Artificial intelligence is playing an increasingly prominent role in shopping. For now, it helps customers search for products, compare prices and analyse reviews. The next stage will involve placing orders and making payments independently, and later also handling returns and complaints.
For online retail, this signifies a shift in one of its fundamental principles: the customer at the checkout will not always be a human being.
According to available estimates, between 30 and 45 per cent of US consumers already use generative AI when researching and comparing products. The transition from recommendations to independent purchasing will not be immediate, but the potential scale of the market is significant. By 2030, purchases made by agents could account for between $190 billion and $385 billion of US e-commerce spending.
A new type of customer
An AI agent can act on the user’s behalf: find a product that meets specific criteria, check availability, compare delivery terms and place an order within a set budget. Over time, these systems may also take on more complex tasks, such as amending an order, contacting customer service or initiating a return.
The problem is that most current e-commerce platforms were not designed with such transactions in mind. For years, security systems have been analysing how users navigate the site, their behaviour history, device, location and the subsequent stages of the shopping journey.
However, a bot does not need to use the shop in the same way as a human. It can connect directly via an API, perform actions more quickly and leave no typical browsing history. From the perspective of traditional anti-fraud tools, it may therefore resemble a malicious bot.
Retail is losing its existing trust signals
AI agents do not need to increase the number of fraud cases to undermine the current security model. It is enough for them to cease conforming to the patterns on which risk detection systems are based.
Until now, the fundamental question has been: is the user genuine? In agent-powered commerce, the questions will need to be framed differently: is the agent in question trustworthy, are they acting on behalf of a specific person, and do they have authorisation to carry out this transaction?
The data shows that this shift has already begun. By the end of 2025, the number of orders generated via prompts from large language models had increased by over 1,000 per cent year-on-year. At the same time, purchases made directly by bots still accounted for less than 1 per cent of all orders.
The scale therefore remains small, but the rate of growth is a warning sign. Ill-prepared shops may start rejecting legitimate transactions simply because the agent’s behaviour differs from that of a human.
A greater risk may be the loss of sales
In the debate about AI agents, the fear of fraud is most commonly raised. From a retailer’s perspective, however, automatically blocking valuable customers could be an equally serious problem.
If the system classifies an authorised agent as a dangerous bot, the shop will lose more than just a single order. It may also be excluded from the AI’s recommendation algorithm. An agent that encounters problems with data availability, payment or transaction processing on several occasions is likely to start choosing rival platforms.
Companies that provide well-organised and machine-readable information on prices, stock levels, delivery, payments and returns may therefore gain a competitive edge. It will be equally important to establish mechanisms to verify the agent’s identity and the scope of the authorisation granted to them.
The system should recognise whether an agent is authorised to purchase specific product categories, what their spending limit is, and whether their actions have been approved by the account holder. Security cannot end at the point of payment; it must also cover subsequent changes to the order, contact with customer service and returns.
Infrastructure will become a competitive factor
Agent-based commerce is likely to begin by handling simple, routine purchases. Everyday items, basic electronics and repeat orders do not require strong emotional involvement, so it will be easier to hand these over to automated systems.
In more complex categories, humans will remain the primary decision-makers for a long time to come. However, this does not alter the direction of the market’s development.
Retailers must prepare for an environment in which some customers will interact with the shop exclusively via software. The most important challenge will not be recognising the machine itself, but verifying that a real customer is behind it, that the decision is a conscious one, and that the authorisation is clearly defined.
Shopping may become increasingly automated. Trust, however, will still rest with humans.
