In Total Retail last week, Kathleen Peters, chief innovation officer at Experian, made a plain argument about AI shopping agents: the technology is ready and the trust is not. People will ask an AI which running shoe to buy. Far fewer will hand it the card. She expects the gap to close the way it closed for mail-order catalogs and mobile payments, one safe transaction at a time.
Her piece is written for the retailers, payment networks, and banks that have to build the guardrails. Read it from the owner's side of the counter, whether that owner runs a restaurant, a clinic, or a regional retailer, and the question changes. When an agent shops on a customer's behalf, what does it see when it looks at you?
Peters expects the first real agentic purchases to be bounded tasks. Her example is an instruction like "If this item comes back in stock for less than $100 and arrives before Dec. 15, buy it." That instruction contains a price, a date, and an availability check. The agent will read your price, your stock, your delivery window, your return policy, and what other people have said about you, and then it will decide.
That is a different kind of customer. It skips a page that hides the price, and a good host cannot win it at the door. It compares facts, and it compares them across everyone in the category at once. Most of what businesses spend on marketing was designed to move a person. Very little of it was designed to be legible to a machine acting for one.
Being legible to an agent is unglamorous work, and most of it is within reach:
- The same name, address, hours, and phone number on every listing, including the ones you did not know you had.
- Prices in text on the page, where software can read them. A photo of a menu or a price behind a form does not count.
- Structured data, the machine-readable labels behind a page that say this is a product, this is its price, this is when it ships.
- A returns policy written in sentences on a page, rather than a PDF.
- Reviews that arrive steadily, rather than in a burst after a campaign.
An agent checks all of it before it lets a human's money move.
Peters also names the retailer's side of the problem. Automated traffic used to mean fraud, so businesses blocked all of it. Now some of that traffic is a real customer's proxy, and she puts the question directly: "How do we know an AI agent is acting on behalf of a real customer?" The payment networks and identity companies will settle that one, and it is worth watching. Until then, find out what an unclassified visitor is before you block it.
For teams with a dashboard, a measurement problem is coming. An agent-mediated purchase will not look like a session. It may arrive as one API call and a payment. A reporting model that only counts humans who clicked will undercount this channel while it is small, then misread it once it is not. Agree now on how you will recognize these purchases, so the first ones do not disappear into "direct."
When we run a Clarity Diagnostic, the AI-answer audit shows a business what ChatGPT and Google's AI already say about its category locally, screenshots included. The agent question is the same audit taken one step further. Could a machine with a budget and a deadline complete a purchase from you without a human filling in the gaps?
Peters closes on history: catalogs and mobile payments were distrusted until millions of safe transactions made them ordinary. Agents will follow the same path. Start with the five facts an agent will check: price, stock, delivery, returns, reviews. Spend on how you sound after those are right.