Agentic Commerce Is Stuck Below 1% of US E-Commerce: Why Liability, Not Technology, Is the Blocker

TL;DR: Agentic commerce is not waiting for better AI. It is waiting for a workable answer to one commercial question: who pays when an authorised AI agent makes the wrong purchase? Bernstein estimates agentic commerce remains below 1% of US e-commerce. Visa has reported “hundreds” of secure agent-initiated transactions in closed beta, while projecting “millions” of consumers using AI agents by the 2026 holiday season. That gap is the market reality. Before Black Friday 2026, payment providers should sell liability controls, verified intent, tokenisation and dispute readiness, not simply another checkout protocol.

Agentic commerce is below 1% because risk has not been allocated

The technology can search, compare, select and pay, but the commercial model for mis-execution is still unresolved.

Bernstein estimates that agentic commerce currently represents less than 1% of US e-commerce. That is difficult to reconcile with the scale of industry investment, but the data tells a consistent story: AI agents are generating attention faster than they are generating completed payments.

Hypertrade has reported a 4,700% year-on-year increase in AI-generated traffic to retail sites. However, that traffic is converting only minimally into actual payments. The discovery layer is working. The transaction layer is not yet trusted enough to scale.

Visa’s own adoption gap makes the point clearly. In December 2025, Visa reported “hundreds” of secure agent-initiated transactions in a closed beta. It projects that millions of consumers could use AI agents for purchases by the 2026 holiday season. Moving from hundreds of controlled transactions to millions of real-world purchases is not a normal product rollout. It requires a new operating model for authorisation, disputes and loss allocation.

The real blocker is liability for an authorised mistake

Merchants will not expose their full catalogue to AI agents until they know who absorbs the cost of a wrong, duplicated or unwanted purchase.

PYMNTS Intelligence data shows that 93% of merchants believe the AI provider should bear the financial loss when an agent makes an incorrect purchase. Yet only 28% are currently willing to make their full product range available to AI agents.

This is a commercial standoff, not a software problem. An agent can comply perfectly with an API and still misunderstand a consumer’s intent. It may select the wrong product variant, ignore a delivery constraint, buy from an unauthorised seller or repeat a transaction after a timeout. Existing card and electronic funds transfer rules were not designed around software acting with delegated authority.

In the US, Regulation E covers electronic fund transfers, liability and error-resolution procedures, but it does not cleanly resolve the position of an AI agent that has been authorised to act and then exceeds the user’s practical intent. The US Treasury Office of Inspector General has flagged this ambiguity.

The AI AGENT Act, introduced in July 2026, focuses on fiduciary duties, access, transparency and logging. It does not fully allocate financial liability for agent misexecution. The Consumer Bankers Association has therefore recommended private network rules to define indemnities, chargebacks and loss-sharing between agents, merchants, issuers and payment providers.

Until those rules are clearer, every participant prices uncertainty into the programme, or declines to participate.

Consumers want agentic commerce, but not without control

Consumer trust is conditional: people will delegate shopping tasks before they delegate unrestricted payment authority.

Visa’s Earning Consumer Trust in the Age of Agentic Commerce report found that only 23% of US consumers trust generic generative AI to handle payment transactions. Trust rises to 61% when Visa is identified as the payment provider.

Around 85% of US respondents want transparency and control over how their data is used. About half would stop using AI agents if that control disappeared. PYMNTS Intelligence separately found that only 14% of consumers trust AI to execute purchases without manual verification.

This has an important implication for payment strategy: the winning experience will not be “set and forget”. It will be bounded delegation. Consumers may allow an agent to buy a replacement item under $100, book a hotel within a defined budget or reorder a known product. They will still expect to see the merchant, price, delivery terms, tokenised payment method and final permission boundary.

That is closer to a trusted mobile wallet than an invisible autonomous buyer. The same consumer behaviours that accelerated mobile wallet adoption, contactless payments and tap to pay (recognisable brands, clear confirmation and easy reversal) will shape AI agent payments.

Europe is building the control layer before mass adoption

Europe’s advantage is not necessarily better AI; it is that authentication, identity and regulatory expectations are being designed into the payment flow earlier.

Europe’s PSD3 and PSR framework assumes a human at the keyboard, which is forcing the industry to move from Know Your Customer towards Know Your Agent.

Live activity is already moving beyond laboratory demonstrations. Santander, Mastercard, ING and Worldline have completed end-to-end agent payments. On 2 July 2026, ING, Worldline and Visa completed an agentic payment in Germany using Visa Payment Passkeys. Visa also reported live agentic commerce transactions across Europe with more than 30 issuers and participating merchants.

Mastercard has enabled all issuers in Europe at network level for Agent Pay and established a Lisbon Centre of Excellence. Mastercard Europe President Kelly Devine described the shift as “a profound change in how commerce is initiated and executed”. Worldline’s Madalena Cascais Tomé has been even more direct: “Agentic commerce is no longer theoretical, it is production-ready today.”

The difference is structural. Europe is making agent identity, strong customer authentication, passkeys and user-defined controls part of the architecture. The US is still debating the loss position after the architecture has been deployed.

Protocol proliferation is adding cost without removing risk

Merchants face a standards problem, but integrating more protocols does not automatically solve liability.

At least five competing standards are now shaping the market: Visa Intelligent Commerce, Mastercard Agent Pay, Stripe’s Agentic Commerce Protocol, Google’s Universal Commerce Protocol and Meta Muse.

Indicative integration costs range from roughly US$5,000 for a basic single-protocol implementation to as much as US$500,000 for a multi-protocol enterprise deployment. That expense may be rational for a large retailer, but it is difficult to justify when agent-originated payments remain below 1% of US e-commerce.

Visa Intelligent Commerce Connect is positioned as a network-, protocol- and token-vault-agnostic single integration. It could reduce fragmentation, but it remains unproven at scale. Mastercard’s first live end-to-end agent payment in March 2026 was explicitly not a commercial rollout.

The market should therefore avoid confusing interoperability with risk transfer. A single API can simplify integration. It cannot decide whether the agent provider, merchant, issuer, acquirer or consumer pays for a bad outcome.

OpenAI’s checkout retreat confirms that discovery is ahead of payment

The strongest near-term model is “discover in AI, buy on the merchant’s site”, not fully autonomous in-chat checkout.

OpenAI retreated from in-chat checkout in early 2026 after Walmart data showed that purchases completed inside ChatGPT converted at approximately one-third the rate of transactions completed on Walmart’s own site.

That does not disprove agentic commerce. It shows that consumers still trust the merchant environment more at the point of payment. AI is effective at product discovery, comparison and intent formation. Retailers remain better at checkout optimisation, fulfilment promises, refunds, loyalty and post-purchase service.

Payment providers should treat the hand-off as a feature, not a failure. A verified agent can pass structured intent, cart details and a tokenised payment credential to a merchant-owned checkout. This preserves the retailer’s customer relationship while giving the agent a useful role.

What this means for payment providers before Black Friday 2026

The immediate product is not autonomous checkout; it is controlled, auditable and appropriately priced delegation.

Payment providers, acquirers, payfacs, ISOs and wallets should prioritise five actions:

Accenture projects that agents could handle 30% of online commerce, around US$3.1 trillion, by 2030. OpenAI’s reported 4% fee on agentic commerce transactions, combined with protocol integration costs, shows how quickly the economics could become material. But scale will follow clear accountability, not forecasts alone.

RivaTech advisory view: sell certainty before autonomy

Agentic commerce will grow, but the first winning providers will monetise trust, control and liability management rather than maximum automation.

The market should stop asking whether AI agents can pay. They can. The better question is whether the payment ecosystem can prove what the agent was allowed to do, authenticate the instruction, contain the downside and resolve the dispute.

For fintechs, POS companies, payfacs, ISOs, wallets, issuers and remittance businesses, the opportunity is to build the control plane between agent intent and payment authorisation. That includes tokenisation, policy engines, passkeys, risk scoring, machine-readable merchant data and closed loop wallet controls.

The RivaTech solutions team helps payments businesses turn emerging infrastructure into practical commercial strategy. Before Black Friday 2026, the strongest position is not “we enable autonomous checkout”. It is simpler and more credible:

We enable verified AI agent payments with clear controls, transparent pricing and a defined answer to who pays when the agent gets it wrong.