Shoppers let AI choose but not pay. CPG brands should design for the handoff

Every few weeks a new chart appears showing AI agents placing a tiny share of orders, and someone concludes agentic commerce is overhyped. A week later another chart shows agent traffic doubling, and someone concludes it is here. Both camps are asking the same question: when will the agent buy on its own?

We think that is the wrong question for a packaged goods brand in November 2026. The data we have points to a different shape, one that is already common and will stay common for a while. The assistant does the choosing. The person does the paying. And in between there is a handoff, a minute or less, when the shopper looks at what the assistant picked and decides whether to believe it.

That minute is where brands should be spending their attention.

People use AI to shop. They do not trust it with the card.

The cleanest numbers on this come from Visa. Its Trust Index, released September 9, 2026, was a Harris Poll survey of 2,065 US adults fielded May 26 to 28.

Show the data

Measure	Share of respondents
Have used an AI assistant	72%
Would trust Visa to handle agentic transactions	61%
Trust generative AI to handle payments on their behalf	23%

Source: Visa, September 9, 2026 (Harris Poll, May 26 to 28, 2026). 2,065 US adults; the payment-brand question was asked of 1,028 to 1,034 respondents per brand tested. Visa commissioned the survey.

Seventy-two percent had used an AI assistant. Twenty-three percent trusted generative AI to handle payments for them. Visa, unsurprisingly, highlights that trust rose to 61% when respondents were asked about Visa handling agentic transactions; it commissioned the survey and sells the payment rails. Set the self-interest aside and the gap is still the story: three in four people use these tools, fewer than one in four will let them pay.

Agents reach checkout often. People are still in the session.

Microsoft's analytics tool Clarity published the best behavioral data we have on what agents do on real sites. It looked at sessions involving AI agents from the OpenAI, Google and Anthropic ecosystems across 230,511 Clarity projects between August 13 and September 9, 2026, and compared the actions recorded with those in human sessions on the same sites, per the study.

Sessions involving an agentHuman sessions

Show the data

Action	Sessions involving an agent	Human sessions
Checkout	25.21%	14.65%
Submit form	19.85%	18.56%
Outbound click	14.28%	26.99%
Search	7.16%	4.79%
Login	6.83%	8.84%
Add to cart	2.33%	3.31%

Source: Microsoft Clarity, What AI Agents Actually Do on Websites. 230,511 projects, Aug 13 to Sep 9, 2026. Events are independent counts, not funnel steps, and agent sessions may include a human handoff.

Checkout was recorded in 25.21% of agent sessions and 14.65% of human sessions. Outbound clicks, the browsing behavior, ran the other way: 26.99% of human sessions against 14.28% of agent ones. Agents show up to get something done.

The caveat Clarity attaches is the important part for this argument. It says an agentic session "may also include a human handoff," and that it "cannot currently separate the agent and human portions into different sessions." It also warns that the events should not be read as steps in a funnel or as proof an agent completed a task on its own. In other words, the best behavioral data on agent shopping is, by its own description, data about agents and people sharing the job.

The trust gap is earned

It would be easy to call the 23% a consumer education problem. Visa's own November 5 benchmark suggests shoppers are being reasonable. Across 22 models in 15 shopping scenarios, 57.2% of runs passed, and 54% of the failed runs still ended in a purchase that missed a requirement or came before one was found. We charted the full results in today's data post.

If you knew your assistant picked the right TV a little over half the time, you would want to look before it paid too. That instinct is not going away this holiday season. It may fade as models improve, but for now the person stays in the loop because the person is right to.

So design for the handoff

If the near-term pattern is assistant picks, person approves, then the handoff is a moment brands can shape. Here is what happens in it. The assistant names a product, maybe a price and a retailer, maybe a reason ("best value for a family of four"). The shopper taps through, or opens the retailer app, or asks a follow-up. Within seconds they are looking for one thing: confirmation that the assistant was right.

Three things break that confirmation, and all three are fixable.

The product does not match. The assistant said "the 24-count" and the page loads the 12-count. The assistant said "fragrance-free" and the page shows a scented variant first. The shopper's doubt about the assistant becomes doubt about the product. Keep product names, pack sizes and variant names identical across your site and every retailer, so whichever page the assistant links to confirms what it said.

The price does not match. Assistants quote prices from feeds and pages that can be hours or days old. When ChatGPT's answer shows one price and the store shows another, the shopper hesitates, and some leave. That will get more visible: on November 10 Search Engine Watch spotted ChatGPT offering to set up a daily price-monitoring task for some shopping queries, per its report. It is unannounced and may be a test. But if shoppers start asking assistants to watch prices, stale or inconsistent prices will be noticed by software, not just by people.

The key fact is not visible. The assistant said "nut-free" or "safe for septic systems." The shopper scrolls the product page looking for it and finds marketing copy. The fact that the assistant used to choose you should be visible on the landing page without scrolling. We wrote a playbook for making those facts explicit today.

Which page gets the handoff matters

Where the assistant sends the shopper is often not your site. In ChatGPT, merchants that sell a product are ranked on availability, price, quality and whether the merchant is the brand owner or primary seller, as Stackline summarized on November 9 in its explainer on the ChatGPT buy box. In practice, many handoffs land on a big retailer. Tip 49 covers how to find which retailer ChatGPT's button sends your shoppers to and whether that listing is ready.

That means the handoff experience for your product is partly in someone else's hands. The retailer listing needs the same name, pack and facts as your own page. If it shows an old image or a vague title, the confirmation fails there, and the shopper may pick a competitor from the same retailer's suggestions.

And let the agent finish when the person says yes

Sometimes the person does say "go ahead." Then the agent has to be able to complete the job. Clarity's numbers say agents get to checkout often; whether they get through depends on your site. Bot protection that blocks every automated session, checkout pages that need a mouse hover, and pop-ups that cover the buy button turn a yes into an abandoned cart. Tip 46 covers testing whether an agent can finish a purchase on your site, and tip 47 covers which shopping agents are on your bot-protection allow list.

What we would not do

We would not build a strategy around fully autonomous agent purchasing for packaged goods this season. The trust numbers do not support it, the benchmark numbers explain why, and the behavioral data shows people still in the loop. We also would not dismiss agents because the order counts are small. As we argued in our earlier piece on agent orders, agents already shape what gets bought even when a person clicks the final button.

The middle position is less exciting and more useful. Treat every AI recommendation as the start of a handoff. Make sure the place the shopper lands confirms the assistant's claim in seconds. Measure it: teams increasingly check what assistants say about each product and where they send people, some with a tool such as SearchPanel, which XENA Intelligence makes for exactly that kind of per-product view, others with a simple monthly sheet.

The brands that do well over the next year will not be the ones with the cleverest agent integration. They will be the ones whose product, in the moment a slightly skeptical shopper checks the assistant's work, turns out to be exactly what the assistant said it was.