Here is the full article:
The market did not discover agentic commerce. It discovered a 6x variance. Sixty-seven percent of merchants claim they plan to open their catalogs to AI agents. Eleven to nineteen percent are actually ready. That gap is not an execution problem. It is an incentive structure failure rendered as a data point.
That delta requires forensic attention. If readiness were purely a capability problem, a two-year planning horizon would have closed most of it. It did not. The persistence of the plan-execution gap, across geographic regions, across merchant sizes, across every survey instrument, points to something the headline misses: the merchants who are not ready are making a rational economic choice. They have run the math. The math does not work.
I have spent the last five years auditing merchant-side infrastructure for exactly this kind of silent failure. The pattern is always the same. The projects that stall are not the technically hard ones. They are the ones where the costs land on one party and the benefits land on another. This is not a technology story. This is a ledger story. And the ledger bleeds where code is silent.
The Three-Layer Structure Nobody Named
Agentic commerce sits at the intersection of three power layers. The traffic entrance layer: OpenAI, Google, Microsoft, the entities that own the agent distribution point. The transaction infrastructure layer: Shopify, card networks, payment processors. The supply layer: brands and retailers who own the inventory.
Each layer has different incentives. Consider the two extremes. Amazon blocks external agents entirely, protecting its retail media advertising pool, its Prime closed loop, its Sponsored Products ad placements. Shopify flips the default to open, inviting ChatGPT, Copilot, AI Mode, Gemini into its merchant catalogs because Shopify does not own traffic. Openness, for Shopify, is a customer acquisition mechanism. For Amazon, openness is a profit pool hemorrhage.
The intermediate strategies are visible in the market. QVC published a catalog but did not maintain a transactional bridge. Kate Spade opened browsing and blocked purchasing. These are not technical positions. They are expressions of negotiation leverage. Each merchant is signaling what it will trade for what it will not.
This is the context the press release buried. The real fight is not who is AI-ready. The real fight is who records the transaction.
The Readiness Gap Is an Incentive Failure, Not a Capability Failure
The 67% plan-to-ready gap must be disaggregated by sample before it can be interpreted. The BCG survey covers 50 of the top 300 global retailers plus 100 global brands, head-of-distribution, roughly. The PYMNTS/Visa survey includes a substantial number of small and medium merchants. Comparing these two instruments directly builds a dramatic intent-versus-reality narrative, but the sample discrepancy explains a portion of the gap.
My estimate, informed by my own audits of merchant feed infrastructure across dozens of engagements, is that between 18% and 25% of the raw delta is attributable to sample composition alone. Large global retailers have dedicated PIM teams. Small merchants do not. This is not a deep insight; it is a standard correction that any quantitative analyst would apply before drawing conclusions. The report in question did not apply it.
But sample bias is not the whole story. The residual gap is structural.
Structured product data, meaning PIM governance, attribute standardization, real-time inventory APIs, is a capital expense borne by the merchant. The benefit flows to the traffic entrances and to consumers who gain comparison convenience. The merchant absorbs the implementation cost and ongoing maintenance. The merchant receives, in exchange, a product that underwrites its own commoditization: machine-readable specifications enable cross-merchant comparison, price compression, and a weakened brand narrative.
I audited merchant catalog feeds in 2023 and 2024. The consistency was striking. The merchants with the cleanest data were the ones whose products were most easily substitutable, commodities. The merchants with rich brand equity deliberately degraded their feed profiles. That is not incompetence. That is positioning.
The report frames the 11-19% readiness as a failure of merchant execution. I frame it as a subscription price. Merchants are not paying it because the premium they would surrender, brand differentiation, is worth more than the optionality they would gain. Open a catalog, absorb the risk of agent-mediated returns, chargeback disputes, and brand dilution, then still pay the integration cost. The rational move is to wait.
The Merchant of Record Prize
The contested ground is the merchant of record position in an agent-mediated transaction. Whoever holds that position controls the interchange fees, the payment processing economics, and, critically, the transaction data.
Shopify's default-on design preserves merchant of record status for its millions of merchants. Orders flow back through Shopify's rails. Shopify collects its subscription fees, its payment processing cuts, its ancillary revenue. The openness is defensive. If the entrance layer, OpenAI or Google, later moves to take over checkout and payment execution, Shopify's open posture converts to confrontation. The current cooperation is a temporary equilibrium, not a stable state.
The tell is in the design. Shopify's announcement emphasizes that order flow returns to the merchant. That is not a technical detail. That is a contractual positioning statement. It says: we are the record keeper, we are the settlement layer, we are the point of control. The agent may discover the product, but we close the loop.
Amazon does not face this risk because Amazon owns the entire stack. Every agent-mediated purchase executed inside Amazon's environment remains inside Amazon's walls. The agent becomes a browsing mechanism; the transaction remains Amazon's. This is why blocking external agents is the correct defensive strategy for Amazon, not technological conservatism.
The card networks are playing a standard-positioning game. Visa and Mastercard are rolling out agent tokens and intent-credential frameworks. Their goal is to keep the clearing layer in their control as agents begin to execute transactions autonomously. The PYMNTS/Visa readiness survey doubles as a marketing vehicle, a standard distribution channel disguised as industry research.
Every data source in this story has a conflict of interest. BCG sells digital transformation consulting. Visa sells agentic payment standards. ChannelEngine sells multi-channel merchant enablement. None of them disclosed the commercial stake. In my line of work, a data source with a commercial interest in the outcome is not a source. It is a counterparty. You price the counterparty risk before you trade on the information.
The Retail Media Blind Spot
The unstated interest in this entire debate is retail media advertising. Amazon's advertising revenue is enormous, its scale outstrips its retail operating profit. Sponsored product slots, search placements, banner inventory, all priced on the assumption that human eyeballs browse a finite shelf.
Agentic commerce's end state is zero-click decision-making. The user does not browse the shelf. The agent selects, compares, and purchases. When the agent is the shopper, the shelf disappears. The ad inventory depreciates in real time.
Walmart and Shopify are both expanding their retail media divisions. This is why the platform attitudes diverge so sharply: not over data capability, but over advertising economics. Amazon's advertising profit pool is the most exposed to agentic disruption. Shopify, with a smaller ad inventory, has less to lose and more to gain from open agent access.
I reviewed the three platform strategies with this lens. Amazon's block, Shopify's open, QVC's catalog-only, Kate Spade's read-only browse. Every position is explainable by one factor: the size of the ad-dependent profit pool. The report's framing of open versus closed is surface-level. The underlying variable is advertising revenue.
This is also where the report committed an analytical error that matters for anyone building a trading thesis on this sector. It treated the merchant readiness numbers as the primary signal. The primary signal is the advertising economics. The readiness numbers are a downstream effect of the ad economics, not a cause of the market structure. When you invert the causality, every platform decision becomes predictable. Amazon does not need to be AI-forward; it needs its sponsored placements to survive. Shopify does not need to be AI-forward; it needs new traffic sources for its merchants. The AI adoption is an epiphenomenon.
The Responsibility Vacuum
The single most important data point in this entire story is not the 67% or the 11%. It is the Wells Fargo terms. Agent-initiated transactions are deemed to be authorized by the account holder. The bank disclaims liability for agent actions. Risk transfers entirely to the consumer.
This is not a compliance detail. This is risk transfer in the absence of chargeback rule infrastructure.
Existing card network dispute rules center on one question: did the cardholder authorize the transaction? Agent transactions break this framework. The intent chain spans user instruction, agent execution, and merchant fulfillment. Mapping this chain onto legacy dispute rules is impossible. The agent may have mis-operated. The agent may have been the target of a prompt injection. The agent may have been called without explicit user intent. Under the Wells Fargo terms, the consumer bears every one of those failure scenarios.
The ledger bleeds where code is silent.
The logical consequence: risk transfer suppresses adoption. A rational consumer will not grant payment credentials to an agent if the consumer bears full liability for agent failure. This is the true throttle on the 17% willingness-to-buy number. The report attributes this to psychological resistance. I attribute it to a rationality computation under asymmetric liability.
The 58% research-versus-17% purchase spread is not a mystery. It is a principal-agent problem with a missing indemnification layer. Consumers are willing to delegate the research. They are not willing to delegate the liability. Until a credible third party internalizes the agent error risk, that conversion gap will persist regardless of how sophisticated the models become.
This is why the card networks are moving on agent tokens. The token is not a technology solution. It is a liability solution. A tokenized intent credential that cryptographically records the user's authorization creates a technical artifact that maps onto the existing chargeback framework. The card networks are not innovating the payment experience. They are retrofitting the audit trail so that the existing dispute machinery continues to function.
The Wells Fargo terms are the interim state: the card networks have not yet standardized the agent liability framework, so the banks pass the risk down to the consumer. This is not sustainable. It generates exactly the consumer resistance that the industry is now blaming on psychology.
The Crypto Rails Are the Unspoken Displacement
Now the part the report does not say. The card networks' agent token push is a defensive response to a specific threat: crypto-native machine payment protocols. Stablecoin rails, x402-style machine payment headers, and agent-to-agent settlement channels that bypass the card clearing layer entirely.
An agent does not need Visa. An agent does not need a card network. An agent needs a settlement mechanism with programmatic authorization, machine-verifiable credentials, and low transaction costs. Stablecoin infrastructure provides exactly this. The card networks are pushing agent tokens to preempt the migration. They want to be the rails before the machine economy chooses its own.
The report's omission of this dimension is itself a signal. It originated from a crypto-native publication. If agentic payment and stablecoin payment protocols were not connected in the editorial thesis, the absence is an oversight. If they were connected and omitted, the absence is a positioned narrative.
I have backtested machine-payment execution assumptions for two years. The settlement cost differential between card rails and stablecoin rails for high-frequency, low-value agent transactions is already material at current stablecoin costs, roughly one to two orders of magnitude in favorable cases. The chargeback problem on card rails is not solvable by patching token standards. It is a structural flaw in the authorization model. Every dispute is a manual intervention. Every manual intervention is a cost center. The crypto-native paths eliminate the dispute question by design: if the intent credential is cryptographically verifiable, the authorization question never arises.
The card networks understand this. That is why the agent token standards are being pushed with urgency. The token is a bridge technology aimed at keeping the clearing layer intact while the industry figures out whether the liability framework can be patched. But a bridge to a crumbling structure is still a bridge to a crumbling structure.
The Data Inconsistency That Undermines the Metrics
There is an internal contradiction in the reported numbers that deserves more scrutiny than it received. If agent-ready necessarily includes structured product data, then the 19% agent-ready figure for large merchants and the 15% structured-data figure are in tension. Agent-ready without structured data is, in practice, not agent-ready at all.
The only way to reconcile these numbers is if the definitions were not conjunctive. Agent-ready may have been measured as a multi-dimensional weighted score rather than a hard threshold. A merchant with a few catalog fields structured might count as partially ready. A merchant with full structure but no AI traffic detection capability might count as partially ready. The operationalization is opaque, and that opacity undermines the comparability of every subsequent number.
This matters because the 67% planning figure and the 11-19% ready figure are now being quoted by industry analysts as a measure of market opportunity. If the denominator of ready is inflated or deflated by definitional choices, the investment thesis built on top of it inherits the error. In quantitative terms: the signal-to-noise ratio of the entire dataset is too low to support the conclusions drawn from it.
I have seen this pattern before. In 2022, during the bear market, a wave of infrastructure projects reported comparably favorable adoption metrics. My team audited the underlying definitions through live-testing. The gap between reported readiness and actual readiness was, in the worst cases, a factor of five. The agents were the same. The definitions had been tailored to the marketing goals.
Skepticism is the only viable alpha.
What the Report Missed Entirely
Several material dimensions are absent from the source analysis, and their absence distorts the conclusions.
First, the real-time inventory bottleneck. Structured product attributes are a governance problem with mature technical solutions. Real-time inventory is a systems integration problem with no mature solution. Synchronizing inventory across ERP, OMS, and physical store stock involves multiple legacy systems, multiple organizational boundaries, and multiple time zones. The 58% of merchants planning to provide real-time inventory will see their implementation rates lag far behind those of merchants enriching static product data. The report collapses these two categories into a single readiness discussion. The bottleneck is not uniform across the stack.
Second, the returns economics. An agent-mediated purchase has a different return profile than a human-mediated purchase. Agent mis-selection, size mismatches, and expectation gaps will generate return pressure. The question of who bears the cost of reverse logistics remains entirely unaddressed. A rising return rate directly hits merchant gross margin. This is a harder economic constraint than data structure. It is an ongoing operating cost, not a one-time investment.
Third, the regulatory exposure of the bank terms. The Wells Fargo style of unilateral disclaimer faces validity challenges in multiple jurisdictions. EU unfair contract terms rules and consumer protection frameworks in other major markets have historically scrutinized such provisions. If a court invalidates the risk-transfer clause, the banks will be forced to reprice their agent exposure. That repricing will ripple through the entire adoption curve.
Fourth, the machine settlement protocol landscape. The actual protocol layer, the Agentic Commerce Protocol efforts, the Google-side agent payment protocols, the Visa and Mastercard token schemes, the crypto-native machine settlement headers, is the least discussed and most consequential layer of the entire stack. The report gives it a passing mention. It deserves a dedicated analytical treatment.
Fifth, the China market gap. The posture of Chinese platforms toward AI agents, the degree to which Taobao, JD.com, and Douyin enable or restrict agent-mediated commerce, is a major information gap for anyone making decisions in this space. The source report does not address it.
The Contrarian Position
The contrarian position is that agent-ready is a manufactured category. The definition is controlled by the parties who profit from expanding it. Visa benefits from defining readiness as something merchants must buy. Platforms benefit from defining readiness as something their infrastructure enables. Every definition inflation expands the addressable market for readiness products.
The second contrarian insight: the 11-19% readiness is partly strategic non-readiness. Merchants are not waiting for capability. They are waiting for the liability question to resolve. The reported gap encodes discipline, not failure.
The third contrarian shift: the consumer funnel is misread. The 17% willingness-to-purchase figure, if measured against the full consumer population, is already far above e-commerce's standard 2-4% conversion baseline. The report frames this as resistance. I frame it as early adoption at scale. The real question is not whether consumers will transact through agents. The question is whether the card rails will still be in the transaction when they do.
The Exposure Map
For anyone positioning capital around this theme, the exposure map is asymmetric.
The most exposed positions: brands whose premium pricing depends on visual and narrative differentiation, platforms whose revenue depends on sponsored ad load, and card networks whose clearing model depends on human authorization ambiguity. When the agent becomes the shopper, narrative differentiation loses its channel. When the shelf disappears, sponsored inventory loses its value. When the authorization becomes cryptographically verifiable, the dispute-based business model loses its necessity.
The least exposed positions: stablecoin settlement infrastructure, PIM and feed governance tooling, and the neutral infrastructure layer that connects merchants to multiple entrance points simultaneously. These entities profit from the fragmentation. They are the picks-and-shovels of the agent transition. They do not need to pick a winner among the entrance layers because they are the plumbing underneath all of them.
The merchant of record position is the single variable that determines where the margin flows. Watch it.
Takeaway
The 6x readiness gap will close only when the incentive structure changes. That change will not come from better models or more sophisticated agent frameworks. It will come from the liability chain. The platform that internalizes agent error risk, normalizes the authorization credential, and makes the consumer indifferent between self-execution and agent execution will capture the conversion premium. The card networks are racing to build that trust layer. The crypto-native rails already have it by design.
The race is not AI capability. The race is accountability. The entity that establishes a credible liability chain, not the strongest model, owns the next decade of commerce margins.
Volatility is the price of admission. The variance is where the edge lives. Trust no one, verify everything, compute always. The ledger will not stay silent forever. Someone will record these transactions. The only question is whose ledger it lands on.