Banks, artificial intelligence companies, and retailers are grappling with a tricky question as chatbots become more prevalent in online shopping: What happens when AI agents acting autonomously on a consumer’s behalf make a mistake or fall victim to fraud?

Shoppers are increasingly turning to AI agents—not only to spot deals on concert and airplane tickets, common household items, and other products, but also to make the purchases when the right deal comes around.

While that can make life easier for shoppers, the companies involved in the transactions are trying to figure out who’s on the hook when customers using AI agents believe they’ve been swindled or demand a refund.

Rules written for an offline era offer some clues for resolving any disputes, hinging on customers’ payment methods, the disclosures they receive, and how they authorize agents to make purchasing decisions. Regulators and courts will likely have to determine how to apply those standards to an AI-powered shopping ecosystem.

“We’re going to have to figure out how to serve this, and I suspect it’ll be things other than just disclosures in the future,” said Kelvin Chen, the head of policy at the Consumer Bankers Association.

It’s a high-stakes question for retailers and other companies offering agentic AI tools for shopping.

By 2030, around $500 billion—or 20%—of online commerce is likely to be carried out by AI agents based on a single prompt with customers’ price, brand, and other preferences, compared to less than 5% anticipated this year, according to a Bloomberg Intelligence report this month.

Another 6.2% of online shopping will be fully autonomous—in which agents independently decide what to buy and when—by 2030, the analysis found.

While online retailers such as Amazon.com Inc. and Shopify Inc. are early leaders developing agentic capabilities for search and checkout, traditional powerhouses such as Target Corp., Walmart Inc., and Kroger Co. are making headway, the report said.

Expected Growth

The expected growth of agentic AI tools is also driven by card networks such as Visa Inc. and Mastercard Inc. making the technology a part of their offerings, according to a separate June report from Bloomberg Intelligence.

For now, most shoppers will ultimately make the purchases themselves, particularly for products such as clothing where it’s difficult for a bot to assess sizing and visual preferences, Bloomberg analysts said.

Allowing humans to make the final purchasing decision avoids some of the liability questions that could arise when bots choose what to buy, whether fully independently or based on a customer prompt, said Carla Sanchez-Adams, a senior attorney with the National Consumer Law Center.

But with so many buying decisions likely to be made by chatbots in the coming years, the liability issues will have to be sorted out now, said Lindsey Johnson, the CBA’s president and CEO.

“We’ve done this before,” she said. “In the sense that we moved from cash to cards, we’ve now moved from cards to online payments.”

Payment Protections

In some ways, the liability questions will depend on the payment methods consumers use, regardless of the AI technology.

If the bot is connected to a person’s credit card, protections under the Truth in Lending Act and the Electronic Fund Transfer Act will still apply. So will chargeback and return standards set by banks and credit card networks, Sanchez-Adams said. Chargebacks are transaction reversals initiated by banks and card issuers.

Debit cards provide somewhat less protection, while stablecoins—which are often seen as a natural match for AI shopping chatbots—don’t allow for refunds or chargebacks, said Eric Goldberg, a partner at Davis Wright Tremaine LLP whose practice focuses on financial technology and consumer finance.

“I don’t know that consumers are going to want agents to go out buying stuff if they have limited protections,” he said.

It remains to be seen whether consumers will understand the different levels of protection. Policymakers may need to consider new disclosure and other regimes where chatbots inform people about the risks of different payment types, Chen said.

Final Say

Authorization is another thorny question posed by AI chatbots making purchases for consumers.

Target said in a March update to its terms of service that once a customer authorizes an agentic commerce agent to make decisions, any selections the chatbot makes will be treated as the customer’s.

Target has partnerships allowing shoppers to use Alphabet Inc.'s Google Gemini, Microsoft Corp.'s Copilot, and OpenAI’s ChatGPT on the Minneapolis-based retailer’s website.

Customers will have to get better at their prompts to prevent purchase mistakes, Goldberg said.

“Poor prompts lead to poor output, and you can see the same thing happening with shopping,” he said.

AI chatbots could be programmed to boost specific products or services based on manufacturer or service provider payments to developers, but customers would need to receive some sort of disclosure about those financial arrangements before they enter their prompts, Chen said.

Fraud is also likely to explode, he added, particularly when phantom retailers trick chatbots into submitting payments despite not having any inventory.

Even if policymakers, banks, AI companies, and retailers can craft strong consumer protections, shopping via AI agent poses new risks that may have to be resolved in court, Sanchez-Adams said.

But those developments will take time, she added.

“If things go wrong, and things will go wrong, there’s likely to be litigation until there is regulation,” Sanchez-Adams said.