AI Agents Can Now Pay for Things — Should We Let Them?

The infrastructure that allows artificial intelligence agents to spend real money has arrived faster than most experts predicted. In just a few weeks during mid-2026, Visa and Mastercard both opened their global payment networks to AI systems, a major cryptocurrency exchange launched a marketplace where software agents hire and pay one another without human involvement, and one of America’s largest banks publicly questioned whether ordinary consumers are ready for any of it.

Morgan Stanley estimates that agentic shoppers — AI systems acting on behalf of people — could drive between $190 billion and $385 billion of U.S. e-commerce by 2030. The industry is racing to build the rails for that future even as basic questions of trust, liability, and control remain unresolved. The technology is no longer theoretical. The harder question is whether society should fully embrace it.

Visa and Mastercard Open the Payment Rails

In June 2026, Visa announced a partnership with OpenAI that embeds its payment network directly into ChatGPT. Users can link a Visa card and set permissions such as spending limits or approval requirements. The AI agent can then research products, compare prices, and complete purchases on the user’s behalf rather than merely recommending items. The system relies on Visa’s established tokenization and fraud-monitoring infrastructure to keep transactions secure.

On the same day, Mastercard launched Agent Pay for Machines. This service is designed for high-speed, continuous transactions between AI agents themselves, including micropayments worth fractions of a cent. Human-granted permissions are recorded on public blockchains such as Polygon, Solana, and Base. Settlement can occur across traditional cards, bank accounts, or regulated stablecoins. More than 30 companies, spanning conventional payment processors and crypto infrastructure providers, signed on at launch. Mastercard described the move as creating conditions for a “superbloom of AI business models.”

Two networks that together handle the majority of the world’s card transactions acted independently on the same day. Agent payments moved from experimental pilots into production-grade infrastructure almost overnight.

When Agents Pay Other Agents

While the card networks focused on enabling agents to buy from merchants, the crypto industry pushed further. Crypto exchange OKX launched a marketplace in which AI agents can hire one another, settle payments autonomously in stablecoins, and accumulate portable on-chain reputations. In this model, one agent might commission another to generate a market report, clean a dataset, design advertising variations, or monitor wallets for suspicious activity. Complex jobs can use escrow arrangements, and successful completions build verifiable reputation records.

This vision of commerce with no human on either side of the transaction is radical. It treats software not merely as a tool but as an economic participant. Whether such agent-to-agent marketplaces become mainstream remains uncertain, yet they demonstrate what an economy optimized for machines rather than people could look like.

The Consumer Trust Gap

Despite the rapid technical progress, senior banking executives remain cautious. Marianne Lake, former CEO of JPMorgan Chase’s Consumer and Community Banking business, observed that while consumers have readily adopted AI for product search and comparison, they have not yet shown the same willingness to hand over the actual payment step.

“I don’t think people are going to delegate their purchasing to agents just yet,” Lake said. When money moves, trust and security matter more. Most people are comfortable asking an AI to suggest three holiday itineraries. Far fewer are ready to let that same system book $2,000 flights and hotels while they sit in a meeting.

The distinction is emotional as much as technical. An AI that offers advice feels helpful. An AI that spends money feels like a loss of control. If the agent books the wrong destination, purchases counterfeit goods, renews an unwanted subscription, or sends funds to a fraudulent merchant, the central question becomes who bears the financial loss. Lake argued that the industry still needs clear frameworks: humans remaining in the loop for significant decisions, full transparency about what agents are doing, and reliable liability rules when mistakes occur.

Potential Benefits and Practical Use Cases

Supporters of agentic payments point to clear efficiency gains. Consumers could instruct an agent to monitor prices for a specific product and complete the purchase the moment it drops below a set threshold. Travel planning could become a single instruction: find the optimal combination of flights, hotels, and activities within a budget, then book everything. Businesses could use agents to restock inventory automatically when levels fall, negotiate better rates in real time, or handle recurring low-value transactions that currently consume human time.

At the machine-to-machine level, the possibilities expand further. An AI managing a logistics route could pay for temporary cold-chain monitoring data, loading-bay access, and warehouse handling fees as a shipment moves. Energy agents could switch electricity providers in response to real-time price signals and settle the exact amount consumed. Developers could let agents purchase API credits or compute resources mid-workflow without constant human oversight.

For merchants, the shift could mean higher conversion rates. An agent that has already researched options and received permission is more likely to complete a purchase than a human browsing multiple sites. Payment networks see an opportunity to capture a new category of high-volume, low-value transactions that traditional systems handle poorly.

Risks, Liability, and the Control Problem

The risks are equally concrete. AI systems still make errors, and the consequences of a mistaken payment are immediate. Fraud models designed for human behavior may struggle against agents that operate at machine speed and volume. Attackers could target the instructions given to agents rather than the payment credentials themselves, creating new forms of social engineering.

Liability remains murky. Existing consumer protection rules assume a human initiated the transaction. When an agent acts within broad permissions granted weeks earlier, responsibility for a bad outcome is less clear. Banks, card networks, AI platform providers, and merchants may all point fingers. Without clear rules, consumers may simply refuse to grant meaningful spending authority.

Privacy concerns also arise. Agents that shop effectively need detailed knowledge of a user’s preferences, budget, location, and habits. Concentrating that data in systems capable of autonomous spending creates attractive targets for both commercial exploitation and malicious actors.

There is also the broader societal question of whether accelerating consumption through tireless agents is desirable. Agents optimized to find the best deals and complete purchases quickly could increase overall spending, with consequences for household finances and environmental impact that are difficult to predict.

Should We Let Them?

The answer is not a simple yes or no. The technology itself is largely neutral; its value depends on the guardrails placed around it. Limited, permissioned agent payments with strict spending caps, mandatory human approval for larger amounts, transparent audit logs, and robust dispute resolution could deliver genuine convenience without surrendering control. Fully autonomous systems operating with broad discretion and weak accountability would introduce unacceptable risk for most consumers.

Regulators, banks, payment networks, and AI companies will need to collaborate on standards for identity, consent, liability, and dispute handling. Consumers will need clear interfaces that make permissions understandable and revocable. Merchants will need ways to distinguish legitimate agent transactions from fraud.

The infrastructure for AI agents to pay is already here. The harder work of deciding the rules under which they should be allowed to spend has only just begun. The companies that succeed will not simply be those that enable agents to transact. They will be the ones that convince people the process is safe, transparent, and ultimately under human control.

For now, most users are likely to keep agents in an advisory role — researching, comparing, and recommending — while retaining the final click themselves. That cautious approach may prove temporary. As the systems demonstrate reliability and the liability frameworks mature, the boundary between advice and action is likely to shift. Whether that shift improves everyday life or creates new forms of vulnerability will depend less on the cleverness of the algorithms and more on the wisdom of the rules we choose to impose.

About The Author