purchase at Walmart on your behalf is a different proposition entirely from Sparky.

So Where Does Meta Sit?

Meta sits upstream of all of that. They are the discovery and intent-shaping layer before a consumer ever reaches a retailer agent. That’s always been their business: they’re an ad network whose core product is behavioral intent prediction at scale. And they’re getting dramatically better at it.

Meta’s Advantage+ system now processes more than 10,000 real-time signals per impression (scroll speed, app usage patterns, time of day, device type, prior engagement) to predict purchase intent before a consumer has articulated it. This is not incremental advertising optimization. It’s the foundation of a sophisticated intent-prediction engine that operates at a scale no retailer can match internally.

In June 2026, Zuckerberg launched Meta Business Agent, deploying AI-powered customer service and sales agents across WhatsApp, Instagram, and Messenger. The pitch: a clothing shop in Birmingham or a bakery in São Paulo can now offer the same always-on, personalized experience as a major brand. That’s real, and it matters for SMB commerce. But the deeper strategic play isn’t the customer-service layer; it’s what happens when you combine that conversational surface with 20 years of behavioral signals.

A Model That Already Exists

Here’s what I think is being underappreciated: Meta’s natural role in the agentic economy isn’t new. It’s an evolution of a structure I wrote about as far back as 2016.

In my Data Leakage post from May 2016, I described the fundamental tension in payment and behavioral data: the parties that hold the richest signals about consumer intent are rarely the parties with the cleanest path to act on them. Google understood this at Google Wallet, they realized they didn’t need to own the payment to close the loop in advertising. The data was the asset; the transaction was just a proof point.

Then in 2017, I covered Amex Advance and Acxiom, a concrete example of what I called collaborative data models: one party’s behavioral signals used to sharpen another party’s targeting, without either party fully surrendering their data. Amex had transaction history. Acxiom had demographic and lifestyle overlays. Neither gave the other raw data. The model combined them to produce targeting that neither could achieve alone.

Meta already does a live version of this today. Their collaboration with Nielsen Catalina Solutions links Meta ad exposure to real-world purchase behavior, allowing CPG advertisers to measure sales lift from social campaigns. The mechanism is a privacy-preserving match: your purchase data never leaves Catalina, your Meta ID never leaves Meta, but the model knows whether you bought the product after seeing the ad. That collaborative data infrastructure (retailers send SKU level purchase info directly to Meta) is production-grade and has been running for years. The agentic layer is simply the next evolution of this construct.

The Play: Preference and Affinity as a Managed Layer

Rather than building a turn-key shopping agent (Gemini’s current challenged play, and one that ran directly into retailer resistance), Meta’s natural posture in an agentic architecture is to operate as a preference and affinity layer: a specialized agent that knows a consumer’s style signals, social context, prior brand engagement, and inferred intent, surfaced at the discovery or cart-abandonment moment.

Think about what this looks like in the Anthropic Commerce Agents blueprint. The blueprint is a scaffold: a merchant agent (owns catalog, inventory, checkout) plus a shopping agent (owns the consumer relationship). What neither agent has, by design, is the pre-purchase behavioral context that determines which merchant the consumer encounters first. That’s the moment Meta is positioned to own.

A Meta affinity layer could operate as a specialized agent in this ecosystem: not closing transactions, but surfacing product-discovery signals to merchant agents based on demonstrated social preference, prior purchase behavior (via collaborative model), and real-time intent signals from Advantage+. Meta brings the targeting intelligence. Anthropic provides the scaffold. The retailer closes. That’s a clean three-party value chain with no checkout competition.

This is fundamentally different from Google’s posture, which wants to own the full stack from query to conversion. Meta’s advertising DNA is wired for influence monetization, not transaction ownership. They’ve always made money by directing intent, not by capturing it. Staying in that lane — but upgrading from ad impressions to agent-to-agent signals — is the logical evolution.

The Collaborative Model Is the Moat

What makes this durable as a competitive position is that the collaborative data infrastructure Meta has built with partners like Nielsen Catalina is genuinely hard to replicate. It’s not a technology moat — the matching algorithms are not exotic. It’s a relationship and trust moat. CPG brands, retailers, and data partners have spent years establishing the legal agreements, data governance structures, and measurement methodologies required to run these models at scale. The agentic layer doesn’t disrupt that infrastructure — it extends it.

Bain’s May 2026 research found that as agentic AI disrupts the top of the marketing funnel, traditional paid search becomes harder to attribute. Retailers who survive that disruption will be the ones with direct agent-to-agent data relationships that don’t depend on a Google or Meta impression to prove ROI. But the brands funding those retailer agents still need to know whether their marketing — including their Meta spend — is working. That measurement need doesn’t go away; it intensifies. Meta’s collaborative model infrastructure is the answer.

What to Watch

A few signals I’ll be tracking:

Advantage+ as an agent interface. Meta’s AI ad system is currently consumer-facing. The interesting move is when it becomes agent-facing: when a retailer’s merchant agent can query Meta’s intent graph directly, rather than running ad creative and hoping the algorithm delivers. That would be a genuine structural shift.

WhatsApp as the agentic surface. WhatsApp has 3+ billion users and is already the primary commerce interface in large parts of Latin America, South Asia, and Africa. Meta Business Agent deployed on WhatsApp is not an ad product, it’s a commerce OS. Watch how payment completion rates evolve. If checkout on WhatsApp becomes frictionless, the competitive dynamics for Western retailers get interesting quickly.

Collaborative model expansion. The Nielsen Catalina partnership is CPG-focused. Expansion into fashion, electronics, and general merchandise, where SKU-level purchase attribution is harder, would signal Meta is seriously pursuing the affinity-layer play I’ve described here.

How Anthropic’s blueprint gets forked. The Commerce Agents blueprint is Apache-licensed. The question is who builds the first integration that explicitly surfaces a Meta affinity signal into a merchant agent’s discovery flow. That would validate the whole architecture.

Meta isn’t going to own agentic commerce. But they don’t need to. They’ve spent twenty years building the infrastructure to shape what consumers want before they buy anything. In a world where the transaction layer is locked up by retailers and the intelligence scaffold is provided by Anthropic, the discovery layer is still up for grabs — and Meta’s position there is stronger than the current narrative suggests.

I’ll keep watching. More as it develops.

My Related posts:

• Data Leakage — May 2016
• Banks as a Data Business: Amex Advance/Acxiom — November 2017
• Retailer Actions in Agentic Commerce — May 2025
• Agentic Commerce Economics and Governance — May 2025
• Agentic Commerce — Inevitable or Unworkable? — September 2025
• Walmart and OpenAI — October 2025
• PayPal Advertising? — October 2025