When Machines Start Trading, AI Agents takeover finances

Something strange is happening in the digital economy. AI systems are starting to buy and sell things on their own, without anyone telling them to press “confirm purchase.” Google, Coinbase, and Ethereum are building the infrastructure for this change, and it’s creating a market worth billions that barely existed a year ago.

The numbers tell part of the story. The machine-to-machine market grew from $33.71 billion in 2024 to $36.52 billion in 2025, an 8.3% jump in just twelve months. But that’s just hardware talking to hardware. What we’re seeing now is different: AI agents that negotiate prices, complete transactions, and move money around while their human owners sleep.

The Problem Nobody Expected

Payment systems weren’t designed for this. Every time you buy something online, there’s friction built in: reviewing your cart, typing in card details, hitting “confirm.” That’s fine when humans are involved but it completely breaks down when an AI needs to make a hundred tiny purchases per second.

Google figured this out and launched something called the Agent Payments Protocol in September 2025. Sixty payment companies signed on immediately. AP2 basically creates a universal translator for AI systems to handle money. It proves that you, the human, actually approved the purchase, then lets your AI use whatever payment method makes sense. Credit card, bank transfer, crypto, whatever works.

Coinbase went a different direction. They dug up an old, forgotten internet status code called HTTP 402, which basically means “payment required.” Their x402 protocol lets AI agents pay for things directly over the internet using digital currency. No banks, no waiting periods, just instant transfers. This matters especially for tiny payments where traditional processing fees would cost more than the transaction itself.

Why This Gets Interesting Fast

Think about what happens when AI agents can handle their own money. A research assistant AI could pay for access to academic databases without bothering you. A smart home system could automatically buy extra electricity during peak hours when rates drop. An AI trading bot could execute thousands of tiny arbitrage opportunities that humans would never catch.

Ethereum’s positioning itself as the backbone for all this. The Ethereum Foundation created an entire team focused on payments and standards for AI agents. One developer, Davide Crapis, proposed ERC-8004, which is essentially a way for AI agents to find each other, verify they’re legit, and do business. Think of it as LinkedIn meets Venmo, but for robots.

The partnership between Coinbase and Cloudflare to launch the x402 Foundation shows how serious major tech companies are about this shift. They’re not building for the next product cycle. They’re building infrastructure meant to support AI agent transactions for decades.

The Problem Nobody Wants to Talk About

Stanford’s Digital Economy Lab ran experiments where they had AI agents negotiate with each other, representing buyers and sellers in simulated markets. The results should worry anyone excited about automated commerce: weaker AI agents lost up to 14% in profit compared to negotiations between equally capable agents.

An analysis of device-to-device economics highlights this exact concern: as AI agents become primary market participants, these performance gaps could create systematic unfairness that compounds across millions of automated transactions. If you’re using a basic AI assistant to negotiate on your behalf and someone else has a more sophisticated system, you’re starting every transaction at a 14% disadvantage. That gap compounds fast across hundreds or thousands of deals.

The Stanford researchers found something else troubling: current language models really struggle with the complex skills needed for reliable negotiation. Reading between the lines, gathering information, and understanding what the other side actually wants. These things humans do intuitively remain genuinely difficult for AI systems.

So you’ve got this weird situation where the infrastructure for AI agent payments is racing ahead, but the AI agents themselves might not be ready. You can have the perfect payment protocol, but if your AI is getting outsmarted on every deal, that protocol just helps you lose money faster.

What’s Actually Being Built Right Now

The managed services segment of the machine-to-machine market tells you how complicated this stuff really is. That market was worth $3.85 billion in 2023 and is projected to hit $17.69 billion by 2031, growing at 21% annually. Companies know they can’t just plug in an AI agent and let it loose. They need professional help to set up systems that don’t immediately start bleeding money.

Nansen announced they’ll roll out AI-powered trading functions by the end of 2025, which gives us a timeline for when these systems move from lab experiments to real products handling real money. Other companies are exploring AI agents that handle inventory monitoring, invoicing, and research. All the boring back-office tasks where automation makes obvious sense.

Financial services are particularly interested. AI agents that can interact with blockchain protocols, swap tokens, manage portfolios, and navigate decentralized finance platforms could handle tasks that currently require expensive human expertise. Whether they’ll actually outperform human traders is still up for debate.

Playing Nice Together

One encouraging development: these companies are building systems that work together rather than walled gardens. Google’s AP2 is compatible with existing standards like Agent-to-Agent communication and the Model Context Protocol. Coinbase’s x402 integrates with AP2. Meridian Finance is building payment rails that work across different blockchain networks.

This cooperation actually matters because agent economies won’t work if every AI system speaks its own language. An AI agent from Google needs to transact with one from OpenAI, which needs to work with specialized financial AI from some crypto startup. The alternative would be fragmented systems where agents can only do business within their own ecosystem, which would severely limit what’s possible.

The security requirements get hairy when you’re dealing with multiple platforms. AP2 creates traceable records for each transaction, which helps with accountability. Blockchain systems provide cryptographic security and permanent transaction records. But integrating traditional payment systems with cryptocurrency creates junction points where different security models have to mesh smoothly, and that’s where vulnerabilities tend to show up.

The Messy Regulatory Reality

The regulatory situation is genuinely confusing right now. Financial regulations assume humans make decisions and can be held accountable. When an AI agent conducts a transaction that violates securities law or consumer protection rules, who’s responsible? The person who owns the agent? The company that built it? The platform it runs on? Nobody has clear answers.

Cross-border transactions make this even messier. An AI agent using x402 could conduct business across multiple countries simultaneously, each with different financial regulations. Traditional regulatory frameworks based on geographic boundaries don’t have obvious ways to handle this scenario.

Consumer protection becomes really tricky when transactions happen at machine speed. Traditional approaches rely on disclosure requirements and cooling-off periods, giving humans time to review what they’re buying and potentially cancel. That doesn’t work when an AI agent completes a purchase in milliseconds based on real-time conditions.

Then there’s the market concentration question that keeps economists up at night. Even if payment protocols remain open, the companies with the best AI agents will have persistent advantages. Access to superior AI technology could become a primary source of market power in automated systems, creating winner-take-all dynamics that open protocols were supposed to prevent.

Where This Goes Next

The machine-to-machine connections market is expected to reach $33.31 billion by 2032, growing at 6.5% annually. But those projections were made before the current wave of AI agent payment infrastructure. If autonomous agents really do become Ethereum’s biggest users, as some people expect, growth could accelerate way beyond current estimates.

What we’re watching is the early stage of a major shift in how economic transactions work. The infrastructure is being built right now: payment protocols, identity verification systems, dispute resolution mechanisms. Whether the AI agents themselves are sophisticated enough to use that infrastructure responsibly remains the big open question.

The Stanford research suggests we should pump the brakes a bit. That 14% profit differential between strong and weak agents represents real economic harm that could affect millions of consumers if AI agent commerce becomes widespread. Payment protocols provide the pipes, but they don’t guarantee fair outcomes if the AI systems using those pipes have vastly different capabilities.

Getting this right will require ongoing cooperation between tech companies, researchers, and policymakers. The technical achievements in payment infrastructure are genuinely impressive, but they’re only one piece of building fair and stable automated markets. The hard work of making sure AI agents actually serve user interests, rather than just moving money efficiently, is really just beginning.

What’s clear is that AI agents buying and selling things isn’t coming someday. It’s happening now. Google and Coinbase are building the rails. Ethereum is positioning itself as the settlement layer. Companies are planning product launches for late 2025. The question isn’t whether this future arrives. It’s whether we’ll build it in a way that actually benefits people, not just the companies with the most sophisticated AI systems.