- Sidekick turned Flow into a prompt, not a drag-and-drop grid. That opens Flow-building to anyone who can describe an automation in a sentence.
- Nearly half of Q1 2026 Flows came from Sidekick per Shopify's earnings, and Flow v4 added 30-day Wait steps plus LLM actions like Summarize and Categorize.
- Use it for order-triggered ops. Do not use it for behavioral events, multi-channel journeys, or lifecycle marketing.
Shopify Flow used to require dragging trigger, condition, and action blocks by hand. Now Sidekick generates Flow automations from a plain-English prompt. Here's how to use Shopify Flow AI to create workflows in the current Sidekick interface, what Flow v4 actually adds, and the point where store-ops automation stops being enough for lifecycle marketing.
The rest of the post covers what changed in 2026, the Sidekick prompt-to-Flow loop step by step, when Flow fits versus when it doesn't, the honest limits of the LLM actions Flow v4 added, and the boundary between store operations and lifecycle marketing that no version of Flow has ever tried to cross.
What changed in 2026
Two shifts in the same year moved Flow from a power-user tool into something a non-technical store operator can drive. Sidekick brought prompt-to-automation to the front door, and Flow v4 quietly closed some of the gaps that used to push teams to an external scheduler or a third-party app.
- Shopify fine-tuned a tool-calling agent (Qwen3-32B) that generates Flow automations from natural language, 2.2x faster and 68% cheaper than the frontier model it replaced, per Shopify Engineering's 2026 post.
- Nearly half of all Shopify Flows generated in Q1 2026 came from Sidekick, not the manual builder, per Shopify's Q1 2026 earnings call.
- You create a Flow with Sidekick by opening Flow, clicking the Sidekick icon, and describing the automation in plain English, per the Shopify changelog.
- Flow's Wait action now supports a maximum duration of 30 days, per the Shopify changelog, enabling delayed follow-ups that used to require an external scheduler.
- Flow v4 added LLM actions (Summarize, Categorize) as steps inside a workflow, so Flow can now read unstructured text from order notes, customer messages, or reviews and act on the result.
The practical effect of the Sidekick shift is who can now build a Flow automation, not just how fast. The old visual builder required understanding trigger/condition/action logic well enough to wire it correctly, which kept Flow-building concentrated among ops-savvy store admins. A prompt-based agent that gets the logic right most of the time opens that same automation up to anyone who can describe what they want in a sentence. That's a real access shift, not just a speed one, and it's the actual reason nearly half of Q1 2026's Flows came from Sidekick rather than the manual builder - the bottleneck wasn't the tool, it was who could operate it.
How to use Shopify Flow AI to create workflows: step-by-step
The Sidekick-driven build loop is short. Open Flow, describe what you want, review the generated shape, adjust in the visual editor, test it once, turn it on. Every step has a place a first-time user commonly gets tripped up, so the numbered list below is the loop with the traps called out.
- Open Shopify Flow from your admin. Flow is included free on all Shopify plans, so if it's not already installed, add it from the Shopify App Store first.
- Click the Sidekick icon in the Flow editor to open the AI prompt bar. This is the entry point for AI-generated workflows - the manual visual builder is still available in the same view if you want to fall back to it.
- Describe the automation in plain English. Be specific about the trigger event, the condition, and the action. Example: 'When an order over $500 is placed by a first-time customer, tag it VIP-first-order and send me a Slack notification.' Vague prompts produce vague flows.
- Review the generated workflow. Sidekick returns a Flow with the trigger, conditions, and actions filled in. Check that the trigger event matches what you meant, the conditions evaluate in the right order, and the actions target the correct fields (order tags vs. customer tags, notification channels, etc.).
- Adjust anything wrong in the visual editor. Sidekick isn't one-shot correct on every prompt, and the visual editor is still where you clean up branches, add a Wait step (up to 30 days), or slot in a Flow v4 LLM action like Summarize or Categorize.
- Turn the workflow on and test it against a real event before leaving it running unattended. A misfired ops flow at scale is a customer-experience incident, not a small bug.
Prompt patterns that actually work
The difference between a good and bad Sidekick prompt is specificity in three places: the exact trigger event, the concrete condition, and the named action target. A prompt that names the field, the threshold, and the destination channel gets a Flow that runs on the first pass more often than not.
| Prompt pattern | Example | Result |
|---|---|---|
| Trigger + condition + action (all named) | "When an order over $500 is placed by a first-time customer, tag it VIP-first-order and send me a Slack notification in the #orders channel" | Sidekick usually returns a complete, correct flow |
| Trigger only, condition vague | "When a big order comes in, do something with it" | Sidekick guesses at threshold and action; expect to edit |
| Multi-step with wait | "When a draft order sits unpaid for 3 days, email the customer a reminder and tag the order as at-risk" | Uses the new 30-day Wait step correctly |
| Uses an LLM action | "When a review is posted, summarize it into a two-line internal note and categorize sentiment" | Sidekick wires Summarize + Categorize actions inline |
Common places Sidekick misses
- Wrong field for the tag: it tags the order when you meant the customer, or vice versa. Two clicks to fix in the visual editor.
- Branch order: if the prompt implies a nested condition, Sidekick sometimes flattens it. Review the branch tree before turning on.
- Wait step defaults: Sidekick picks 24 hours when the prompt doesn't state a duration. Set it explicitly.
- Notification channel: it defaults to email; if you meant Slack, name the channel.
- LLM action model choice: the Summarize action has a default output length that often needs shortening for internal notes.
When Flow fits, when you need a lifecycle tool
This is the boundary the rest of the post is about. Flow is a great store-operations tool. It is not a lifecycle marketing tool, and Shopify has never claimed it is. The table below is the honest split by trigger, action count, channel, branching, and the state each side needs to hold.
| Question | Shopify Flow fits | Lifecycle marketing tool fits |
|---|---|---|
| Trigger type | Store event (order placed, inventory threshold, tag applied) | Customer behavior over time (signup, browse abandonment, days since last purchase) |
| Action count | One action or a short branch | Multi-touch sequence across days or weeks |
| Channels | Store-internal actions (tags, notifications, fulfillment) | Email, SMS, push, coordinated across channels |
| Branching logic | Multi-branch conditions for ops decisions | A/B-tested branches tied to marketing intent and audience |
| State it needs | The current order or product | Full customer profile, history, and segment membership |
| Timing model | Fires on the event | Fires on the event, then waits, then adapts based on what the customer did next |
A quick test for which job you actually have
Three questions settle the tool choice faster than comparing feature lists. The test is deliberately not about capability - Flow v4's LLM actions can technically write an email, but the test isn't "can it" but "is this the tool designed for this shape of work."
- Does the trigger fire on a store event (order placed, inventory threshold, tag applied) or a customer-behavior event (signup, browse abandonment, days since last purchase)? Store event = ops, Flow's job. Behavior event = lifecycle marketing, a different tool's job.
- Does the automation end in one action, or does it need to branch and re-engage over days or weeks based on what the customer does next? One action = ops. A multi-touch sequence with branches = lifecycle marketing.
- Does it need to run across channels (email and SMS together, coordinated) or stay inside store-internal actions? Cross-channel = lifecycle marketing, not something Flow's operations scope covers regardless of how capable its AI gets at ops tasks.
What Flow v4's LLM actions are (and aren't) for
The Summarize and Categorize actions inside Flow v4 are genuinely useful for a narrow job: reading unstructured text that already lives in the store data (order notes, customer messages, reviews) and turning it into something a downstream Flow step can act on. That's a real gap Flow closed. What it isn't for: generative writing that goes to a customer.
| Use case | Fits the LLM action | Belongs in a lifecycle tool |
|---|---|---|
| Summarize a long order note into a two-line internal comment | Yes | |
| Categorize a review as positive/neutral/negative and tag the profile | Yes | |
| Extract a shipping instruction from a free-text order note | Yes | |
| Write a personalized email body to send to the customer | Yes | |
| Choose which of three creative variants to send based on segment | Yes | |
| Generate ad copy based on the SKU | Yes |
The rule of thumb: if the output is read by an internal user (an operator, a tag, another Flow step), the LLM action is fine. If the output is read by a customer, it belongs in a tool designed for variant testing, hold-out groups, and QA at scale.
The failure mode nobody warns about
Shopify Flow AI does not have a testing environment separate from production. Sidekick generates a Flow, you turn it on, and it runs on the next real event. That is fine for a low-volume workflow. It is a customer-experience risk on a high-volume ops flow, especially one that touches customer-facing fields like discount codes, order tags visible in the customer account, or notification channels the customer sees.
- Duplicate the workflow into a staging state (turned off) before editing a live one - Flow keeps a version history, but Sidekick edits inline.
- Test with a single filter first (e.g. add a condition that fires only on your own test account) before removing the filter and letting it run at store scale.
- Add a Slack or email notification on the first three real firings so a human sees the first outcomes before the flow runs quietly.
- For anything touching discount codes, refunds, or customer-visible tags, treat the first week as observation-only - not fire-and-forget.
What Shopify Flow AI does not do
The honest boundaries, so the tool gets deployed where it earns its keep and not where it's guaranteed to disappoint.
- It does not run cross-channel journeys. Email plus SMS plus push, coordinated on one customer identity, is lifecycle marketing infrastructure, not Flow.
- It does not hold customer state across time in the way a marketing platform does. Flow reacts to events; it doesn't maintain a segment membership that changes as the customer behaves.
- It does not test creative variants. There's no A/B holdout inside Flow.
- It does not attribute revenue back to a campaign. Attribution lives in analytics, not ops.
- It does not replace the manager judgment on whether the workflow should exist. Sidekick will happily generate a Flow that misfires quarterly if you ask it to.
- It does not test itself. There is no separate staging environment, which is a real operational risk on high-volume flows.
Most teams end up running both, Flow for the operational plumbing, a lifecycle tool for the customer-facing journeys, because they're genuinely different jobs rather than the same job at two maturity levels. Knowing how to use Shopify Flow AI to create workflows quickly is the easy half - knowing which of those jobs a given automation belongs to is the harder half, and the one worth getting right.
The short version
Shopify Flow AI genuinely changed who can build a Flow. Sidekick's prompt-to-Flow loop is fast, half of Q1 2026's Flows came out of it, and Flow v4's 30-day Wait plus LLM actions closed a few of the gaps that used to force an external tool. What the new capability does not do is turn Flow into a lifecycle marketing tool - the boundary between store operations and customer-behavior journeys is a real one, and running both is the right answer for any store that outgrows the ops-only stage.
Frequently asked questions. Answered.
Shopify Flow AI is the Sidekick-powered layer over Shopify Flow that generates automations from a natural-language prompt. You describe the workflow ("when an order over $500 is placed by a first-time customer, tag it VIP-first-order and send me a Slack notification") and Sidekick returns a Flow with trigger, conditions, and actions filled in. It's built on a Shopify-fine-tuned Qwen3-32B agent per [Shopify Engineering's 2026 post](https://shopify.engineering/fine-tuning-agent-shopify-flow).
