AI agents and grocery shopping: how much time do they really save?

In a trial reported by Clubic, ChatGPT filled a basket with no errors reported, but took thirty minutes and required several interventions. What this result does — and does not — tell us about the time actually saved.

Delegating the filling of a basket to an AI agent can free up time without speeding up the order. In a trial published by Clubic, ChatGPT found the requested products and added the specified quantities to an online grocery shopping site. But preparing the basket took around thirty minutes, permissions interrupted the process and the user still had to confirm the order. The result describes a specific task, not unsupervised automation.

A filled basket, with a list already prepared

The trial reported by Clubic involved transferring a shopping list from one retailer to another, then adding a few products. The brands and quantities had already been specified. To guide the agent, the journalist provided it with a file, seven screenshots and links to the products he was looking for. Preparing these materials took him less than five minutes.

This step is far from incidental. A usable list cuts down on searching and leaves the agent with fewer decisions to make. Clubic estimates that writing the list, specifying the brands and quantities, and checking what was missing would have taken at least a quarter of an hour instead. This is the journalist’s estimate, not a controlled measurement; it nevertheless shows why the result also depends on the work done before the AI gets involved.

The instructions were explicit: do not substitute a product, avoid duplicates and keep the items already in the basket. The agent could browse and add products, but not confirm the order or pay. On its first pass, it left some items out because they were unavailable or there was no exact match. The selected store was also not the one intended for the next stage; the user had to specify this before continuing.

At the end of the trial, Clubic reports that no errors were found in the completed basket: the products and quantities matched the request. This is a useful result, but one limited to a given list and a particular process. It does not allow us to conclude that the agent will always find the right product, that the displayed stock is up to date or that all sites will work in the same way.

Elapsed time is not time spent

For the entire operation, including setup, the journalist estimates the elapsed time at around thirty minutes, from writing his request to confirming the order. He estimates that he would have spent about twenty minutes filling a comparable basket himself. Judged simply by the clock, then, the trial shows no time saved: the automated process took longer.

The difference lies in what was happening during those minutes. The journalist was able to carry on working while the browser moved through the pages. The total time before a basket is ready and the time spent clicking through it oneself do not measure the same cost. An agent may be useful if the aim is to keep your hands free for another activity, even when the task does not finish any faster.

But the time was not entirely freed up. The user had to authorise access to the account, various pages and control of the site, then respond to further prompts. These approvals have a practical cost: the user has to remain available and keep track of what the agent is doing. They also serve as a check, especially when the system is operating in a personal account. In this trial, placing and paying for the order remained outside its remit.

One trial is no guarantee

In its launch post for ChatGPT agent, published on 17 July 2025, OpenAI described a system capable of using a virtual computer to perform tasks and requiring authorisation before certain important actions. The company also said that users could interrupt the work or take back control of the browser. This document explains the principle announced by OpenAI; it is not an independent assessment of the reliability of a purchase, nor an up-to-date description of all access conditions. The page itself notes that its content is outdated for current conditions.

The Clubic trial provides a different kind of information: an account of a specific use by a journalist. It is not a benchmark — a repeated evaluation carried out according to a defined protocol — and the available sources do not provide a series of orders that would allow an error rate to be measured. A correct basket in this trial therefore cannot be used to calculate the probability of success on a future order.

The distinction matters because the agent does not simply produce a written response. It acts in a browser and depends on the pages, accounts and permissions available. Poorly prepared instructions, an unavailable product, an ambiguous match or a changed interface can complicate the process. Human approval still has a role: the basket has to be checked and the decision to order made.

The value also depends on the tools already available

For repeat purchases, saved lists, order history and a retailer’s suggested regular products may be more straightforward. Clubic points out that the trial had a particular feature: products had to be transferred from one store to another, something the saved preferences in the first account could not do automatically. The initial search may therefore be more useful than repeating a familiar basket.

Once the products have been found and saved with the new store, the benefit of using an agent again becomes less obvious. For a small list that is already prepared, automation can avoid a series of searches while the user gets on with another task. For a meal requiring dozens of products, the journalist says he would go straight to the online shop and be quicker doing it himself.

Clubic also reports that the operation used nearly half of the journalist’s ChatGPT Plus quota over a five-hour window. This is his observation during this trial, not a general official limit or one that is necessarily still valid today. It nevertheless highlights a dependency to include in the calculation: an agent may use up a subscription resource that is also needed for other purposes.

So the right question is not just whether AI can fill a basket. In this specific case, it did so correctly after being given instructions, switching stores and receiving several permissions. To decide whether the service is useful, you also need to compare the time actually freed up, the supervision still required, the tools already offered by the retailer and the cost of access. The basket can be ready without the user ceasing to be responsible for it.