Agentic commerce—online shopping managed by AI agents—is poised to change the way we purchase nearly everything from toothpaste to car insurance. Online retailers, AI platforms and credit-card companies are all vying for a piece of the agentic commerce pie. According to McKinsey & Co., by 2030 we could see up to $1 trillion in orchestrated revenue in the retail market alone.
Although agentic commerce is still in its infancy, the technology is advancing quickly and issues are already rising around regulation and trust. As companies build out the mechanisms for seamless shopping, it is unclear how this will be regulated and how soon people will trust that they can hand over their shopping to AI agents.
Where Agentic Commerce Is Today
Agentic commerce is the next step in online shopping, connecting the search and recommendation capabilities of an LLM chatbot with the low-friction user experience of e-commerce. Consumers can ask an AI agent to re-up regularly used household items, select the best price, and negotiate deals—all with little human intervention.
But to prepare for agentic commerce, businesses will need to move beyond traditional e-commerce models and adopt a new way of doing business, says Bob Hedges, a Digital Fellow at the MIT Initiative on the Digital Economy (IDE) and former Chief Data Officer at Visa. While at Visa, Hedges contributed to a 2025 white paper, “Earning Consumer Trust in the Age of Agentic Commerce,” which examines the state of agentic commerce and the requirements for achieving consumer trust.
“There are formal rules today between banks, merchants and payment networks that outline who has the responsibility to ensure the system works efficiently and address any breakage in the system,” Hedges says. “The rules address what happens in the case of charge-backs, who has liability in a fraud scenario…Today, there is a lot of negotiating as to whether, and how, those rules should apply to AI agent environments.”
To that end, several alliances have been formed by payment service providers and AI agent platforms, including:
- In 2024 Perplexity announced a “Buy with Pro” feature with PayPal integration and access to over 5,000 merchants.
- Last year OpenAI announced an Agentic Commerce Protocol (ACP), co-developed with Stripe, that lets users complete purchases directly within the ChatGPT interface.
- Visa is positioning its Intelligent Commerce network as the backbone for agentic commerce, providing “AI-ready cards” with tokenized credentials to a wide range of AI platforms.
- Last year Mastercard launched Agent Pay in collaboration with Microsoft, also using tokenization capabilities for security.
Hedges expects other deals will continue to be announced as agentic shopping evolves.
“The companies that have the capabilities to make agentic commerce work are still figuring out the most appropriate business models,” he says.
While dramatic progress is being made on the tech front and important strategic partnerships and standard setting are also in motion, less clear is how the economics of AI agents will work. One important question is who the agent works for. To build trust and ensure consumer confidence in agentic commerce, having transparency around the issue of “who’s paying the agent” will become vital.
But the real challenge, Hedges says, may be how to make the business models work for all players. As tech solutions mature, addressing the risks and costs—as well as balancing the interests of consumers, retailers, payment networks and AI platforms—will become strategic.
How Do AI Agents Manage Consumer Shopping?
Asking an LLM to recommend or select the perfect product for a consumer is straightforward. But once an item is found, AI agents also require infrastructure and integrations to access the online retailers’ shopping interface, make a selection, and issue payment securely. There’s also the challenge of authenticating whether the AI agent has permission to act on behalf of a particular consumer.
To accomplish all this, agentic commerce is being built around four key components:
Model Context Protocol (MCP)
Created by Anthropic and donated to the Linux Foundation’s Agentic AI Foundation in December 2025, MCP standardizes how an AI agent connects to external tools, data sources and services. In this way, MCP solves “bot sprawl.” Without MCP, retailers would need to build and maintain separate AI systems for every function. Instead, MCP allows a single agent to query multiple backend databases through a unified interface.
For example, consider Walmart’s customer-facing agent handling this request: “Plan a unicorn-themed birthday party for eight kids.” Thanks to MCP, the agent could pull data simultaneously from the party supplies, bakery and toy departments.
Agent-to-Agent (A2A) Protocol
Launched by Google in April 2025 and maintained by the Linux Foundation under Apache 2.0, A2A manages the lifecycle of a request with a three-step sequence:
- Discovery: Finding capable agents
- Delegation: Assigning the task
- Lifecycle updates: Tracking progress to completion
Where MCP handles the relationship between an agent and its tools, A2A provides the communication layer for multi-agent workflows, enabling complex orchestration.
Agent Payments Protocol (AP2)
This protocol specifically addresses how AI agents can safely participate in commercial transactions. It builds trust through three core mandate types: intent, cart and payment. These mandates are digitally signed statements that define what an agent is allowed to do, such as create a cart, complete a purchase, or manage a subscription. They are portable, verifiable and revocable, allowing multiple stakeholders to coordinate safely.
Tokenization
This approach makes it safe to let an agent spend on your behalf. Rather than giving an AI agent access to a live credit card, the payment method is represented as a tokenized credential. This is a single, specific payment method selected by the credentials provider and confirmed by the user. It’s accompanied by a risk payload containing signals required by merchants, payment processors and issuers.
Who Owns My Shopping Agent?
How will AI shopping agents disrupt certain industries? A recent study by the IDE’s AI, Marketplaces and Labor Markets research group theorizes that AI shopping will lower transaction costs to almost nothing. In their paper, The Coasean Singularity? Demand, Supply, and Market Design with AI Agents, in which the authors argue that agent ownership will have a big influence on how agentic commerce plays out.
Under one model—known as the “bowling-shoe” agent—the platform or online retailer builds and maintains the agent conducting the transaction. This type of agent is easy to use and is optimized for a particular environment. But like rented bowling shoes, consumers don’t own it. The paper’s authors argue this creates an incentive to design the agent in a way that benefits the seller or online retailer.
But a “bring-your-own” agent flips that dynamic. User-controlled and portable across platforms, it’s configured around a consumer’s preferences rather than a vendor’s goals, so it works for the consumer regardless of where they’re shopping. However, the paper’s authors argue, a BYO agent may also present consumers with management and configuration challenges, and it may not be compatible with all retailers. There are also the important issues of who pays for the agent and whether consumers will pay directly for an agent’s advice.
How the economics of “consumer-owned” agents works remains to be determined. A wide range of players—including fintech start-ups, consumer advocacy groups, and organizations like Consumer Reports—are actively exploring options.
At the moment, AI platform-run agents are leading the way by borrowing a business model from digital advertising and search. However, Hedges warns, there is a risk that the tempting economics of advertising and “product placements” will undermine these agents’ objectivity. Given current market dynamics, there is little incentive for the AI platforms to provide transparency on how and why specific products are recommended.
“Without a requirement of transparency to disclose what influences an agents’ product recommendations, AI platform agents can accept compensation for advertising without the obligation to tell consumers what’s driving their recommendation,” Hedges says. “The economics of digital advertising are highly attractive, and the short-term temptation is great, particularly to AI platforms prioritizing revenue growth. But there is great risk to a platform’s objectivity being compromised and damaged. Consumer trust, once compromised, is hard to repair.”
Challenges to Adopting Agentic Commerce
What’s blocking more complete adoption of shopping by AI agents? Mainly, four serious challenges:
Trust
According to the 2025 “Earning Consumer Trust” report from Visa, nearly eight in 10 consumers (79%) express concerns about data privacy with AI commerce—and that’s among those who are already interested in it. Trust also varies around the world in both context and cultures. Until providers can demonstrate their agents are acting in a user’s genuine interest, adoption will remain uneven.
Accountability gaps
When an AI agent makes a poor decision, determining accountability is complex. Who’s to blame for that faulty transaction? The platform that developed the model? The brand that deployed the agent? The user who approved it? Currently, there is no global consensus. Until clear legal frameworks emerge, this liability gap creates both reputational and regulatory exposure for every player.
Systemic risk
Autonomous agents could eventually make dozens of decisions a day for a single consumer. But at that scale, decision-making introduces systemic risk: When agents are interconnected across multiple systems, minor errors can have exponential impact. A single faulty prompt can trigger a cascade of unintended consequences.
Legacy payments infrastructure
As the Visa report notes, the payments stack—gateways, fraud engines, KYC/AML systems—was designed to have humans in the loop. But agentic commerce disrupts this paradigm: The customer is now an AI agent acting on behalf of a person. This necessitates a new approach for delegating authorization, setting programmable spend policies, and attesting consent. This shift will also require the system to verify not only users, but also agents. For many companies, this will require a major infrastructure overhaul.
Will People Let AI Agents Do Their Shopping?
Companies that aim to lead on agentic commerce must build and launch quickly to remain competitive. This leaves them limited time to earn trust among consumers, especially those wary of agentic commerce.
“Building and maintaining consumer trust requires transparency, explainable processes, and some degree of consumer control,” explains Hedges. “Consumers want objective and appropriate recommendations and guidance.”
This means AI platforms now have a strategic choice to make: Do they want to serve consumers, or do they want to monetize them?
“In the long run, investing to build enduring consumer trust is the right strategic choice,” Hedges says. “We’ll learn which players have the strategic discipline to serve the long-run interests of consumers.”