What does the agent need before it can answer?
Real order data, policy context and internal tools. The agent is live on the storefront and combines those sources rather than relying on a scraped help page.
Routine questions include order status, returns eligibility and exchange rules.
How does a request reach an answer?
The agent combines order lookup, policy context and internal tools. Low-risk requests close on their own. Sensitive ones go to the team.
Claude API, tool calling, Shopify and the WMS are the core stack.
Does an AI agent get to move orders or money?
No. Agents answer tickets. Anything that moves orders, inventory or money is deterministic code with hard rules.
Money decisions and sensitive cases go to a person, every time. The agent has a bounded job.
What changes for the support team?
Fewer tickets a human ever has to touch. Low-risk requests can close on their own, while sensitive requests still reach the team.
The useful distinction is which requests can finish safely and which require a person.
What is running in production?
Fewer tickets a human ever has to touch.
Have the same handoff?
We build automation and AI agents for ecommerce teams whose orders move through a real warehouse.