storepilot · automations for shopify, with a brain

001 — the thesis

The automations layer Shopify Flow should have been. With an LLM instead of if/then.

Build workflows that trigger on Shopify events, reason with an LLM, write back to the store, and stop for approval before anything risky ships. Every step, every tool call, logged.

002 — anatomy

A workflow, piece by piece.

Every workflow has one trigger and an ordered list of steps. Each step can be an action, a condition, an LLM reasoning call, or a proposal gate.

1triggerorders/create
2llm.reasonscore fraud risk → {fraud_score, rationale}
3conditionfraud_score > 0.6
4proposal.gatetag customer 'fraud-review' (needs approval)
5notify.emailalert ops@company.com

003 — what people build

Things you cannot build in Flow. Not easily, not at all.

trigger: orders/create

Fraud triage on every order

An LLM reads the order, scores it, routes anything above a threshold to a review queue. Rules can't read a shipping address like a human can — this does.

trigger: schedule · 24h

Daily restock plan with approvals

Pulls low-stock items, asks an LLM for reorder quantities based on 30-day velocity, drops each recommendation in your Proposals inbox.

trigger: review created

Auto-reply bad reviews, propose the refund

Drafts a tone-matched response, proposes a discount code or refund amount, routes the approval to whoever's on ops rotation.

trigger: schedule · weekly

Slow-mover discount generator

Finds products with 45+ days of stock, reasons about competing SKUs, proposes a code + a Klaviyo email segment to target.

trigger: orders/create

Tag VIPs the moment they cross $500

Simple, but one line of YAML in Flow can't do the downstream: personalized thank-you email and a one-time discount generated per customer.

trigger: inventory_levels/update

Restock-alert loop with human in the loop

When a SKU crosses its low threshold, the workflow emails purchasing with an LLM-drafted PO. The PO only sends after approval.

004 — versus everything else

We read the room. Then we did something different.

Shopify Flow

rule-based (if X then Y). Can't reason. No LLM steps.
LLM steps output structured JSON downstream steps consume.

Shopify Sidekick

conversational. Acts once, then you're done. No schedule, no state.
persistent workflows with triggers, retries, audit log, history.

Zapier/Make with GPT blocks

generic. No Shopify-aware actions. No approval queue.
first-class Shopify Admin API writes. Proposals inbox for anything risky.

005 — trust

Autonomous where you want it. Supervised where it counts.

Proposals inbox

Mark any step as 'requires approval'. The workflow stops, an approval lands in your inbox with the resolved action config, and only then does it run. Reject and the run continues with a skipped step.

Versioned workflows

Every save bumps the version; runs are pinned to the version that executed them. You can see exactly what rule produced a change a month ago.

Tamper-evident audit log

Every write an agent makes lands in audit_log with the actor, tool, args, and result. Your ops team gets a real paper trail.

006 — next

Hook up a Shopify store, pick a template, ship a workflow.