Agentic Commerce Trust Gap: Shoppers Use AI, Keep the Wallet

Chat is becoming a storefront

The industry's push to make conversation the place where shopping happens picked up speed this autumn. On October 1, Shopify introduced Canvas, a store builder where merchants describe changes to Sidekick, the company's AI agent. Sidekick then edits the store's actual theme code and shows a live preview across devices.3 DoorDash rolled out an AI agent that takes food orders by text message, which moves the transaction out of the app and into a chat thread.3 Taken together, the launches suggest that large consumer brands now treat messaging as a real point of sale rather than an experiment.3

These products serve different users, since Canvas is for merchants and DoorDash's agent is for diners. Both rest on the same assumption: people will increasingly hand tasks to an AI over chat. The open question is whether consumers will let that AI finish the job, which means spending their money.

Adoption is broad but shallow

AI is clearly part of how people shop. An IBM-NRF survey of more than 18,000 respondents in 23 countries, published in January 2026, found that 73% of consumers use AI somewhere in their buying journey.1 The same research puts the share using AI for at least part of the buying process at 45%. It also reports that AI application usage rose 62% over two years.1 In the US, Capital One Shopping data indicates 59% of Americans have used generative AI tools for shopping.1 McKinsey found that 53% of US generative AI users also shop with it, and 44% of users prefer AI over traditional search.1

The numbers drop at the point of purchase, though. Morgan Stanley reports that only 23% of Americans made an AI-assisted purchase in the past month.1 That gap between "uses AI while shopping" and "buys through AI" says a lot. Consumers seem comfortable letting a chatbot research, compare and recommend. Industry surveys say 58% now see AI replacing search engines for product recommendations, and one in four think ChatGPT's recommendations beat Google's.1 Discovery and delegation are not the same thing, however. The figures are consistent with AI acting as an adviser while the final click stays with the human.

The security case for hesitation

That caution may be well founded. Visa's Payment Ecosystem Risk and Control team recorded a more than 450% increase in dark web posts mentioning "AI Agent" in the first half of 2026 compared with the previous six months.2 According to WorkOS's analysis, those posts focus on compromising agents, not building them. The goal is to hijack the delegated payment credentials agents hold for users and make purchases that look legitimate in every respect except intent.2

The attack vectors are specific. An agent holds access tokens, payment tokens and session credentials. These could be stolen through a compromised device, a malicious browser extension or a supply-chain attack on the agent's software dependencies, and the thief would then inherit all of the agent's authorized permissions.2 Impersonation is a second risk. DataDome counted nearly 8 billion AI agent requests across its network in January and February 2026, with significant rates of impersonation.2

This changes the usual fraud model. Traditional defenses look for transactions that seem wrong. A hijacked agent produces transactions that seem right, made with valid credentials by an authorized party. When the only signal is whether the user actually wanted the purchase, human review is one of the few dependable controls left.

Reading the gap

The sources look at agentic commerce from different angles: the market-sizing optimism in adoption statistics, the threat modeling of security vendors, and the product momentum covered in startup news. They point to the same conclusion. Usage is high and the infrastructure is shipping, but autonomy is lagging.

One caveat applies. Several of the adoption figures come from vendor-compiled roundups and broadly described "industry surveys," so the exact percentages deserve some skepticism. The direction is consistent across sources, though. Most shoppers welcome AI as a guide, and far fewer let it complete purchases.

I read this less as a temporary hurdle and more as a design requirement. Chat commerce probably grows fastest where the stakes and the friction are both low. A text-message food order is a good example, since the user confirms in the same thread. Fully autonomous buying will likely need visible guardrails before consumers accept it. Spending limits, confirmation steps and credential protections that can be shown to resist hijacking would all help. Platforms that treat human review as a feature rather than a bottleneck may earn trust first. Given what fraud teams are seeing, that approach is also the safer one.