Agentic commerce: preparing for agents that buy on behalf of customers
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Agentic marketing has mostly meant agents working for the brand: drafting, targeting, and following up. A second kind of agent is appearing on the other side of the transaction, one that works for the shopper and may complete the purchase itself.
AI and payments companies have been building ways for assistants to check out and authorize payments for a person, and some retailers are experimenting with them. We do not know how quickly shoppers will hand over final purchases, and many will want to stay in control for a long time. The direction is clear enough that marketers should start preparing now.
What changes when an agent shops
A person shopping online scans images, reads a headline, skims reviews, and decides partly on feel. An agent shopping for that person works from a brief. It reads structured data, policies, and review text, and it checks each option against the constraints it was given.
That shifts four things for marketers.
The audience includes software. An agent comparing products is not moved by visual design or a well-told brand story in the way a person is. It needs facts it can read and compare.
Constraints are explicit. A shopper might tell an assistant to find running shoes under a set budget, in stock in a given size, delivered by Friday, with free returns. A product that does not clearly meet each of those conditions may be dropped, even if the person would have liked it.
Trust depends on consistency. An agent acting for a customer needs confidence that the price, availability, and policies it reads are the ones that will apply at checkout. Conflicting information is a reason to choose another seller.
Brand still matters, upstream. People often ask agents for a specific brand, or for brands they have heard good things about. Preference formed through other channels shapes what the agent is asked to find in the first place.
How an agent decides, step by step
It helps to picture the decision as a pipeline. We think most purchasing agents will work through some version of these stages, though the details will vary by product and platform.
- Interpret the request into a set of constraints: category, budget, size, timing, and policy needs.
- Gather candidates from feeds, marketplaces, retailer pages, and prior purchases.
- Filter out anything that fails a hard constraint or cannot be confirmed.
- Compare the remaining options on the attributes the shopper cares about.
- Act or confirm, either completing the purchase or presenting a short list for approval.
Most of the risk for a brand sits in step three. A product is not excluded because it is worse. It is excluded because the agent could not confirm that it fits. Missing delivery estimates, vague return terms, and size charts that live only in an image are all ways to fail a filter you never knew you were in.
A worked example
Consider a hypothetical outdoor retailer selling a waterproof jacket. A shopper asks an assistant for a waterproof jacket under $200, in a medium, delivered within four days, with free returns.
The retailer's product page says the jacket is waterproof and shows the price. The size chart is an image. Delivery times appear only at checkout, after an address is entered. The returns policy says "free returns on most items" and links to a separate page with exceptions.
To a person, this jacket might look like a good fit. To an agent, two of five constraints cannot be confirmed from readable data, and a third is ambiguous. A competitor whose feed states sizes, a delivery window by region, and a plain returns policy is easier to choose.
Nothing in the fix is exotic. Put sizes in structured fields, publish delivery estimates in the feed, and state the returns policy in one plain sentence with any exceptions listed by category.
How to prepare
The preparation is mostly data work, with a few conversations across teams. This is the sequence we recommend.
- Make product data complete and accurate. Prices, availability, shipping times, and attributes should be current and match everywhere they appear. See Product Feeds Are the New Landing Pages.
- State policies plainly. Returns, warranties, and delivery commitments should be easy to find and unambiguous, with exceptions spelled out.
- Support structured data. Machine-readable product and offer information helps agents interpret your catalog. See Schema, Structured Data, and Machine-Readable Brands.
- Ask your platforms. Work with your commerce platform and payment providers to learn which agent checkout approaches they support and what each requires from you.
- Bring in fraud and payments early. Agent-initiated transactions raise new questions about authentication, authorization, and disputes.
Where purchases touch consumer protection, disclosure, or payment rules, confirm the specifics with counsel before you change checkout behavior.
Failure modes to test for
Agentic checkout can fail in ways a human shopper would catch. We suggest testing a handful of scenarios before agents become a meaningful share of orders.
- Price drift: the feed shows one price and checkout shows another after a promotion ends.
- Promotion mismatch: a discount applies only through a code or a banner the agent never sees.
- Phantom stock: an item reads as available in the feed but is out of stock in the chosen size.
- Hidden conditions: free shipping or free returns depend on a membership the agent cannot detect.
- Duplicate orders: a retry after a slow response places a second order.
The last one is a familiar pattern from agent design, covered in Tool Errors Are Marketing Errors. The difference here is that the agent placing the order belongs to someone else, so your systems need to recognize and reject duplicates on their side.
What to measure
Agent-driven purchases may be hard to see directly. Some will arrive through new checkout paths, and some will look like ordinary orders.
We recommend starting with what you control. Track the share of your top products that have complete, consistent data for price, availability, delivery, and returns. Track how often feed values disagree with what your site shows. Where your platform marks orders as agent-initiated, track their volume, return rate, and dispute rate separately from other orders.
Attribution gets harder when an agent sits between you and the customer. Our post on attribution in an agentic funnel covers how decision logs help on your own side of the transaction.
Questions for marketing leaders
Several strategic questions do not have settled answers yet. We think they are worth putting on a planning agenda now.
- Relationship: if an agent buys without visiting your site, how do you build a relationship with the customer afterward?
- Promotions: which offers and loyalty benefits can an agent understand and apply on a shopper's behalf?
- Measurement: how will you account for sales influenced by agents you cannot observe?
The question we are studying next is how purchasing agents handle conflicting information across sources. Our working hypothesis is that consistency carries more weight than any single persuasive detail, and we plan to test it with controlled product data.
How to apply this: pick 10 best-selling products and check whether an agent could confirm price, availability, delivery time, and return policy for each from structured, consistent data alone. Fix whatever is missing, then test one agent-style order end to end.
Related research
Measuring share of answer: tracking brand presence in AI responses
AI answers vary from one request to the next, so measuring brand presence takes repeated sampling. Here is a practical method for tracking share of answer.
Machine-readable brands: why structured data is part of being findable
Structured data tells machines what your pages describe. As more discovery runs through AI systems, being machine-readable becomes part of being findable.
Product feeds are the new landing pages for shopping agents
AI shopping assistants lean on structured product data more than persuasive pages. Feed quality becomes a marketing priority, not only an operations task.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
