An agent that can answer a product question does not automatically have permission to act. Machine Interface design should separate retrieval from state-changing operations.
What is a read interface?
A read interface exposes facts without changing commerce state. Examples include retrieving product attributes, checking variant availability, reading policies, comparing offers or obtaining a delivery estimate.
Read access still requires governance. Customer, order or operational data should not be exposed merely because an AI system can use it. The interface should provide the minimum information required for the approved purpose.
What is an action interface?
An action interface allows an authorized system to perform a bounded operation: choose a variant, add an item to a cart, apply an eligible offer, create a draft artefact, reach checkout or escalate to a human workflow.
Why action contracts must be explicit
Every action needs defined inputs, preconditions, permissions, expected result, failure states and an audit trail. A request such as “buy the cheapest suitable option” contains both reasoning and action. The agent may compare products autonomously, but payment or another high-impact step may require direct customer confirmation.
Progressive authority
A practical architecture grants capability in stages:
- Discover and retrieve public commerce information.
- Compare and recommend with stated uncertainty.
- Prepare a reversible action such as a cart.
- Request confirmation at the appropriate boundary.
- Execute only the specifically authorized operation.
- Return a clear result or recoverable failure.
Failure handling is part of the interface
“Something went wrong” is not enough for an agent. Errors should distinguish unavailable inventory, invalid variants, expired offers, address restrictions, authentication failure and permission denial. This enables the agent to recover safely or escalate.
Human review and controlled automation
Codified Commerce treats assessment and recommendation as separate from implementation. A machine-readable opportunity does not authorize a storefront, price, policy or campaign change. CAT governs approved implementation through Define → Decompose → Orchestrate → Execute → Validate → Learn.
Where MI fits in the Codified Commerce system
Machine Interface exposes information and capabilities. Machine Experience evaluates whether machines can understand and use them reliably. Machine Learning interprets the interaction and outcome signals. Moving Target detects meaningful change, while Next Product Move determines what the merchant should consider changing.
Continue with Machine Experience, Machine Learning for Shopify, Moving Target, and Next Product Move.
Explore the Machine Interface cluster
- What Is a Machine Interface in Agentic Commerce?
- Machine Interface vs Machine Experience: What’s the Difference?
- The Components of a Shopify Machine Interface
- How AI Shopping Agents Access Product, Price and Inventory Data
- Read Interfaces vs Action Interfaces in Agentic Commerce
- Structured Data, APIs, Feeds and Agent Tools: How They Work Together
- Machine Interface Failures: Stale Data, Ambiguity, Permissions and Errors
- How to Audit a Shopify Store’s Machine Interface
Explore Machine Experience →
Discuss a controlled Shopify automation →
