Designing for Delegation

At Stripe Sessions, a more useful lens emerged: a five-level maturity model for agentic commerce. Stripe deserves credit for turning a broad, often ambiguous conversation into something structured and actionable. Instead of treating AI as a feature, the model reframes it as a progression, from assistance to autonomy.

More importantly, it shifts the goal. The end state is not a better interface. It is a different relationship between user, system, and decision-making. One where intent replaces interaction, and delegation replaces execution.

For digital leaders, this creates something rare: a roadmap.

Traditional commerce systems are built around interaction. Users move through defined journeys— browse, search, filter, add to cart, checkout. Every step requires specific input from the end user.

AI, in its current form, improves these steps. It accelerates search and refines recommendations and it even reduces the number of clicks. But the structure of the flow remains intact, more or less the same as it always was.

Agentic commerce changes the structure itself.

Instead of guiding users through workflows, systems begin to accept intent and act on it. The progression is subtle at first, then profound. Users move from clicking, to instructing, to ultimately delegating.

A helpful way to visualize this shift is as a compression of effort:

• At early stages, systems respond to explicit inputs
• At mid-stages, systems retain context and assist continuously
• At advanced stages, systems act independently within defined guardrails

Each step reduces effort for the user, and increases responsibility for the system.

Stripe’s model outlines a clear progression:

Level 1 - Informational Assistance Systems answer questions: Think chat interfaces, basic copilots, and AI-powered search. Useful, but reactive, and most organizations have at least dabbled here.

Level 2 - Guided Experiences Systems begin to shape decisions: Recommendations improve, flows adapt, and copilots assist with tasks. Most organizations who are doing any AI experimentation and implementation are here today.

Level 3 - Persistent Context The system remembers: Preferences, history, and intent persist across sessions. Interactions become continuous rather than episodic.

Level 4 - Delegated Execution Users grant permission for systems to act: Agents complete tasks, like placing orders, managing subscriptions, and resolving issues, within defined boundaries.

Level 5 - Anticipatory Systems Systems act before being asked: They infer needs, predict intent, and execute proactively. The experience becomes outcome-driven rather than interaction-driven.

Each level removes friction, but it also shifts control and a consistent pattern emerges. As the interface starts to fade, the system takes on more of the decision-making burden.

Despite rapid investment in AI, most enterprises remain anchored in guided experiences.

Why? Structural constraints.

First, systems are fragmented. Commerce platforms, customer data, fulfillment, and service layers often operate independently, and without shared context, persistence is impossible. Second, architectures are largely stateless. They excel at responding to requests, but not at maintaining continuity. Memory—a prerequisite for Levels 3 and above—is missing or bolted on. Third, trust models are underdeveloped. Organizations are comfortable suggesting actions, but not executing them. Delegation introduces risk, and most systems are not designed to manage it. Finally, many teams are still optimizing interfaces rather than rethinking orchestration. Their focus remains on improving journeys instead of redefining outcomes, and that failure of vision can quickly become the biggest limiting factor an org faces.

To progress through the levels you need more than better models, you need a different architectural foundation, and this is where composable systems play a critical role.

APIs become action surfaces, exposing capabilities that agents can invoke. Event-driven architectures enable real-time responsiveness, allowing systems to react to changes as they happen. Identity and data layers evolve to support persistent context, ensuring that memory is not confined to a single session.

And on top of all this, a new layer emerges: orchestration.

Agent orchestration coordinates decisions across systems, balancing autonomy with control. It determines when an agent can act, what constraints apply, and how outcomes are validated. As discussed in our guide to composable orchestration, this layer becomes the control plane for agentic systems.

Importantly, MACH principles provide the foundation, but not the full solution. They enable modularity, but they do not inherently provide memory, reasoning, or delegation. Those capabilities must be designed explicitly.

As systems take on more responsibility, the risk profile changes:

• Who is accountable when an agent makes a decision?
• What constraints define acceptable behavior?
• How are actions observed, audited, and explained?

Forget edge cases, these are central design considerations.

Organizations must introduce guardrails that balance flexibility with control. Observability becomes critical—not just for performance, but for decision transparency—and consent models must evolve to ensure users understand what is being delegated and why. Trust moves out of the category of brand attribute to become a core system capability.

The strength of the five-level model is its practicality, providing a sequence that organizations can map against.

Start by assessing where current capabilities sit. Most will cluster around Levels 1 and 2. From there, the focus should shift to enabling layers:

• Establish persistent identity and memory
• Invest in API coverage and event-driven flows
• Introduce orchestration mechanisms for coordinated action

With those pieces in place, you can confidently pilot system delegation in constrained, low-risk domains. Areas like subscription management, reorder flows, and service resolution are often strong starting points. Remember: the goal is not to jump straight to Level 5. It’s to build the conditions that make higher levels possible. Progression is earned through capability, not declared through ambition.

The real transformation in commerce is not intelligence, but autonomy. Yes, AI will continue to improve interfaces, but the bigger opportunity lies in reducing the need for them altogether. Systems that can understand intent, retain context, and act responsibly will redefine how value is created and delivered.

Stripe’s model provides a useful lens for this shift. It turns a complex future into a navigable path.