Two agentic commerce forecasts for 2026, a hundred times apart

There are two numbers in circulation for the size of agentic commerce in 2026. One is $8bn. The other is $944bn. They are not competing estimates of the same thing, and anyone building a business case off a headline this quarter should know which one they have picked up.

Two numbers, one label

Juniper Research puts agentic commerce transaction value at $8bn in 2026, rising to $3.5tn by 2031 [1]. PHD, working with WARC, puts agent-facilitated consumer spending at more than $944bn in 2026, reaching $3.35tn by 2030, which it calculates as 3.8% of global consumer spending [2].

The distance between them is a definition. Juniper is counting value that passes through an agent as the transacting party. PHD and WARC are counting spending an agent shapes, and their own framing is explicit about it: agents increasingly determine what gets seen, shortlisted and bought [2]. One is a payments number. The other is a marketing number. Both get filed under agentic commerce in the coverage, and the ratio between them is about 118 to 1.

The PHD and WARC forecast is concentrated rather than general. The top ten markets account for close to 68% of the total, with the United States at $1.1tn by 2030, or 31.9%, China at roughly $505bn, and the United Kingdom at 3.9% [2]. The fastest-moving sector is telecoms and utilities, forecast to go from $57.6bn in 2026 to $410.3bn by 2030 [2]. Renewals and switching, in other words, which is the part of buying people are gladdest to hand over.

Cards lead, and banks are being told the lead is temporary

On the payments side of the split, cards are where agentic volume currently sits, and the American Banker read is that Visa and Mastercard have bought themselves an early-mover position rather than a permanent one [1]. The argument for it being temporary is account-to-account, forecast to grow 113% between 2025 and 2030, from $91.5tn to $195tn globally [1]. The warning attached is blunt: failing to support local methods, including digital wallets and account-to-account, will hold the market back [1].

The autonomy case is strongest where no consumer is watching

Sunrate and Mastercard published a joint white paper, "Beyond Automation: Defining Agentic Global Payments", at the World Artificial Intelligence Conference on 24 July [3]. It argues cross-border payments are moving past digitisation and automation into a stage it calls Autonomy, where agents plan and execute end-to-end payment and treasury workflows inside defined governance frameworks [3]. The paper maps 16 pain points across the B2B payment lifecycle and sets out 13 agent use cases, among them supplier onboarding, payment routing, foreign exchange management, compliance screening and fraud detection [3].

Nothing in that list requires a consumer to trust anything. Fragmented workflows, disconnected systems, foreign exchange leakage and reconciliation are internal problems, and an agent working on them answers to the company that deployed it [3].

Read from the rails

Pick your denominator before you write the business case. If your revenue depends on an agent completing a payment, you are building into the $8bn number, and the honest version of that plan says most of the volume arrives after 2028. If your revenue depends on being the thing an agent selects, you are building into the $944bn number, and it is already here. Those are different products with different timelines, and I have watched teams commit to the first while quoting the second in the board deck.

The second thing worth taking from this week is that the two credible routes to real autonomy both avoid the consumer. One is the seller side, agents running listings, pricing and media. The other is corporate treasury, which is what Sunrate and Mastercard have written up. Ten years in payments operations left me with a fairly reliable rule: capability lands first wherever the number of parties who must consent is smallest. In treasury that number is one. At a consumer checkout it is the customer, the merchant, the issuer and the scheme, and each of them has a reason to say wait.

So the near-term build is unglamorous. Make your catalogue, your pricing and your availability legible to an agent that is choosing on someone's behalf, and automate the internal payment work where you are the only party who has to trust the thing. Neither needs a protocol to be finalised, and neither needs a consumer to get comfortable.

An $8bn market and a $944bn market are both being called agentic commerce this month. One counts the agent paying. The other counts the agent choosing. Only one of them is already happening at scale.