Can AI Agents Actually Pay for Things? Inside the Push for Agentic Commerce with Stablecoins

There is an HTTP status code the internet reserved back in the 1990s and then never used. Code 402: Payment Required. For around 30 years, it sat in the spec as a placeholder, a door nobody had a key for. But in 2026, it became the backbone of agentic commerce and how AI agents pay for things.

So can an AI agent actually complete a purchase, start to finish, with no human in the loop? Yes. As of this year, it can. The catch is that it mostly can’t do it on the payment system you already use, because cards and bank accounts assume a human is on one end. That gap is the whole story of agentic commerce right now. And stablecoins are the workaround the biggest tech companies quietly agreed on. I want to walk through what is actually working, what is hype, and where I think this goes.

Key Takeaways

• AI agents can transact autonomously today, but almost entirely on new blockchain rails, not on cards.
• Agentic commerce could drive up to $17.5 trillion in commerce by 2030, per Deloitte.
• Stablecoins are the settlement layer of choice because card fees make sub-dollar agent payments impossible.
• Real agent payment volume is still tiny: one filtered estimate puts genuine activity near $1.6 million a month.
• The blockers are not technical anymore. They are liability, wallet custody, and compliance.
• Gartner projects machine customers could drive up to 20% of revenue by 2030.

What Is Agentic Commerce, And Why It Matters Right Now

Agentic commerce is when an AI agent does the buying. Not recommends, not drafts a cart for you to approve, but searches, decides, pays, and settles the transaction on its own. The agent pays for what it needs to finish a task, like an API call, a data feed, compute time, or the output of another agent.

That last line is where people trip over. An AI shopping assistant that suggests three toasters and waits for you to buy them isn’t agentic commerce. It’s just a recommendation system with extra steps. But the moment the software itself authorizes and completes the payment, you have crossed into agentic payments.

Why is this landing in 2026 and not three years ago? Two things had to be true at once.

• Agents got good enough to do valuable work: Once agents could genuinely complete multi-step jobs like research, procurement, and coding, how they pay for resources stopped being a thought experiment.
• The payment protocols matured: The x402 protocol gave software a standard way to pay for a web resource without a human entering card details, and by early 2026 Stripe, AWS, Google, Mastercard, Visa, and Cloudflare were all involved.

For scale, global ecommerce is projected to hit $6.88 trillion in 2026. If agents start mediating even a slice of that, the rails underneath AI commerce matter a lot.

How Agentic Payments Actually Work

An agentic payment follows a real-time decision system. The agent evaluates what it needs, checks what it’s allowed to spend, and transacts on that judgment. No person approves it in the moment.

Strip away the jargon and the lifecycle has four beats:

1. Intent: The agent hits a wall mid-task. Like it needs a paid API, a dataset, or another agent’s output to continue.
2. Authorization: The agent proves it’s allowed to spend, and that the spend fits the limit the owner has fixed.
3. Settlement: Value moves, usually in stablecoins, and gets confirmed in seconds.
4. Reconciliation: The payment gets logged in a way a human or auditor can verify later.

Agentic payments tend to collapse billing and payment into a single primitive. Instead of using a service and getting billed later, the request itself carries the payment. The agent asks for a resource, includes proof of payment as its credential, and the service runs the instant that payment clears. This is exactly what HTTP 402 was waiting for.

Norwest also splits agentic commerce into three flows.

1. Human-to-agent (you delegate a budget to your assistant).
2. Agent-to-business (an agent buys compute from a cloud provider).
3. Agent-to-agent (two autonomous systems settle with each other).

The pattern across all three is the same. Software is executing transactions, not suggesting them.

The authorization problem for AI agents

Here’s the wall all of this runs into. Card rails were built on one assumption that a human is on one end of every transaction. Card networks need a cardholder. Bank accounts need a legal identity. But an AI agent has neither. It can’t legally open an account, can’t pass identity verification, and surely can’t click through a wallet pop-up at 2 am on a Sunday.

So the real question is who authorized this agent, what is it allowed to spend, and who is liable when things go south. Delegated authority, hard spending caps, and revocable permissions have to live somewhere where the payment system can read it. But cards don’t have a slot for that. This is the exact gap stablecoins and smart contracts step in to reduce.

Why Stablecoins Became The Rail For AI Agents

Stablecoins won this role for an unglamorous reason: the math on card fees doesn’t work for agents.

I pulled the numbers from the sources below and ran the break-even myself, because this is the part that makes everything else click.

Cost line	Card rail	Stablecoin rail (Base)
Fee per transaction	2.9% + $0.30	Under $0.0001
Typical agent payment	Around $0.31 per transaction	Around $0.31 per transaction
Settlement time	ACH T+2, banking hours	Seconds, 24/7

Do the arithmetic on a $0.31 payment. The card fee is $0.30 plus about a cent, which eats the entire payment before you have bought anything. Card economics fall apart below roughly $5 per transaction, and most agent payments will never get near that. On a stablecoin rail, the fee is a rounding error, well under a tenth of a cent.

Beyond price, stablecoins give agents three things cards don’t:

• Programmability: A stablecoin payment can be wrapped in code that enforces rules automatically.
• Always-on settlement: Blockchains confirm in seconds at any hour. An agent running overnight doesn’t have to wait till Tuesday for ACH to clear.
• Viable micro-transactions: Solana settles for about $0.00025 per transaction with 400ms finality, which makes sub-cent crypto payments actually work.

The liquidity is there too. Stablecoin transaction volume hit $33 trillion in 2025, up 72% YoY. But almost none of that $33 trillion is done by agents. AI agent activity is a rounding error inside stablecoins so far, which I will get into.

The Infrastructure Layer: Crypto Wallets and Smart Contracts Architecture

For an agent to pay, it needs somewhere to hold value and a way to enforce the rules. That is where crypto wallets and smart contracts come in.

An agent-controlled wallet is what it sounds like. The agent holds a balance and can move it. Genuinely nerve-making to run, because if the agent holds the private keys, then a compromised model or a breach can drain your whole wallet. Serious setups use distributed key custody, where the full private key never sits in one place.

Smart contracts are where the guardrails live. Here, instead of trusting the agent to behave, you encode the limits into code that runs on-chain. A smart contract can enforce things like:

• A hard per-transaction cap, so a single payment can never exceed, say, $5.
• A daily or monthly ceiling on total spend.
• An allowlist of which services or agents can receive funds.
• An automatic escalation to a human above a threshold.

Picture an agent with a smart contract that refuses any payment over $2 and blocks anything outside an approved list of APIs. The agent can act freely inside that box, and the box is enforced by math. That is the actual appeal here. You aren’t trusting the agent. You’re trusting the contract that fences it in. Identity standards like ERC-8004, which had over 24,000 agents registered within weeks of its January 2026 launch, are trying to give agents verifiable credentials to plug into these rules.

Who Is Building Agentic Commerce Today?

This stopped being speculative fast. In 2026, AWS, Stripe, and Google all shipped agent payment systems built on blockchain rails and stablecoin settlement, within about a month of each other. Three of the largest tech companies on earth hit the same wall and built the same door at nearly the same time. That is the clearest signal this is real and not a crypto-Twitter fever dream.

The protocol landscape is genuinely an alphabet soup, so here is a map of who is doing what.

Protocol / Player	Who is behind it	What it does	Status
x402	Coinbase, now governed by the Linux Foundation	Payment-as-credential over HTTP 402	Live, 35M+ transactions on Solana
AP2 (Agent Payments Protocol)	Google	Machine-native payment handshakes across fiat and crypto	Live, 60+ partners
ACP (Agentic Commerce Protocol)	OpenAI and Stripe	Enables AI agents to browse and complete purchases	Live
Agent Pay / Intelligent Commerce	Mastercard, Visa	Extends card networks into AI agent environments	Rolling out
MPP (Machine Payments Protocol)	Stripe / Tempo	Uses payment as the authentication credential	Emerging

The one to watch is x402. It has already processed more than 35 million transactions on Solana and handles roughly $600 million in annualized volume across chains, with the Linux Foundation taking over its governance in April 2026. Google’s AP2 launched in September 2025 and pulled in more than 60 organizations including PayPal, Coinbase, Mastercard, and American Express. When card networks and crypto protocols both show up to the same standard, that is consensus forming.

The Hard Problems Blocking Agentic Commerce Adoption

Now the part the vendor blogs skate past. The technology mostly works. The rest doesn’t, and I wouldn’t deploy an agent with a wallet until these are answered.

Start with the volume. Despite the billions in stablecoin liquidity, actual AI agent payments are very small. One analysis pegged agent activity at around $50 million across 40,000 on-chain agents, which is about 0.0001% of stablecoin volume. Filter out wash trading, and one estimate drops genuine agent payments to roughly $1.6 million a month. This is a real thing but at a very tiny scale.

The genuinely unsolved problems:

• Liability: When an agent overspends or buys the wrong thing, who eats it? You, the model vendor, the merchant? Card networks have decades of chargeback rules for this. Crypto payments have none.
• No chargebacks: Stablecoin settlement is fast because it’s final. That finality means if an agent gets tricked or goes rogue, there is no dispute this charge button.
• Compliance: The GENIUS Act made stablecoin issuers financial institutions under the Bank Secrecy Act, and Treasury and FinCEN published implementing rules in April 2026. The problem is existing compliance stacks screen human customers at onboarding. Agents have no identity to onboard, so the industry is shifting from KYC to “Know Your Agent,” and that framework is barely built.
• Custody risk: As covered, an agent holding keys is a security surface. Get it wrong and the wallet empties.
• Idle float: Money sitting in an agent’s wallet between transactions earns nothing and, worse, may not be running AML screening at all.

None of these are reasons to write off agentic commerce. They are reasons to treat “the agent can technically pay” as the easy 20% of the problem.

Where Agentic Commerce Goes From Here

I’ll try not to make any predictions here, so let me split this into what I would actually bet on versus what is a maybe.

Likely in the next 12 months: Agent-to-business payments grow first, inside controlled settings. An enterprise’s own agents paying for compute, APIs, and data with tight smart-contract limits. The risk is contained because the company controls both ends. Expect the protocol soup to start consolidating around x402 and AP2, since that is where the card networks and cloud giants are converging.

The 5-year maybe: Agents handling meaningful consumer spending, negotiating A2A in open marketplaces, and driving a real share of the $17.5 trillion.

For any of it to go mainstream, three things have to become boring and standard. A clear liability framework so people know who pays when an agent goes south, custody standards that make agent wallets genuinely safe, and merchant acceptance broad enough that agents can pay in the wild.

Final Thoughts

So, back to the question I started with. Can AI agents actually pay for things? Yes. Not a maybe, but a full yes. The rails exist, the biggest players shipped them, and an agent can move stablecoins through a smart contract without a human touching it. But can pay and should be trusted to pay at scale are different sentences. The current volume is tiny, with tons of questions on liability and compliance. And the smart move today is small, fenced, auditable deployments rather than handing an agent your credit card.

That dormant HTTP 402 code finally has a use after 30 years. Whether agentic commerce becomes the default way software buys things, or stays a niche for machine-to-machine plumbing, is the thing worth watching over the next few years.

FAQs