The machine customer is buying boring things first

Agentic commerce is usually illustrated with a cheerful assistant buying shoes, groceries, or a holiday. Consumer shopping is easy to picture and hard to prove. The more useful signal is already sitting farther down the stack: agents paying for narrow resources needed to finish another job.

Stripe's March 18, 2026 launch of the Machine Payments Protocol names the current examples. Stripe says Browserbase lets agents start headless browsers and pay per session. Parallel lets them pay per API call for web access. PostalForm lets them pay to print and send physical mail. There is also a sandwich service, because every new payment rail eventually finds lunch. The serious pattern is not the sandwich. It is an agent encountering a bounded resource, receiving a price, authorizing payment, and continuing its work.

This is vendor-published launch material. Stripe gives no transaction count, repeat rate, revenue, or autonomous-versus-human volume for these examples. One protocol announcement is not a demand curve. It is still better evidence than a concept video because named businesses have exposed named units that software can buy.

A resource beats a catalog

The near-term machine customer does not need 14 subscription tiers and a pricing quiz. It needs a resource with a stable contract: what it receives, what it costs, how long delivery takes, what credentials are required, and what happens if the call fails. A browser session, a verified record, a parcel quote, a document conversion, a risk check, or a slice of compute can fit that shape. A broad promise to transform the enterprise generally cannot.

Stripe describes the MPP sequence plainly. An agent requests a resource from a service, API, MCP server, or other HTTP endpoint. The service returns a payment request. The agent authorizes it. The service delivers the resource. That makes the commercial unit smaller than a software seat and more legible than a consulting engagement. It also makes metering, idempotency, timeout policy, and delivery evidence part of the product rather than the plumbing somebody remembers during an incident.

On the June 2 episode of The Startup Ideas Podcast, Greg Isenberg framed the broader agent buyer around identity, tools, inbox, memory, wallet, and receipts. The accompanying transcript adds a useful seller checklist: structured documentation, schemas, policies, examples, endpoints, MCP tools, SDKs, OAuth, checkout, sandboxes, and receipts. The market-size claim in the title is a prediction. The checklist is work a founder can do on Monday.

Payment is the easy-looking hard part

A machine-readable price does not make a safe machine customer. The seller still needs to know which agent arrived, whose authority it carries, how much it may spend, whether the request was altered, and where the money settles. The buyer needs proof that the resource arrived, a way to stop a runaway series, and a refund or dispute path when the machine buys the same thing 4,000 times with admirable consistency.

Mastercard's June 10, 2026 Agent Pay for Machines announcement is revealing because most of the product is control surface. Mastercard says every agent is credentialed; organizations can set authorization rules and spending limits that are enforced programmatically; and settlement can run across cards, accounts, and stablecoins. It listed more than 30 initial participants and supporters, while also saying the group is collaborating to validate priority use cases and establish common rules. That is ecosystem formation, not proven volume. The distinction can survive a press release.

Stripe is making a different but complementary point. It says MPP transactions can appear in the same Stripe API and dashboard as human payments and can use the existing machinery for tax, fraud protection, reporting, accounting integrations, refunds, balances, and payout schedules. The new interface may be agent-to-service. The back office remains stubbornly interested in ledgers.

The agent acts like a customer; the principal still owns the bill

Calling the agent a customer is useful product language and incomplete business language. The agent discovers, evaluates, invokes, and pays. A person or organization still owns the budget, the policy, the liability, and the complaint. Sellers therefore need two records: which machine made the request and which principal authorized that machine to make this request under these limits.

Australian Payments Plus published a small real-world example on June 24. AstraSync AI adopted ConnectID, Australia's bank-backed identity network, for developers registering on its Know Your Agent platform. With consent, the developer authenticates through an existing Australian bank; limited verified identity attributes are matched to the developer profile and cryptographically bound to the Know Your Developer record. AP+ describes the resulting chain as bank to developer to agent to counterparty.

That chain does not establish that agent commerce is large. It shows what has to exist before a serious counterparty opens the gate. Identity, delegated authority, intent, payment, delivery evidence, and recourse are separate jobs. A glossy agent profile solves approximately one of them, provided the photo is very reassuring.

Package one machine-grade SKU

Start with one resource that another agent already needs in the middle of a valuable workflow. Give it a machine-readable description with inputs, output schema, exclusions, latency target, version, price, and examples. Expose one bounded endpoint or tool. Make retries idempotent so a network wobble does not become a surprise growth channel.

Then add the commercial envelope. Require a verifiable agent and principal where the risk justifies it. Accept an authorization containing a unit cap, total spend cap, expiry, and allowed purpose. Quote the price before execution. Return a receipt containing the request identifier, resource version, delivery status, time, charge, and enough evidence to reconcile the result. Publish cancellation, refund, dispute, and support rules in a form software can parse and a human can defend.

Pilot with a small allowlist and a prepaid or tightly capped budget. Measure paid requests, successful delivery, duplicate suppression, retry rate, payment failure, refunds, disputes, human support minutes, gross margin per resource, and repeat purchase by principal. If the experiment needs a sales engineer to bless every call, you have a useful demo and an unusually punctual services business. Keep narrowing.

The operator takeaway

Founders should look for resources that are expensive for an agent to stop and ask a human to obtain: a browser session, a data lookup, an identity check, a current quote, a document transformation, a compliance search, a physical dispatch, a test environment. The best first SKU is low enough risk to authorize in advance, useful across many workflows, cheap enough to purchase repeatedly, and concrete enough to verify automatically.

Buyers should ask for paid autonomous volume, not total API traffic or a partner-logo collage. Request the last 90 days of machine-initiated transactions, unique principals, repeat rate, average and 95th-percentile purchase value, refund and dispute rates, failed deliveries, manual interventions, abuse losses, and contribution margin after payment and support costs. An agent can create a heroic number of requests without creating a customer.

Operators should treat the machine route as its own channel. Give it explicit pricing, permissions, service levels, fraud rules, reconciliation, and an owner. Compare it with the human route on acquisition cost, completion rate, support load, repeat use, and margin. The first machine customer is not proof that the economy has changed. It is a new kind of account that should still be able to pass an audit.

The machine customer does not need a brand anthem. It needs a reliable endpoint, a bounded wallet, and evidence that something happened.

That is less like ecommerce theater and more like a utility meter. Utilities, inconveniently, are businesses.

sources
  1. The Next $100B Market: Selling To AI AgentsThe Startup Ideas Podcast, accessed July 11, 2026
  2. The Next $100B Market: Selling To AI Agents transcriptPodscripts, accessed July 11, 2026
  3. Introducing the Machine Payments ProtocolStripe, accessed July 11, 2026
  4. Mastercard launches Agent Pay for Machines to unlock super-fast, always-on paymentsMastercard, accessed July 11, 2026
  5. AstraSync AI adopts ConnectID to verify the humans behind AI agentsAustralian Payments Plus, accessed July 11, 2026