Agentic commerce market size 2026: forecasts vs what is actually measured

The forecasts, the rails, and what they count

Two figures anchor every 2026 deck. Gartner forecasts that AI agents will intermediate more than $15 trillion in B2B spending by 2028, with 90% of B2B purchases handled by agents (Gartner IT Symposium/Xpo 2025; reported by Digital Commerce 360, 2025-11-28). McKinsey forecasts $3–5 trillion of global agentic commerce sales by 2030, including up to $1 trillion of orchestrated U.S. retail revenue (reported 2025-10-20). Both are forecasts. Neither measures anything that exists today.

Read the verbs. Gartner says intermediate; McKinsey says orchestrated. The object counted is the full ticket of a purchase in which an agent participated — it found the supplier, compared options, filled the cart, perhaps pressed buy. The money still moves on cards, ACH and invoices. Nothing in either forecast requires a new rail, a stablecoin, or a wallet held by software. They forecast how much existing commerce will pass through an agent, not how much money agents hold or move themselves. That distinction is the whole article. Our working definition of agentic finance →

The rails have moved faster than the ledgers. As cited in the Agentic Finance Graph research pack (2026-09-04): Visa and Mastercard describe agent-payment programs built on verifiable credentials — signed vouchers a merchant can check — aimed at the larger tickets and refunds that cards already handle well; Stripe and Tempo describe a Machine Payments Protocol (MPP) that routes one HTTP payment flow across cards, stablecoins and Lightning; Google’s Agent Payments Protocol (AP2) specifies mandate-style authorization, meaning what the user approved, for how much, and when; and x402, born at Coinbase, moved under the Linux Foundation as the x402 Foundation in July 2026 with 40 founding members, among them Visa, Mastercard, Stripe, Google, AWS, Circle and Cloudflare. These are standards and intentions. A published spec is not a settled dollar. Only x402 exposes a public ledger anyone can count, which is why the measured numbers below come almost entirely from it.

What is actually measured

Here is what can be counted on public rails today. Every row carries its object, window and source. The windows do not match, and we do not blend them.

Two arithmetic notes. $73M over 176M payments is about $0.41 each; the report’s stated average is ~$0.31, which tells you an average depends on which payments were included. And ~$1.45M over ~26M receipts is about $0.056. The typical machine payment in September 2026 is a few cents.

Each of these numbers counts a different object. Ranked from largest to smallest:

• Intermediated spend (GMV). The forecasts: the full ticket of any purchase an agent touched, on any rail. The largest object, and the only one with no public ledger.
• Settled value. Money that actually left an agent-controlled address: $73M in twelve months. Auditable, and small.
• Receipts. Transaction counts: 176M in a year, ~26M in 30 days. Dust and farming count the same as real demand. The most impressive number, and the easiest to inflate.
• TVL. A stock, not a flow: roughly $19M parked in AgentFi products across Base, Arbitrum and Ethereum (Cambrian, Q2 2026). How much machine money is managed, not how much is spent.
• Registrations. Identity mints: 385,998 in July 2026, of which 2.24% had a working service. No money is involved.
• Float. ~$311B of stablecoins and ~$33B of RWA (cited as 10x in two years) exist. Almost none of it is agent money. It is the pool agents could draw on.

Every trillion-dollar headline sits on the first rung. Every number we publish sits on rungs two to five. The L0–L9 liveness ladder exists so that a registration is never presented as a payer, and the ERC-8004 brief explains why 385,998 badges are not 385,998 agents.

Claim, object, measurement, condition

A bottom-up sizing framework (our estimates)

Sizing from the bottom is less exciting and more honest. The identity is simple:

annual settled value = economic agents × transfers per agent per day × average ticket × 365

Each factor has a measured anchor: ~21.2k paying buyers, ~41 transfers per buyer per day (26M ÷ 21.2k ÷ 30) and a ~$0.056 ticket from the x402scan window; a ~$0.31 reported average from Keyrock; 8,631 answering services from ERC-8004. The scenarios below are our own estimates, not anyone’s forecast, and they are run-rates, not cumulative totals.

Haircut scenarios A and C by the 25–30% farming estimate and the economic figure drops by a quarter. Then note what the forecasts require. At a $0.31 ticket, $3 trillion a year is about 26 billion payments a day; at 100 transfers per agent per day, that is roughly 265 million paying agents. The forecasts are not micropayment forecasts. They close only if agents carry human-sized tickets on ordinary rails — exactly what the card networks are building, and exactly what no public ledger yet shows. Scenario E, not scenario D, is the path to the forecasts.

The input–output problem

Gross figures mislead for a second reason: agents pay agents. Suppose agent A pays B $1 for data, B pays C $1 for compute, and C pays A $1 for a summary. Gross receipts: $3. Value that crossed the loop’s boundary: zero. National accounts solved this a century ago with input–output tables — GDP counts final demand and value added, not every intermediate invoice. Gross B2B spend exceeds GDP for the same reason: a chip is invoiced to a manufacturer, then again inside a laptop. Gartner’s $15T is gross intermediated spend. It should never be read as $15T of new economy.

Machine commerce needs the same split, and the loops are tighter because agents are cheap to spawn. On the graph, a payment is final demand when it exits the agent population — to a provider that is not itself an x402 payer, to a vault, to a fiat off-ramp — and intermediate when payer and payee share a funding parent or sit on a detected cycle. The gross number stays available, labeled. The economic number is the default. How we filter x402 volume →

Verdict: where the market is in September 2026, and what to watch

On rails that can be audited, the agent economy settles tens of millions of dollars a year, at a few cents per payment, almost entirely in USDC, with thousands — not hundreds of thousands — of agents that answer when called, and a treasury layer in the low tens of millions. That is a real economy with an unusual shape: high frequency, sub-dollar, machine-to-machine, priced below what cards can carry. It is not a trillion-dollar economy, and no honest measurement says it is.

The forecasts describe something else: existing human commerce, at existing ticket sizes, with an agent inserted into discovery and checkout. That may well arrive, largely on card and invoice rails; the credential work at Visa, Mastercard and Google is its plumbing. Both pictures converge on one question — whether agents become principals that hold and spend, or remain intermediaries that recommend and click. The measured rails say principals exist and are small. The forecasts say intermediaries will be enormous. Neither side is wrong about its own object. Both are wrong the moment they borrow the other’s number.

What to watch, and the reading that would move us:

• Agent-flagged card volume. The first network disclosure of agent-initiated transaction value. Until then, the retail forecasts have no measured counterpart.
• Average ticket on machine rails. Sustained movement above $1 would mean agents are carrying purchases, not pings.
• Economic share of receipts. Farming below 10% of x402 windows under repeat-counterparty tests.
• Working services. ERC-8004 endpoints that answer above 10% of registrations.
• Agent Treasury Utilization. AgentFi TVL crossing $100M against the stablecoin float — one basis point of machine money managed on purpose. Where idle USDC goes →
• Settlement mix. USDC below 95% of measured settlement, meaning a second issuer carries real machine volume.

The data panels carry the float and TVL context live; the statistics page holds every measured figure above with its window; the research hub links the rest. Operators who want the labeled, filtered series can register for the Agentic Desk.

Sources

• Gartner forecast, $15T agent-intermediated B2B spending by 2028, via Digital Commerce 360 — 2025-11-28 — link
• McKinsey forecast, $3–5T global agentic commerce sales by 2030, via Digital Commerce 360 — 2025-10-20 — link
• Keyrock, “Who Pays the Agent” (Coinbase and Tempo data) — May 2025 – Apr 2026 — link
• x402scan explorer, 30-day window — observed 2026-09 — link
• x402 Foundation under the Linux Foundation, 40 founding members — July 2026 — link; protocol source link
• Coinbase CDP x402 docs; the 25–30% farming estimate is attributed to Coinbase as cited in Agentic Finance Graph research pack — 2026-09-04 — link
• Token Dispatch, “Who Checks the Agents?” (385,998 registrations, 2.24% working), as cited in Agentic Finance Graph research pack — July 2026 — live counts at 8004scan link
• Guryanov, Base agent-registry liveness crawl (6.8% strict endpoint response) — through 2026-08-16 — link
• QuickNode ERC-8004 explorer, 30-day registrations, feedback and validation requests — 2026-09 — link
• Cambrian, Agentic Finance Landscape Q2 2026 (AgentFi TVL by chain) and DefiLlama stablecoin float (~$311B, 2026-09-04), as cited in Agentic Finance Graph research pack — 2026-09-04 — link; RWA context at rwa.xyz link
• Visa/Mastercard agent-payment credentials, Stripe/Tempo MPP and Google AP2, announcements as cited in Agentic Finance Graph research pack — 2026-09-04 — link

FAQ