Skip to main content
Start your store

What an AI Agent Can Run in Your Store Today, and What Still Needs You

What AI agents already do in online shops, which jobs still need the seller's yes and why, and how FaStart is building Ally around approvals, receipts and autonomy that waits for evidence.

All articles

Enes Özkan, Founder and CEO7 min read

Example shop

In this article
  1. What can AI agents do in a store today?
  2. What still needs you?
  3. How we're building Ally around your yes
  4. What goes on a receipt?
  5. Where autonomy waits for evidence
  6. What can you see today?
  7. FAQ
  8. Sources

Today an AI agent can draft a lot of a store's routine work, from listings and replies to reorders, and the agents shoppers use can find products and buy them once the shopper confirms. What still needs you is anything that spends money, can't be undone or goes out of your store, because marketplaces, payment rules and the law hold the seller responsible for it.

Note: What follows about Ally, the agent inside FaStart, is how we're building it; the Ally page shows where each of its jobs stands today. The outside rules come from each company's or regulator's own page, read on 20 October 2026.

What can AI agents do in a store today?

Start with the shopper's side, where most of the public examples are. Google's agentic checkout tracks a price and, when it drops, can "buy it for you on the merchant's site using Google Pay," with a condition Google spells out: "We'll always ask for your permission first, and only buy after you've confirmed the purchase and shipping details" (Google). Google's Universal Cart, launched in the US on 15 October 2026, saves items from several stores, tracks price drops and checks out with Google Pay or hands the cart to the merchant's own site (Google). In ChatGPT, shoppers browse and compare products from merchants' feeds, and the purchase is completed on the merchant's site or app by default (ChatGPT). Our guide to UCP and ACP explains what each of these programs asks of a store.

On the seller's side the work is mostly drafting. Amazon says independent sellers created more than 12 million sales-ready listings with generative AI in 2025 (Amazon). Drafting is the safe end of the job: a draft that's wrong costs you a minute of editing. The risk starts when an agent's output leaves the draft folder.

What still needs you?

Every job where the seller carries the consequence. That's most of the jobs that matter:

JobWhat an agent can doWhat stays with you, and why
ListingsDraft titles, descriptions and attributesChecking the facts. OpenAI's rules for ChatGPT merchants say "Merchants are responsible for ensuring their products and content do not violate these restrictions or any applicable law"
Prices and offersSpot a price that drifted and prepare a change with exact amountsApproving it. Amazon's pricing policy applies to your prices whoever set them
Refunds and disputesGather the order, the photos and the messagesThe decision. The merchant of record handles refunds and chargebacks, OpenAI's commerce docs say
Replies to shoppersDraft an answer from your catalog and policiesWhat goes out. On Amazon, any agent you authorize must "clearly identify themselves as automated systems"
ReviewsAsk buyers after delivery, sum up what they saidNever writing them. The FTC's rule bans fake reviews, AI-generated ones included
StockFlag what's running low and draft the reorderPlacing the order. Under Google's UCP, "it is your responsibility to perform real-time inventory checks"
AdsPropose a budget inside the cap you setThe spend

When the FTC announced its crackdown on deceptive AI claims in September 2024, its chair said the cases "make clear that there is no AI exemption from the laws on the books" (FTC). Amazon's agreement says the same about accounts, making sellers "solely responsible for all activities" that use their API credentials, whoever carries them out (Amazon).

How we're building Ally around your yes

We've written one rule down before writing the code that will enforce it. The dashboard's autonomy screen already says it: "Whatever you choose here, anything involving money, anything that cannot be undone, and anything sent outside your store will always wait for you." It will hold at every level of autonomy a merchant can choose. Raising Ally's level will never let it spend, delete or message a shopper on its own.

Here's what that will look like, with an example. Petrakis, a deli in Sonoma (a sample shop, not a real one), gets a message: a tin of early-harvest olive oil arrived dented. Ally will draft a reply, prepare a $19 refund, and show the shopper's photo with it. The proposal will say what changes, the exact amount, and that a refund can't be taken back once it's sent. Then it will wait. Petrakis can approve, edit the reply or say no. If the amount changes after the approval, say because a second tin turns out to be dented, the plan will come back for a new yes.

Asking first is common: Google's checkout agent does it too, as quoted above. What we care about is what the ask contains, so that a yes means you saw the action, the amount and whether it can be undone.

What goes on a receipt?

Every job Ally runs will end in a receipt: a record of who did what and what came of it, kept with the job. It will list each step that ran, anything that didn't, the evidence from the service that did the work (the payment provider's reference for a refund, the carrier's for a label), and how to reverse it where that's possible. If a provider doesn't answer, the step will show as unknown until it's settled, never as done.

We want them because problems surface late. A shopper asks about a refund a month on, or a marketplace questions a price change, and someone has to say what happened and on whose say-so.

Where autonomy waits for evidence

Some jobs could reasonably run without asking every time. Alt text for product photos, a typo in a description, a missing size on a product page: all reversible, all inside your store. Our plan is that a job earns that freedom from its own record, one kind of work at a time. A month on the calendar, a bigger plan or a new release will never raise it by itself.

Say Petrakis approves every alt-text draft unchanged for a month. Ally could then suggest letting it write alt text without asking, showing that record and everything the change would open. Petrakis would decide, in the dashboard. The evidence can also go the other way: if Petrakis starts correcting or undoing what Ally did on its own, Ally will hold that work and ask again, saying why. Refunds, prices, messages to shoppers and anything at a marketplace stay with the merchant whatever the record says.

What can you see today?

On fastart.co every feature shows its own status, checked in code rather than typed by hand, and the trust page explains how. Two demonstrations run on sample products: Ally on your catalog, showing product records an AI assistant can read, and Ally answers your shoppers, showing a shopper answered from the catalog and handed to a person. The Ally page shows where each of Ally's jobs stands.

FAQ

Can an AI agent run my store on its own? It can prepare much of the routine work. Spending, permanent changes and anything sent to shoppers or marketplaces should wait for your yes, since under marketplace agreements and US law you answer for them.

Where can I see what Ally does today? On the Ally page, which shows where each of Ally's jobs stands, and on the product page, which does the same for the rest of FaStart.

Who is responsible when an agent makes a mistake? The seller. Amazon holds you responsible for everything done with your credentials, the merchant of record handles the refunds and chargebacks, and the FTC has said AI is no exemption from the law.

Sources

All read on 20 October 2026.

Ready to open your own store?

FaStart is the online store platform with an agent built in.

All articlesStart your store