Language
Book a discovery call
AI Strategy · Agentic Economy

How machineswill pay machines.

A Tuesday morning, sometime in 2029: overnight, your company’s purchasing agent negotiated prices with three supplier agents, confirmed the best terms via a signed mandate, and settled the cent-sized fee for a database query in stablecoin on the side — logged, budgeted, while everyone slept. Science fiction? The payment protocols for it are already live: x402 counts over 165 million agent transactions, ChatGPT has an instant checkout, Google signs purchasing mandates for agents. Here is how agent-to-agent payments work, the three phases in which they become everyday reality — and where your opportunities lie.

Summary
  • The payment stack for machines already exists: x402 (Coinbase, stablecoin settlement), ACP (OpenAI/Stripe, checkout), AP2 (Google, signed mandates), MPP (Stripe, micropayment sessions) — plus Visa and Mastercard making their card rails agent-ready.
  • Three phases to everyday reality: today the agent recommends and the human confirms. In 2027/28 agents buy within delegated budgets. From ~2029, machines negotiate with machines — in milliseconds.
  • The economic consequence: usage-based billing displaces subscriptions, prices get negotiated in real time — and whoever is not machine-readable does not exist for buying agents.
  • The mid-market opportunity: become agent-readable now (selling side), test a purchasing agent (buying side), and build clean budgets and logs (foundation).

The prompt for this look ahead comes from Visa: the world’s largest payment processor declared at its Payments Forum that AI agents will shop over card rails and settle among each other in stablecoins — for sub-dollar amounts, card fees are simply too expensive, while newer chains settle for fractions of a cent. You can file that as a crypto story. Or you look closer and see: the cash register for an economy in which software shops at software is being built right now. And it is further along than most people think.

The payment stack for machines — it already exists

Payments between agents need three things: proof that the agent is allowed to buy (authorisation), a way to complete the purchase (checkout), and the actual movement of money (settlement). Every layer already has protocols with real traction:

ProtocolLayerStatus July 2026
AP2 (Google)Authorisation: cryptographically signed “mandates” — what may the agent spend, in which category, until whenLaunched by Google with a broad partner roster; the power-of-attorney model of the agentic economy
ACP (OpenAI/Stripe)Checkout: the agent negotiates cart and purchase directly with the merchantPowers the Instant Checkout in ChatGPT — buying without leaving the chat
x402 (Coinbase)Settlement: HTTP status 402 “Payment Required” becomes real — server names a price, agent pays in stablecoin, server deliversOver 165M agent transactions, Stripe integration since February 2026, Cloudflare support
MPP (Stripe/Tempo)Settlement: “sessions” — the agent pre-authorises a limit and streams micropayments in stablecoin or fiatLaunched March 2026
Visa & MastercardCard rails for purchases on behalf of humans — plus trust infrastructure (Agentic Directory, Agent Score)Visa Intelligent Commerce with OpenAI partnership; ~$7bn annualised stablecoin settlement

The protocols do not compete, they stack: one purchase can use AP2 (mandate) + ACP (checkout) + x402 (settlement) in a single flow. Sources in “Sources and context”

This is what separates this development from most AI trends: these are not whitepapers, they are running systems with transaction volume. The question is no longer if, but at what pace.

The three phases to everyday reality

PhaseWhat happensWho pays
TodayAgents research, compare, fill the cart; instant checkouts inside chat interfacesThe human confirms every purchase — the agent is an advisor with the till in sight
2027/28 — delegated budgetsSigned mandates become normal: “Buy office supplies up to €200/month from these merchants.” Companies run wallets and limits per agentThe agent — within hard, machine-readable boundaries; the human reviews the log
From ~2029 — machine-to-machine marketsBuying agents negotiate with selling agents in milliseconds; services without human frontends; pay-per-use as the norm instead of subscriptionsMachines among themselves — humans set goals and guardrails

Framing: Digital Maker — the timeline is a scenario, not a promise; the pace depends on regulation and trust infrastructure

Important context: phase 1 and the beginning of phase 2 are no longer a forecast — instant checkouts are live, mandates are specified, and at small scale phase 2 has long been reality in every agent operation. We run 54 productive agents with budgets per agent (in German) — today for API calls and compute, tomorrow for real purchases. The mechanism is the same; only the merchant changes.

What changes economically — three shifts

  • From subscription to usage. Once sub-cent payments become economical, nobody needs a monthly subscription for a service their agent uses three times a quarter. Software, data and compute get paid per call — good for buyers, a rebuild for every provider with a subscription model.
  • Prices get negotiated — in milliseconds. When both sides are agents, list prices are only the starting point. Availability, volume, timing, payment speed — everything flows into machine negotiation. Dynamic pricing, the exception today, becomes the norm.
  • Visibility to agents becomes the new SEO. A buying agent reads no banners. It reads structured data, price lists, APIs — and trust signals like Visa’s Agent Score. Whoever does not show up there does not exist. And yes, the Agent Score raises the same question as every gatekeeping layer: who rates, by which rules, with what power? Choosing dependencies consciously applies here too.

What this means for the mid-market — three moves, by urgency

  • 1. Selling side: become agent-readable. Structured product data, transparent prices, clean interfaces — that decides whether buying agents find you and can buy from you. The side effect: exactly this homework already improves your visibility in AI search today. You prepare for 2029 and win immediately.
  • 2. Buying side: test your first purchasing agent. Price comparisons, reorders, consumables — an agent that compares around the clock and prepares orders (the human confirms) saves from month one and builds the know-how that becomes an edge in phase 2.
  • 3. Foundation: budgets, limits, audit trail. Every payment-capable agent needs the controls of a company card — budget per period, allowed purposes, complete logging, clear responsibility. Build it today and you can plug in any payment rail tomorrow. Skip it and the agentic economy arrives as loss of control.

We know the pattern from every technology wave: the winners are rarely the first to invest — they are the ones who are ready when it tips. Agent-to-agent payments will not arrive on a fixed date; they seep in: first the checkout in the chat, then the delegated budget, then the negotiating agent. Whoever stays compatible at every step never has to catch up.

Sources and context

Based on the announcements of the Visa Payments Forum 2026 (Visa Intelligent Commerce, Agentic Directory, Agent Score, OpenAI partnership; ~$7bn annualised stablecoin settlement, 160+ stablecoin card programmes) and the Visa/Artemis report on the macro/micro-commerce split. Protocol details as of July 2026: x402 (Coinbase; 165M+ agent transactions, Stripe integration February 2026, Cloudflare support), ACP (OpenAI/Stripe, Instant Checkout in ChatGPT), AP2 (Google, signed mandates), MPP (Stripe/Tempo, launched March 2026) — figures and maturity levels are snapshots from public vendor and trade reports. The phase timeline is a Digital Maker scenario based on today’s protocol landscape, not a promise; the pace depends on regulation and trust infrastructure. Practice examples (54 budgeted agents) from our own operations; no legal or investment advice.

Frequently asked questions: agent-to-agent payments

How will AI agents pay in the future?

Over a dedicated payment stack that is being built right now: cryptographically signed mandates define what an agent may spend (Google’s AP2), a checkout protocol handles the purchase with the merchant (ACP by OpenAI/Stripe), and settlement runs over classic card rails or — for small amounts — stablecoins with sub-cent costs (Coinbase’s x402, Stripe’s MPP sessions). For sub-dollar amounts, card fees are simply too high — which is why Visa sees stablecoins as the default rail there.

Which agent payment protocols already exist today?

Four with real traction, as of July 2026: x402 (Coinbase) turns HTTP status code 402 “Payment Required” into a working payment standard — over 165 million agent transactions, with a Stripe integration since February 2026. ACP (OpenAI/Stripe) powers the Instant Checkout in ChatGPT. AP2 (Google) governs authorisation with signed mandates: what may the agent buy, up to what amount, until when. MPP (Stripe/Tempo, since March 2026) lets agents pre-authorise a limit and stream micropayments. On top, Visa (Intelligent Commerce) and Mastercard (Agent Pay) are making their card rails agent-ready.

When will agent-to-agent payments become everyday reality?

In three phases. Today: agents research and fill the cart, the human confirms — the first instant checkouts are already live. 2027/28: delegated budgets become normal — agents buy autonomously within signed mandates, companies run wallets and limits per agent. From around 2029: true machine-to-machine markets, where buying and selling agents negotiate prices in milliseconds and services are billed per use instead of per subscription. The pace depends on regulation — the technology is ahead of the rules.

What changes for merchants and providers?

The biggest shift: your customer will often no longer be a human but their agent. It does not read banners and is not swayed by store design — it compares machine-readable prices, availability and terms. Whoever does not offer their services in agent-readable form (structured data, clear prices, API access) simply does not exist for buying agents. Visibility to agents becomes the new SEO — and usage-based billing displaces the subscription in many areas.

How does a mid-market company prepare concretely?

Three moves, ordered by urgency: first, make your selling side agent-readable — structured product data, transparent prices, eventually an API; that already pays into your visibility in AI search today. Second, test the buying side: a purchasing agent that compares prices and prepares orders saves money from month one. Third, the foundation: budgets, limits and an audit trail per agent — whoever builds that cleanly today can plug in any of the new payment rails the moment they become relevant.

Should agents be able to buy from you — or buy for you?

Both start with the same homework: machine-readable offers, a first purchasing agent, clean budgets and logs. We run 54 budgeted agents and build exactly that for clients. In a discovery call we look at where your business stands — 30 minutes, no slides.

Book a discovery call