AI shopping agents: cart and checkout preferences can differ

Product.ai had already published a useful comparison: 18.7% of the 550 respondents would let an agent add items to a cart without asking, while 16.9% would let it place an order without asking. Those figures describe each task separately. They do not reveal how much the two groups overlap. Our contribution is to join the two answers given by each respondent. ^{[1] [2]}

The broader distinction is already established. NMI’s 2026 research describes consumers who welcome shopping assistance while wanting to retain control over the purchase. This study asks a narrower question: when the same person answers both the cart and order questions, which preferences go together? We use one source and one question pair rather than comparing percentages from surveys with different participants and wording. ^[7]

There were 103 people in the first group and 93 in the second. Only 53 belonged to both. Another 50 accepted cart additions without asking but wanted to be asked before ordering or would never allow ordering. Meanwhile, 40 gave the reverse combination: ordering without asking, paired with an ask-first or never answer for cart additions.

That matters when a brief describes a single audience as comfortable with autonomous shopping. Similar totals can hide different people and different expectations. A team reading only the two headline percentages would not know which respondents expected the same treatment of both actions.

The table below keeps all nine combinations visible. Read across a row to see how a particular cart group answered the order question. Read down a column to see the cart answers within an order group. The cells are people in this selected panel, not stages completed in a shopping funnel or the settings of a live assistant.

Start with the 103 people who would let an agent add items without asking. Their answers to the order question split into three groups: 53 would also allow ordering without asking, 40 wanted the agent to ask first and 10 would never allow it. The corresponding shares are 51.5%, 38.8% and 9.7%. ^[3]

Here is the worked calculation: add the 40 ask-first answers to the 10 never answers, then divide 50 by 103. Multiplying by 100 gives 48.5% after rounding. The denominator is the group that accepted cart additions without asking. It is not all 550 respondents, all US shoppers or everyone who uses an AI shopping agent.

The distinction between the 40 and the 10 is useful. Asking first still leaves a role for the agent in placing an order. Never allowing that task is a different answer. Combining them describes a boundary on acting without asking; it does not mean that all 50 reject AI shopping assistance.

For a merchant, the result suggests separating two research questions that a broad enthusiasm measure can blur. Does a customer want help putting a selection together? What, if anything, would they want to review before the assistant places an order? The survey gives a reason to ask those questions separately. It does not prescribe the review screen or show that adding one improves sales.

A neat story would place everyone on one ladder: first allow research, then cart additions, then ordering. These paired answers do not justify that ladder. Of the 259 respondents who wanted to be asked before cart additions, 32 would allow ordering without asking. Of the 188 who would never allow cart additions, eight would allow ordering without asking. Together, those are the 40 reverse answers.

The remaining off-diagonal cells matter as well. There were 52 people who wanted the agent to ask before adding to a cart but would never allow it to order. Another 28 would never allow cart additions but would allow ordering if asked first. Keeping these answers avoids quietly turning the survey into a single scale of willingness.

Across the whole paired group, 380 people chose the same category for both tasks and 170 chose different categories: 69.1% and 30.9%. Agreement here means identical answers to two different task questions. It does not measure consistency, trustworthiness or the quality of a respondent’s answer. ^[3]

We cannot tell whether the reverse answers reflect different interpretations of a cart, a particular shopping scenario, measurement noise or genuine preferences. None of those explanations was tested in this analysis. Rather than discard those people or guess their motives, a merchant can use the ambiguity to improve a future interview: describe the action, then ask the customer what they think it would do.

The practical implication is to make the task clear before treating a response as support for a feature. “Would you use an AI shopping assistant?” leaves too much of the proposed behavior unstated. Someone can welcome help choosing an item while wanting to review the order themselves. Someone else may imagine a repeat order that needs no new cart-building step.

In a merchant brief, describe cart assistance and order placement in separate sentences. For the first, say whether the assistant would suggest an item or actually add it, and whether the customer could inspect the selection. For the second, say what “place an order” means in the proposed experience and when the customer would see the details. These are questions for the team to specify, not details supplied by this survey.

Then keep the three possibilities distinct: do the stated task without asking, ask before doing it, or do not do that task. Avoid summarizing the first two as a single preference for automation. An ask-first answer contains a condition that the broader label would erase.

Keep the two answers attached to the same person in the research summary. Separate percentages remain useful, but they answer a different question. A paired table lets the team see whether the people comfortable with one action are also comfortable with the other, including combinations that do not fit the team’s expected sequence.

These recommendations concern research and communication. A stated preference is not a grant of payment authority, a legally sufficient consent process or an instruction to change a customer’s settings. Those operational decisions need their own requirements. The narrower benefit here is that the brief asks a clearer question before the team promises what an assistant will do.

Imagine a homeware merchant preparing a creator demonstration of a shopping assistant. This is an invented example, not a business observed in the survey. The initial brief asks the creator to show the assistant helping a customer choose a desk lamp and complete the purchase. That sentence leaves the customer’s involvement at each step unclear.

The revised brief separates the actions. One part shows the selected lamp being added to a cart. Another shows the order details and the point at which the proposed experience would ask the customer to review them. The merchant would still need to verify that the actual product behaves as described. The survey does not establish that this particular design is the right one.

Before filming, the team can show the proposed wording to prospective customers and ask them to explain what they expect the assistant to do. A useful follow-up is whether their answer changes between adding the lamp and placing the order. If someone gives a reverse combination, the interviewer should ask what each task means to them rather than mark the answer wrong.

The creator then needs a precise promise, not the survey percentage as a sales claim. For example, a demonstration brief could require the creator to distinguish selecting an item, adding it to a cart and confirming an order. Any statement that the customer is asked first must be checked against the real experience before it appears in published content.

The 50-of-103 finding earns its place in the internal rationale: it shows why a single cart preference may be insufficient evidence for an order preference. It does not belong in the demonstration as “half of customers need this feature.” The study used a selected panel, and it did not measure what happened after anyone saw the proposed interface.

A reusable brief can start with this line: “We are evaluating how customers want the assistant to handle cart additions and order placement.” Keep the following fields together so the eventual campaign summary can be traced to the questions people actually answered.

Tasks and context: “For this version of the experience, adding to a cart means [describe the action]. Placing an order means [describe the action]. Before answering, the customer sees [describe the explanation or demonstration].” The blanks matter: a response to an unspecified assistant is not interchangeable with a response to a demonstrated product flow.

Questions and answers: “For each task, would you want the assistant to act without asking, ask first, or not do the task?” Record both answers for the same person. If a customer gives a combination the team did not expect, retain it and ask them to explain what they understood. Do not convert the response into a more convenient category.

Results and denominator: “Among [number] eligible people who answered both questions, [number] chose each paired combination. Among [number] who would allow cart additions without asking, [number] gave each order answer.” Include missing answers separately. Keep ask-first and never distinct even if a headline also combines them.

Decision and owner: “These answers will inform [the specific wording or demonstration decision]. [Named reviewer] will check the campaign’s promises against the actual product behavior. Before deciding [a separate operational question], we still need [the relevant evidence].” This prevents a customer-research summary from quietly becoming a specification for autonomous ordering.

For the hypothetical homeware launch, the immediate decision might be which moments the creator must explain and which claims the merchant can substantiate. It would not be a forecast of how many customers will buy. The useful habit is to preserve the relationship between two answers before turning either one into a broad claim about an audience.

We used Product.ai’s Edition 2 release, published October 14, 2026. Fieldwork ran August 27 through September 1, 2026, using Alchemer and a panel of US adults who shop online or in stores. The publisher reports age, gender and regional quotas, with no weighting to Census distributions. Recruitment details, inclusion probabilities and incentives are not supplied. Product.ai sells verification products for commerce and publishes the research, a relevant commercial interest in the topic. ^{[1] [5]}

The released file contains 1,493 complete responses. Q12 identifies 550 people who said they had used an AI assistant to research a product they were considering in the preceding 90 days. We retained that source-defined group, including online and store-only shoppers. Recent product research does not establish that someone has used an autonomous shopping agent or completed an AI-assisted order.

All 550 eligible respondents answered both selected Q36 items with one of the three allowed options. The other 943 respondents had blanks because they were routed past the AI-user questions. We checked those routes, unique identifiers, source dates and response values before constructing the paired table. No blank became a never answer, and no answer was removed because its combination seemed unexpected. The two published task margins reproduced at their stated one-decimal precision. ^{[2] [5]}

The analysis fixed the cart/order pair, all nine combinations, the conditional split within the autonomous-cart group and the reverse combinations before calculating the joint results. Each cell is available as a count, a share of all 550, a share of its cart row and a share of its order column. Those three percentages have different denominators and should not be substituted for one another. ^{[3] [4]}

We report descriptive counts and percentages without population confidence intervals, significance tests or causal claims. Complete answers within the routed group do not remove selection or measurement uncertainty. Q12 concerns past research; Q36 concerns hypothetical tasks with no recall period. The answers do not form an observed sequence from a cart to a completed purchase.

A separate implementation reproduced the prespecified aggregates from the original released CSV. The download includes the complete table, source controls, definitions and code, with attribution and the original-file hash. It contains no respondent records or identifiers. Product.ai’s named research artifacts are licensed CC BY 4.0; the calculations, interpretation and translations are by Influence IQ, without implied endorsement. ^[6]