AI and ecommerce represent the twin digital booms of the past decade โ so it was perhaps inevitable that they would collide in the form of agentic commerce. Amazonโs โBuy for Meโ is already completing transactions on behalf of consumers, while over 25% of consumers already feel comfortable allowing agents to shop for them. The era of agentic commerce, where AI handles the entire shopping journey, is arriving fast.
At ChargebackX 2025, Justt CEO and Co-founder, Ofir Tahor moderated a live panel with Jamie George (VP of Account Management and Partnerships at Ravelin) and Shahar Tal (CTO at Justt), on the practical realities of agentic payments. Their conversation produced a clear message: while agentic technology promises convenience and efficiency, it also introduces complex challenges around fraud, liability, and chargeback management that merchants must address sooner, rather than later.ย
Watch the full webinar here, or read our highlights and key takeaways below:
The Current State of Agentic Commerce: Still WIP, but Evolving Fastย
Agentic AI refers to intelligent software capable of making decisions, interacting with digital environments, and taking autonomous actions. In ecommerce, this means AI agents can navigate websites, compare prices, and complete purchases based on simple instructions โ saving consumers huge amounts of time while opening vast new commerce channels.
Despite widespread enthusiasm, current implementations reveal deployment roadblocks. During the webinar, Tal shared his experience using ChatGPT’s agentic mode to order fries from McDonald’s. The process took over 40 minutes, while the agent navigated captchas, switched between delivery platforms, and eventually succeeded. Similarly, George attempted booking hotels in Manhattan for his team, spending 45 minutes only to be offered a motel in Jersey City.
While AI is progressing at breakneck pace, itโs likely that online shopping infrastructure will also need to evolve to better support agentic transactions. Current storefronts are optimized for human conversion, not AI comprehension. “Storefronts were actually built to maximize conversion rates by taking all the fine print away from the funnel,” Tal explained. “But when you think about AI agents and LLMs, they need context in order to succeed, and this is something many storefronts today simply do not have.”
Major industry players are moving fast. While individual experiences remain clunky, tech giants are racing to establish standards and platforms. Google’s Agent Payment Protocol (AP2) attempts to solve one of agentic commerce’s biggest challenges: proving customer intent. The protocol provides frameworks for sharing purchase mandates through smart contracts, allowing merchants to demonstrate that customers authorized specific transactions.ย
Similarly, Amazon’s Buy for Me feature (currently in beta) keeps customers within Amazon’s ecosystem while enabling an AI agent to purchase products theyโre looking for, and which arenโt available on Amazon, from external merchants. The strategy maintains Amazon’s market dominance while expanding product availability. From a risk perspective, merchants participating in the program retain liability under their own terms and conditions โ even though they receive minimal customer data and Amazon controls the entire transaction flow.
More Fraud is Coming
George’s prediction was unequivocal: “I am willing to place a bet with everyone here that fraud goes up for agentic commerce. I’ll give good odds on it.” Nobody took him up on these odds โ the panel agreed that the rewards of agentic technology are matched by huge risks of true and friendly fraud. The reasons are numerous, including:ย
- Fraud signal disruption: Traditional fraud signals like device fingerprints, IP geolocation, and behavioral analytics become unreliable when agents operate from virtual devices with cloud-based locations. This means that, among other counterfraud programs, Visa’s Compelling Evidence 3.0 requirements, which rely on matching IP addresses and device IDs to prove legitimate transactions, will not work in agentic environments.
- Agentic malware: As George noted, โThe first people to adopt a new technology are always fraudsters. In a month, you will probably be able to go on the dark web and buy a bad agentic modelโ. These malware agents will likely be used to intercept or corrupt benign models, and even to carry out hyperintelligent, large-scale cyberattacks.
- Agentic Honeytraps: Bad actors may design scam websites to attract agentic ecommerce. These scam sites could potentially steal vast amounts of data and money from unsuspecting users.
- ATO Fraud: Account takeover fraud has long been a hacker mainstay, but it may soon become far more lucrative, as attackers could direct agents to conduct high-speed transactions in multiple locations at once.ย
- Agents of friendly fraud chaos: Perhaps most pressingly, refund abuse and service-not-as-described friendly fraud chargebacks are likely to surge. While friendly fraud chargebacks already comprise 75% of all disputes, the added layer of abstraction in agentic commerce will provide bad actors with lucrative new ways of disputing legitimate transactions, causing this figure to rise.ย
The Liability Question: Who Will Bear the Burden?
The question remains: when AI agents order the wrong colors, book incorrect hotels, or become tools for fraud, who will foot the bill? Unfortunately, itโs likely to be the merchant.ย
“The card schemes aren’t going to take liabilityโ, remarked George. โNeither are the customers, their issuers, or the AI model, who technically hasn’t made any money. That doesn’t leave many options โ itโs going to be the merchant.”
The problem compounds when considering agent-to-agent transactions, which George described as “the most obvious case for Chinese whispers.” As requests pass through multiple AI interpreters, errors multiply, increasing the potential cost of liability for fulfillment failures.
This liability shift represents a paradigm change in the customer-merchant relationship. This is because the cardholder doesnโt see the merchant behind the transaction โ as far as theyโre concerned, the agent has โbugged outโ. As Tal explained, โthis makes it mentally easier to file chargebacks. We’re definitely going to see friendly fraud rise.”
Preparing for Agentic Commerce: Data and Automation Are Essential
The panel agreed that comprehensive data collection is the most effective defensive strategy at this stage. George noted that “going forward, making sure you get as much data as possible out of these agentic flows is going to be absolutely key.” Merchants should capture:
- Order context: items, prices, fulfillment timelines
- Execution signals: timestamps, session tokens, API headers, agent IDs
- Account-level information: customer history, login consistency, payment patterns
“It’s really a game of data,” Tal stressed. “Merchants need to choose partners that are really technological and can cope with the change. If you work manually or with outdated products that don’t have the ability to collect this data at scale, you need to rethink ASAP.”
Beyond data collection, merchants must embrace automation. Manual chargeback management systems cannot scale to handle the complexity and volume that agentic commerce will generate. “People would just not be able to cope with it,” Tal said. “AI created the problem, but AI also gave us tools to fight it.”ย
As agentic commerce continues its inevitable rise, merchants should look to smart chargeback management systems that can automate evidence collection, representment creation, and submission โ taking charge of this burden before it becomes overwhelming. However, merchants must also ensure that their chosen solution is adaptable enough to respond to the novel dispute scenarios that agentic commerce will generate. โMachine learning capabilities have never been so indispensableโ, noted Tal.ย
Read our previous blog post to learn more about agentic ecommerce and chargeback risk.
Timeline: Sooner Than You Think
While current agentic transaction volumes remain minimal, “fractions of a percent,” according to George, the trajectory suggests rapid acceleration. A 30% quarterly increase in AI tool usage, combined with growing consumer comfort with agent-driven purchases, indicates that widespread adoption could arrive within months rather than years.
“I don’t know how soon it will happen,” Tal concluded, “but I know that when it happens, it will happen very fast. Merchants really need to start to prepare for it immediately.”
George echoed this advice for staying current: “Keep close to all of the changes that are happening and be as dynamic as possible. If you think you’ve got your whole strategy nailed today, you’re probably going to be behind the curve.”
