Until September 2026, agentic transaction data wasn’t on most merchants’ minds. Agentic checkout was a “later” project. Early pilots were confined to single AI interfaces, where shoppers largely bought inside chat windows, and open standards such as OpenAI’s Agentic Commerce Protocol (ACP) or Google’s Universal Commerce Protocol (UCP) passed the transaction straight to the merchant.

However, since then, the game has changed. Agents like Muse, Instinct, and Grok aren’t waiting for an invitation. They’re using headless checkout if the merchant has built one — and if not, they’re browsing the merchant’s site the way a person would.

In the month after they arrived, agent orders on Forter’s network were 229% higher than the month before.

How quickly agentic orders are growing

About 1 in 10,000 orders on Forter’s network is now placed by an AI agent. That’s already 25x more agent orders per day than the first half of 2026. In the past 30 days, 60% of Forter merchants have seen at least one agent order, and 9% have seen more than 100.

Personal assistants that shop by clicking through the merchant’s site went from 1% of agent orders in early August to 77% by mid-October. Protocol agents kept growing during that period, and personal assistants grew on top of them. As of November 12, Muse accounted for 49% of agentic traffic, followed by ChatGPT’s protocol agent (22%), Grok (14%), and Instinct (9%).

Agentic commerce transactions are now 7.2x more frequent than in the first week of August, and 1.7x as large as human orders at the same merchants: $227 vs. $133. That gap shows up in nearly every vertical — most notably food and delivery ($117 vs. $57, or 2.1x) and digital goods and services ($145 vs. $99, or 1.5x).

Agentic transaction data: Human vs. agent AOV across physical goods, food & delivery, digital goods & services, and travel

“These are your customers using their agents to shop with you. So when these agents are being blocked, you’re actually blocking your good customers.” — Galit Michel, VP of Product, Forter

How authorization rates compare, and where checkout breaks

On orders that reach card authorization, banks are not turning agent traffic down. While completion rate is low, authorization rate is stable between agents and humans. This isn’t necessarily a “win.” Many issuers still cannot tell that these are agentic commerce transactions, and authorization only applies once an agent has submitted a payment.

A recent Visa study of 22 AI models from 12 labs across 15 shopping journeys found just a 57.2% average success rate, meaning nearly half of their test runs failed.

In Forter’s experiments, a headless checkout completed in about 5 to 10 seconds — against 5 to 10 minutes for browser automation — with a higher success rate and fewer AI tokens. Headless checkout already runs through ACP, UCP, MCP Apps, API-based headless architecture, and WebMCP.

“The bank authorization rate only applies when the agent was successful in getting as far as submitting a transaction.” — Adam Davies, Director, Forter AI

Agent fraud and the first chargebacks

The share of fraud attempts using automation has doubled in the past two months. Across Forter’s network, we’re seeing two patterns emerge: fraudster-created agents, powered by stolen identities and payment credentials, and takeovers of legitimate agents.

Agentic transaction data: Attempted fraud caused orders to jump significantly on one site

On one site, agent orders had jumped from roughly a dozen to more than 130 in a single day. Hundreds of “shoppers” were linked to just a few dozen devices. It was ultimately one operation, using one AI shopping assistant. Forter declined the transactions, and the activity dropped off.

And within this new agentic shopping wave, the first chargebacks are beginning to arrive. The 20 largest claims in the last 30 days were led by a $5,514 dispute, and the highest-value claims — and vast majority of chargebacks overall — have been service disputes, such as “item not as described.”

So far, the chargeback rate on personal-assistant orders is 0.012 percentage points from the rate on human orders at the same merchants. That comparison, however, is still early, as the increase in agent orders only happened in the past four to five weeks, so the window is still open.

“Just as consumer adoption for this technology grows, fraudulent adoption is growing as well.” — Adam Davies, Director, Forter AI

See more agentic transaction data

This data was originally shared during Off the Rails, our monthly series dedicated to helping you understand the most pressing problems in payments. By watching the full recording, you’ll receive the deck with all of the data and insights, including:

  • How agent baskets compare with human orders in each vertical
  • A walkthrough of a Muse checkout, and the three beauty sites where the agent was blocked
  • The live figures Forter is tracking across its network, including for cyber week, at ai.forter.com 
Published on November 18, 2026   •  
3 minute read   •  
Author: Forter Team