Stripe's Five Levels of Agentic Commerce

Stripe's model defines five levels of increasing agent autonomy: Level 1 (Eliminating Web Forms), Level 2 (Descriptive Search), Level 3 (Persistence), Level 4 (Delegation), and Level 5 (Anticipation). This progression, illustrated with a consistent back-to-school shopping scenario, effectively maps the technological capabilities required for agents to take on more complex commercial tasks. The framework correctly identifies that the industry is currently at the nascent stages (Levels 1-2) and rightly emphasizes that progress depends on interoperable protocols, such as their proposed Agentic Commerce Protocol (ACP). However, this capability-focused ladder reveals a profound transaction bias, viewing commerce from the checkout perspective rather than the human relationship that precedes it.

This bias leads to critical omissions that are the central concern of agentic experience design (AXD). The first is the absence of a robust trust architecture. Stripe's model treats trust as a given precondition for delegation, mentioned only once, rather than as the primary, dynamic material that must be designed, calibrated, and repaired. It ignores that a user's willingness to grant autonomy is not static but fluctuates based on context, stakes, and past performance. The model provides no language for designing the delegation design itself - the granular, revocable permissions that govern an agent's authority. A user might permit Level 4 autonomy for groceries but demand Level 1 for electronics, a contextual reality the linear model cannot accommodate.

Furthermore, the framework presents a success-only narrative, completely omitting a failure architecture. It describes what happens when the agent purchases wisely but not what happens when it fails - buys the wrong size, misses a deadline, or violates an unstated preference. In agentic systems, where actions are taken in the user's absence, designing for failure, recovery, and trust repair is not a secondary concern but a primary one. The experience of an agent failing at Level 5 (Anticipation) is not a simple error; it's a potential violation of trust and financial stability that demands a sophisticated, pre-designed response, a concept central to Absent-State Design.

If Stripe's capability-focused model is adopted without the complementary human-centric frameworks of AXD, the development of agentic commerce will stall. Product leaders and designers must recognize that capability is not permission. Building a Level 4 agent is a technical achievement, but getting a human to trust it is a design challenge. This requires shifting focus from a linear progression of capabilities to a dynamic matrix of trust and autonomy. The primary design surface is not the transaction; it is the relationship, governed by a robust trust architecture that includes clear delegation design, transparent agent observability, and comprehensive failure architecture.

Practitioners should use the Five Levels as a shared vocabulary for technical maturity but build their products around the principles of AXD. This means designing for trust regression, not just progression, allowing users to fluidly increase or decrease an agent's autonomy based on context and confidence. It means that for every new capability an agent gains, a corresponding set of guardrails, oversight mechanisms, and recovery paths must be designed. The future of agentic commerce will be defined not by the raw power of autonomous agents, but by the quality of the trust relationship we design with them. The most successful systems will be those that match the level of autonomy to the level of trust, in real time, for each specific context.