Order status, returns, exchanges: Celeste on the ecommerce queue

Every ecommerce support queue converges on the same three messages. Where is my order. How do I return this. Can I exchange it for the other size. They arrive at all hours, they spike after every promotion, and each one is a customer holding their money in one hand and their patience in the other.

This is the anchor use case Celeste was built around, and the one where every part of the product shows up in one queue.

Where is my order, answered from the order

Connect your order system, Shopify, your ERP, whatever holds the truth, as a tool, and “where is my order” gets a real answer: shipped Tuesday, this carrier, this tracking number, out for delivery. Not a link to a portal. The answer.

The scoping is the part competitors gloss over. The customer who asks verifies their identity first, and every lookup the agent runs is constrained to that customer’s own orders by code that runs outside the AI. Buyer A cannot retrieve buyer B’s address or order history, and neither can someone spoofing buyer A’s email, because an unverified sender never gets an autonomous reply containing account data. The constraint lives in the tool call itself, where a persuasive prompt cannot reach it.

Policies answer themselves

Return windows, exchange rules, shipping thresholds, warranty terms: you already wrote these. Ingest your policy pages and the agent answers what they actually say, consistently, instead of a paraphrase that varies by rep and by hour. When something is not covered, say a customer asks about a product with no return policy line, the agent routes it to a person and logs the gap by name so the policy gets written once.

Refunds are a decision you make once, deliberately

A refund moves money, and Celeste treats every money-touching tool the way it treats all writes: classified by you before the agent may execute it, held as a drafted action until then. Day one, refund requests arrive pre-worked in your queue: order verified, policy checked, reply written, refund prepared, one click to approve. When weeks of drafts have proven clean, you can classify the routine case, in-window, under a threshold you set, as safe to automate, and leave the edge cases held. The dial moved because you moved it, not because a confidence score felt good that day.

Words you never want automated, “chargeback,” “fraud,” “lawyer,” “never arrived,” go on the escalation keyword list, and a message containing one is barred from an autonomous reply and flagged for your team with the trigger shown.

Practice on your own history first

The agent does not learn on your customers. Point it at your past support mailbox and it studies how your shop actually talks: what you comp, when you apologize, how you handle the size exchange. Then it runs in shadow mode against live traffic, drafting silently while you read what it would have said. Order status usually graduates first, policy questions next, refunds last, each on its own dial.

The volume math is why this is the anchor. The big three are most of the queue in nearly every shop, they surge exactly when your team is busiest, and every one the agent resolves at 2am is a customer who did not spend the next twelve hours composing a one-star review. What is left for your people is the genuinely tangled stuff, arriving pre-triaged with the context gathered.

See how it works for the practice-first model, the product for the queue view, or pricing if you want to run the trial against your own order questions in shadow mode this week.