Shopify actions

Real actions. Hard guardrails.

Most support AI can only talk about a refund. Suvenna executes it against Shopify, through a guardrail stack that makes the unsafe path impossible, not just discouraged.

What it can do

The four actions that close tickets

Order status, refunds, cancellations, and returns cover the bulk of what a Shopify support queue actually asks for. Suvenna handles each one end to end.

Order status

“Where is my order?” answered from the customer's real Shopify order: fulfillment state, carrier, tracking. Lookups are bound to the verified customer's email, so nobody can fish for someone else's order.

Refunds

Full or partial refunds executed against Shopify, bounded by the order's actual refundable amount at the moment of execution. If a prior refund already reduced what's left, the bound reflects it.

Cancellations

Cancels the order in Shopify, refunds the original payment method, and restocks the items. The customer confirms first; your rules decide whether a human signs off too.

Returns

Opens a real return in Shopify with an RMA, so your warehouse sees it in the same system as everything else. No side spreadsheets, no “just ship it back” emails.

The guardrail stack

Four layers between a request and your money

Every money action passes through all four. The stack is enforced in the engine itself, not in a setting someone can flip off.

Layer 1

The customer confirms

Before anything moves, the customer explicitly confirms the exact action and amount in plain language. “Refund $42.50 to your original payment method, correct?” No confirmation, no action.

Layer 2

Your rules decide: auto or approve

You set the policy. A refund under your auto-approval ceiling can proceed on its own; anything above it, or anything your automation rules flag, is held for a human. The AI never decides its own permissions.

Layer 3

One-click human approval

Held actions land in the approval queue with the order, the amount, and the full conversation. Your team approves or rejects in one click. Nothing risky executes without a person choosing to let it.

Layer 4

Execution bounds and the audit log

At execution time the engine re-checks the bound against Shopify: a refund can never exceed the order's refundable amount, whatever was said earlier in the thread. Every action, approved or rejected, is written to an append-only audit log.

The approval queue

Approve with the whole picture

Each pending action shows the order, the amount, and the conversation that led to it. One click executes it against Shopify and writes the audit entry.

The Suvenna approval queue: a pending refund with the order details, amount, and conversation context beside one-click approve and reject buttons

Off by default

Live actions ship off for every store.

Autonomy is opt-in, per store. Until you flip a toggle, Suvenna drafts and proposes but never touches an order. Turn on exactly what you trust, when the QA scores earn it.

Questions

Guardrails, in detail

Can the AI refund more than the order is worth?

No. At execution time the engine checks the order's actual refundable amount in Shopify and refuses anything above it. That bound is enforced in code, on every refund, and cannot be disabled by any setting.

What stops a customer from asking about someone else's order?

Order lookups are bound to the verified customer's email. On email, that's the sender of the thread. In the chat widget, the shopper's identity is verified by Shopify itself. If there's no verified identity, order tools simply don't run.

Do I have to let the AI act on its own?

No. Every live-action toggle ships off, and even when you enable one, you choose the split: actions under your ceiling can auto-approve, everything else queues for one-click human approval. Most stores start with everything queued and loosen from there.

Get priority access

Suvenna onboards in limited waves. Join the list and you're first in line for the next one, with founding-member pricing locked in.

Request early access See pricing