Shoppers Welcome AI to Browse, Not to Buy

The agentic checkout trust gap is the single most useful number in ecommerce right now: a 2026 Logica Research survey for Commerce and PayPal found that 63% of Australian shoppers are interested in trying agentic AI shopping tools, while only 4% want AI assistance at the point of payment. Shoppers are happy to let AI help them find things. They are deeply reluctant to let it pay.

That asymmetry is not a rounding artifact or a bad survey question. It shows up again across markets, across studies, and across the way people describe what they actually want from AI in a store. The interest is genuine and growing; the trust to delegate a purchase is not there yet. For merchants, the gap between those two facts is where the next few years of ecommerce strategy will be decided.

This guide reads the consumer data carefully — what people will let AI do, what they will not, and what they say it would take to change their minds. Then it maps the emerging trust infrastructure (the payment rails, identity proofs, and control mechanisms being announced through 2026) onto the journey, and lays out what a merchant can do today to be ready when the gap starts to close.

Start with the two numbers that frame everything else. In the Logica Research survey for Commerce and PayPal (March–April 2026, 1,000 Australian respondents), 63% said they were interested in trying agentic AI shopping tools — primarily for discovery and comparison. In the same survey, only 4% said they wanted AI assistance at the point of payment. That is the agentic checkout trust gap in a single line: openness at the top of the funnel, near-zero comfort at the bottom.

Most coverage reports one of these figures in isolation — the enthusiasm number or the security-anxiety number — and misses the shape of the thing. Read together, they describe a consumer who treats AI as a capable researcher and a suspect cashier. The chart below traces that gradient across the Australian journey.

The same survey makes the openness unmistakable: 79% of Australian respondents said they want AI to play a greater role in their future online shopping, and among current AI users, 53% reported spending less time searching for products. The appetite is there. What collapses is the willingness to let the agent act unsupervised when money moves. To make the drop concrete, the table below recomputes each stage as a distance from that 63% interest baseline.

Forty-three points lost between interest and actual use; fifty-nine points lost between interest and payment comfort. That second drop is the one merchants have to engineer around. The good news in the data is that the gap is about trust, not refusal — these are not people who hate AI, they are people who have not yet been given a reason to trust it with their card. Understanding why is the next section.

The cleanest way to read the data is as three distinct mental modes, each with a very different comfort level. Consumers welcome AI in the first, accept it conditionally in the second, and resist it in the third. The line they draw is not about capability — it is about who holds control at the moment value leaves their account.

There is a useful historical parallel here. When ecommerce was new, consumers trusted the products they were buying long before they trusted typing a card number into a web form. The friction was never the merchandise; it was the payment mechanism. Agentic shopping is repeating that pattern one layer up. People already trust AI to tell them which blender to buy. What they do not yet trust is AI to buy the blender. The resolution last time was not better marketing — it was payment infrastructure that made the risk feel contained, plus years of uneventful transactions. The same earning-the-right dynamic applies now.

That framing matters for how merchants build. The instinct is to chase the convenience story — fully autonomous, hands-off buying. The data says that is the least-wanted version. The version consumers are actually asking for keeps a human in the loop at the moment of payment and earns autonomy gradually. If you want the broader context on how these journeys are shaped end to end, our guide to agentic commerce covers the discovery side in depth.

The reluctance is specific, and the survey data names it. When Commerce and PayPal asked UK consumers what worried them about agentic shopping, the answers clustered around losing control of money rather than vague unease about AI. Spending without approval topped the list, followed by account security and getting the wrong thing.

The flip side of fear is a precondition. Across markets, consumers are clear about the bar agentic payments have to clear: security at least as good as what they use today. In the same survey family, 79% of Australian, 82% of US, and 83% of UK respondents said they expect agentic AI services to offer payment security equal to or better than existing methods. That is not a nice-to-have — it is the floor below which the agent never gets to touch the checkout.

The Australian figures are the most-quoted, but the broader survey covered the US and UK too — 3,000 respondents, 1,000 per country, consistent methodology. Interest in agentic shopping ran from 63% in Australia to 67% in the US, and the security expectation sat in the high seventies to low eighties everywhere. A separate Visa study of 14 Asia-Pacific markets, run earlier in 2026, lands in the same place from a different angle.

The Visa row is worth reading carefully — it measures something slightly different (AI use for discovery, and willingness to adopt with stronger security), so it is corroboration, not an apples-to- apples fourth data point. But the direction is identical: high comfort discovering with AI, a meaningful share reluctant to share payment details, and security as the lever that moves the reluctant group. When two independent studies, different vendors, different markets, point the same way, the pattern is probably real rather than an artifact of one questionnaire.

The trust gap is a problem the payments industry has noticed, and 2026 brought a cluster of announcements aimed squarely at it. Most coverage treats them as separate launches. It is more useful to read them as layers of a single trust stack — the rails, identity proofs, and control mechanisms a merchant will need to make agentic checkout feel safe. The table below maps consumer expectations at each stage of the journey against merchant readiness and the standard that addresses it.

The most-watched piece of this stack is Mastercard Verifiable Intent, announced March 5, 2026 — an open-source cryptographic framework, co-developed with Google, that aims to create tamper-resistant proof that a consumer actually authorized a given agent transaction. It links identity, intent, and action into one privacy-preserving record and is built on open standards from the FIDO Alliance, EMVCo, IETF, and W3C. On the rails side, Stripe Shared Payment Tokens let agents initiate payments without exposing a buyer's card credentials, scoped to specific merchants. And the Agentic Commerce Protocol (ACP) and Google's agent payment protocols define how the agent, merchant, and payment provider talk to each other in the first place.

Here is the part almost no editorial coverage mentions. Every debate about payment trust assumes the agent has already found your products. For most merchants, it has not. A PayPal merchant survey (498 US merchants, early 2026) found that only about 20% have 80% or more of their catalog in structured, machine-readable form — the format an AI agent needs to read a product, compare it, and put it in a cart.

That reframes the whole problem. If an agent cannot parse your catalog, the trust gap is irrelevant to you, because you never enter the consideration set in the first place. The discovery stage — the one part of the journey where consumer comfort is already high — is gated by a data problem that sits entirely on the merchant side. The same survey found data security ranked as the number-one barrier merchants cited across every business size, and roughly two-thirds agreed a standardized liability framework is urgently needed before agents transact at scale.

The data points to a clear sequence. You do not have to solve autonomous checkout to benefit from agentic commerce — you have to get discoverable, get the payment rails ready, and design for the control consumers keep asking for. In priority order:

None of this requires betting on a single protocol winning. It requires getting your house in order so that whichever rails reach critical mass, you are ready to plug in. For merchants who want to move on this deliberately rather than reactively, our ecommerce practice helps with exactly this kind of readiness work — and if your near-term problem is the conventional funnel, our guide to checkout optimization still applies to the human-driven majority of your traffic.

Project the data forward and a picture emerges. Today agentic transactions are a sliver — Checkout.com put AI agents at roughly 3% of UK and US transactions in mid-2026 — but the intent curve is steep: 57% of Australian non-users say they are likely to try within a year, and 89% of merchants in the Checkout.com study were already preparing for agentic commerce, with 72% expecting consumers to adopt faster than merchants can keep up. Adobe has reported that AI-sourced traffic to US retail sites rose sharply year over year in early 2026, though that figure is vendor-stated and worth treating as directional rather than independently verified.

The likely path is not a sudden flip to autonomous buying. It is a slow widening of the spending envelope. Checkout.com found consumers would, on average, allow an agent to spend about £177 per transaction without sign-off — while merchants assumed the figure was nearer £200. That small gap is the whole story in miniature: trust is granted in increments, and it is granted to whoever proves reliable first. Expect the 4% payment-comfort number to climb not because consumers change their minds about control, but because the trust stack makes giving up a little control feel safe. Merchants who are discoverable, on agent-safe rails, and visibly in favor of consumer control will be the ones who capture that climb. The reluctance the data captures is not a verdict — it is a brief, and the brief is winnable.