Agentic commerce is the purchase of goods and services by, or with the help of, AI agents acting for a buyer. A principles paper published on September 22, 2026 by six banks (ASB Bank, Bank of America, Capital One, Commonwealth Bank of Australia, ING Group and NatWest Group) defines it as the "use of AI agents to help make or facilitate payments between a consumer and a merchant for a product or service", occurring "across any payment rail, including cards, account-to-account and tokenized money", and treats it as "a subset of agentic payments", a broader category that also covers peer-to-peer payments and accounts payable and receivable (Building Trust in Agentic Commerce: Customer-centric principles (six banks, 2026)).
Degrees of autonomy
The banks' paper distinguishes two variants by the autonomy given to the agent. In human-in-the-loop agentic commerce the agent searches and a human reviews and executes the purchase, or the human pre-approves a specific purchase for the agent to carry out later. In autonomous agentic commerce the agent identifies a need and buys on its own after an initial prompt, subscriptions being the example given. The paper states that agents differ from traditional e-commerce in that they "can provide advice to and make decisions on behalf of customers, input payment credentials and other data on behalf of customers, and execute purchases when customers are not present", and that risks "may increase as greater autonomy is given to AI agents" (Building Trust in Agentic Commerce: Customer-centric principles (six banks, 2026)). The general gradation of agent autonomy is treated at Agent Autonomy Spectrum (5 Levels).
Protocols
The banks describe the focus to date as "on protocols, mostly led by payment networks and technology platforms, and concentrated on product discovery, checkout and payment authorization", and state that "some transactions are operating outside of formal protocols" (Building Trust in Agentic Commerce: Customer-centric principles (six banks, 2026)). The Agentic Commerce Protocol, co-developed by OpenAI and Stripe, was released on September 29, 2025 alongside Instant Checkout in ChatGPT and specifies how an agent passes an order to a merchant that remains merchant of record (Source: openai.com). Google's Agent Payments Protocol (AP2) addresses overlapping ground (Source: metarouter.io). Whether a merchant or platform may exclude agents that do not identify themselves under such standards is contested in *Amazon v. Perplexity* (Brief of the Software & Information Industry Association as Amicus Curiae Supporting Rehearing En Banc — Amazon v. Perplexity (August 28, 2026)).
Deployments
Retail and platform deployments are recorded at Retail — AI Deployment. On May 27, 2026 AWS packaged the architecture behind Alexa for Shopping into a service for retailers to launch their own shopping assistants, which AWS framed as "first-party agentic commerce" (Source: cnbc.com). DoorDash launched a conversational assistant for ordering food and booking reservations on June 11, 2026 (Source: techcrunch.com). The banks' paper states that agentic commerce "has already shifted from closed trials to live pilots and early scaling", while calling the pace and scale of adoption "uncertain" (Building Trust in Agentic Commerce: Customer-centric principles (six banks, 2026)).
Risks named by issuers
The six banks list consumer concerns that agents "may buy the wrong thing or spend too much – or even worse, lose their money to scams and fraud", merchant concerns about disputes and chargebacks "for reasons outside their control", and "potential for higher rates of scams, fraud and disputes". They name as unsafe provider practices requesting card details and entering them directly into websites, prioritising payment methods with lower protections, and not complying with payment scheme rules, and describe the present allocation of liability as "unclear and inefficient". Their five voluntary, non-binding principles (transparency, safety, privacy and data, choice, interoperability) call for disclosure of when an agent is acting and for whom, including sponsored options; consumer control over delegations; authentication that "any party holding a potential liability" can require; auditable records of consumer instructions, intent and transaction decisions; dispute processes with "liability outcomes reflecting where risks and errors are introduced"; and access to agentic services without "gatekeeping fees without reasonable commercial grounds". The authors plan a follow-up paper on implementing the principles through protocols, standards and policies (Building Trust in Agentic Commerce: Customer-centric principles (six banks, 2026)).
The purchase-agent liability question (who is responsible if an agent buys the wrong item) is also raised on Retail — AI Deployment, and the general allocation of responsibility for agent actions at AI Liability. The banks' transparency principle addresses the conflict of interest in agents that "may prioritize products, payment methods or other services that deliver the best financial outcomes for their providers" (Building Trust in Agentic Commerce: Customer-centric principles (six banks, 2026)); see Principal-Agent Problem Applied to AI.
Relationships
- instance-of: Agentic AI — agents acting in payment and purchase flows
- depends-on: Building Trust in Agentic Commerce: Customer-centric principles (six banks, 2026) — definition, autonomy variants and the five principles
- related: Agentic Commerce Protocol (ACP) — the OpenAI–Stripe transaction standard
- related: Agentic Economy — agents as market participants more broadly
- deploys-in: Retail — AI Deployment, Financial Services — AI Deployment
- related: AI Liability, Principal-Agent Problem Applied to AI, Agent Autonomy Spectrum (5 Levels)
- related: Amazon v. Perplexity AI — agent access to merchant platforms
- related: OpenAI, Stripe, Amazon