$15TB2B purchases handled by AI agents by 2028 (Gartner)

33%Of enterprise software apps will include agentic AI by 2028 (Gartner)

$4.4TAnnual generative-AI economic potential, upper bound (McKinsey, 2023)

~0%Of agent-to-API interactions currently monetised at the tool/API level

The capability threshold moment

Every transformative technology market has a capability threshold — the point at which a technology becomes reliable enough and affordable enough that it transitions from experimental to infrastructural. For AI agents, that threshold was crossed somewhere between late 2024 and early 2026.

The evidence is not subtle. McKinsey put the annual economic potential of generative AI at $2.6–$4.4 trillion back in 2023 — and the tool-calling, API-consuming agents that actually realise that potential are now a production workload, not a research curiosity. Enterprise AI agent deployments — systems that autonomously call external APIs, search the web, write and execute code, and take actions in external systems — are no longer pilot programs. They are production infrastructure at Fortune 500 companies. The Model Context Protocol (MCP), introduced by Anthropic in late 2024, gave this ecosystem a standardised tool interface and catalysed an explosion of publicly available tool servers.

What has not kept pace is payment infrastructure. The API economy that emerged from 2006 to 2016 was built for human developers: sign up, enter a card, get an API key, hit rate limits, upgrade a plan. None of those steps are available to an autonomous AI agent. The agent has no email address, no browser session, no hands to enter a credit card number. It has a context window, a set of tool calls, and a task.

ℹ

The mismatch is not a product gap that will be closed by adding an "agent login" flow. It requires rethinking the payment primitive from first principles — which is precisely what protocols like x402, L402, and MPP are attempting to do.

The historical parallel: APIs and the web services era

To understand where agent payments are headed, it helps to look at the analogous moment in the API economy. In the mid-2000s, most web services had no programmatic payment layer. Google Maps launched its API in June 2005, Amazon S3 launched in March 2006, and Salesforce's developer APIs were older still — but for the first few years, access was either free, invitation-only, or billed through enterprise contracts negotiated by humans.

Stripe launched in 2010 and changed the fundamental economics of the developer API market by making programmatic payment trivially easy. Within two years, the pattern of "developer builds product, monetises via API, bills via Stripe" was the default. A Stripe-like infrastructure layer unlocked an entire category of businesses that could not exist before.

We are at an equivalent moment for the agent economy. The agent equivalent of Stripe — infrastructure that makes it trivially easy for an AI agent to pay for a tool call, and for an API operator to receive that payment — does not yet exist as a category default. The companies that establish this infrastructure position will occupy the same structural role that Stripe holds in the human web economy.

2005–06

Google Maps API (June 2005) and Amazon S3 (March 2006) launch

API economy begins — no native payment layer

2010

Stripe launches

Developer-native payments unlock the API monetisation market

2022

ChatGPT launches; plugin ecosystem emerges

First wave of AI tool calling — payments still ad-hoc

2024

MCP released; agentic frameworks mature

Standardised tool interface; agent-native payment gap becomes structural

2025

x402 and L402 reach production use

First agent-native payment protocols in active use

2026

MPP launched (Stripe + Tempo, March 18); enterprise agent adoption

Major incumbents enter with an open standard; market structuring phase begins

Why the market is emerging now, not in 2022

The timing is not arbitrary. Three things converged in 2024–2026 that make agent payments a structural inevitability rather than a speculative bet:

1. Sufficient agent capability

The GPT-4 class of models, followed by Claude 3 and their successors, crossed the reliability threshold for tool use. Earlier models hallucinated tool parameters, failed on multi-step tasks, and required constant human correction. Current models can reliably plan, execute, and recover from tool failures across complex multi-step workflows. Agents that actually complete tasks autonomously are now deployable at production scale.

2. Protocol standardisation via MCP

Before MCP, every LLM provider had a different function-calling interface. OpenAI's function calling, Anthropic's tool use, Google's Gemini function declarations — all structurally similar but incompatible in detail. MCP provided a single standard that tool builders can target once and have it work across any MCP-compatible LLM. By standardising tool access, MCP also clarified where payment belongs: at the tool server level, intercepting tool calls before they execute.

3. Enterprise adoption at scale

In its June 2025 press release, Gartner forecast that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024, and that 15% of day-to-day work decisions will be made autonomously. Enterprise adoption creates the spending pressure: when a corporate agent is authorised to spend $50/day on external API calls, that $50 needs to flow to someone, through a system that has audit trails, approval workflows, and budget governance. The human-centric payment stack simply cannot handle this.

The size of the market opportunity

The addressable market for agent payment infrastructure is not just the transaction fees on agent-to-API payments. It is the entire stack: payment routing, balance management, per-agent spend controls, audit and compliance logging, protocol translation, and seller payout infrastructure.

The underlying transaction volume is not hypothetical. In its October 21, 2025 Top Predictions release, Gartner forecast that “by 2028, 90% of B2B buying will be AI agent intermediated, pushing over $15 trillion of B2B spend through AI agent exchanges.” Even a 10 basis-point infrastructure layer on top of that flow — roughly one-thirtieth of Stripe's blended card take rate — represents $15 billion in annual revenue from a standing start. The infrastructure layer is historically more defensible and higher-margin than the applications built on top of it.

✦

For comparison: Stripe's own total payment volume in 2024 was approximately $1.4 trillion. If AI agents handle $15T of B2B spending by 2028 as Gartner projects, agent-native payment rails clear roughly ten Stripes of volume — before counting balance services, compliance tooling, or enterprise contracts.

The EU AI Act creates a new forcing function

In Europe, the market pressure is not just commercial — it is regulatory. The EU AI Act, which entered enforcement in 2026, imposes transparency and audit requirements on high-risk AI systems. An enterprise AI agent that autonomously spends money is, by most reasonable interpretations, a system that requires auditability: which agent, which action, how much, when, for what purpose, approved by whom.

No current agent payment protocol natively produces audit trails in the format compliance teams require. x402 transactions are on-chain but not human-readable in the context of enterprise expense management. L402 macaroons carry cryptographic proofs but not business-context metadata. Stripe's transaction logs are human-readable but not agent-attributed. The compliance gap is real, and enterprises that cannot demonstrate governance of agent spending will face regulatory exposure.

This creates a strong pull for payment middleware that solves compliance as a first-class feature — not as an afterthought bolted onto a crypto protocol.

Who builds the infrastructure layer

The agent payment infrastructure market is at the "pre-Stripe" moment of the API economy. A few categories of player are competing to establish the default:

Payment incumbents (Stripe, Visa, Mastercard, Ant Group) — extending existing rails with AI-native APIs. Advantage: existing merchant relationships and compliance infrastructure. Disadvantage: fiat-only, slow to support crypto-native agent ecosystems.
Crypto-native protocols (x402, L402, Coinbase) — building from first principles for crypto agents. Advantage: sub-cent micropayments, permissionless. Disadvantage: no fiat support, inaccessible to enterprise buyers.
Protocol-agnostic middleware (Custena and similar) — abstracting over all protocols, targeting sellers who want to support all agent types without building payment logic. Advantage: full coverage, fastest to deploy. Disadvantage: not building a protocol, so dependent on protocol ecosystem health.

The historical parallel suggests the infrastructure abstraction layer wins in the long run. Stripe did not win by backing OAuth or SAML or any specific auth protocol — it won by making it irrelevant which protocol your bank used. The agent payment infrastructure winner will similarly abstract over the protocol wars.

The structural inevitability

Agent payments are not a feature request. They are a structural consequence of the agent economy that is already underway. The APIs that cannot accept agent payments will become progressively less accessible as the agent-to-human ratio of API consumers shifts. The payment infrastructure that emerges from this transition will be as fundamental to the next decade of software commerce as Stripe was to the last.

The question for API operators, MCP server builders, and infrastructure investors is not whether to engage with this market — it is how quickly to move and which bets to make. Given the protocol fragmentation currently in play, the safest bet is infrastructure that does not require betting on a single protocol at all.

Why AI Agent Payments Are the Future of Commerce Infrastructure