How to Auto-Segment Customers Using AI and Klaviyo

Let Miraxa analyze customer behavior and automatically assign segments in Klaviyo with Spojit.

What This Integration Does

Manual segmentation goes stale fast. A customer who was a VIP three months ago may now be at-risk, and a one-time buyer might be quietly turning into a regular. This workflow runs on a schedule, uses a Connector node in Agent mode to classify each profile against your taxonomy (VIP, new, at-risk, bargain-hunter, loyal, lapsed), and writes the result back to Klaviyo as a custom property your flows and segments can use.

Each run is incremental - it scopes to profiles updated since the last run, classifies them, and upserts the segment property. Because the property is a single string, downstream Klaviyo segments stay simple: one filter on ai_segment equals vip.

Prerequisites

A Klaviyo connection with profile read and write access.
A connection to your store of truth for purchase history (shopify, woocommerce, or your data warehouse).
A written segment taxonomy with definitions (what makes a profile a VIP vs. loyal vs. at-risk).

Step 1: Schedule Trigger

Add a Trigger node set to Schedule. Daily at 03:00 is a sensible default - the segments are stable enough that hourly is overkill, and daily means flows and campaigns always see fresh tiers in the morning.

Step 2: Pull Profiles to Score

Add a Connector node on klaviyo and use list-profiles. Filter to profiles updated in the last 24 hours, or paginate through everyone on a full weekly re-score. Pair this with shopify list-customers or woocommerce list-customers if you need order history fields that are not already mirrored to the Klaviyo profile.

Step 3: Loop and Compute Features

Wrap the next steps in a Loop. Inside the loop, add a Transform node that computes the input features for the AI:

orderCount
lifetimeValue
daysSinceLastOrder
averageOrderValue
discountUsageRate

This is what keeps classification consistent - the agent gets the same shape of input for every profile.

Step 4: Classify the Profile in Agent Mode

Add a Connector node on klaviyo and switch it to Agent mode. Agent mode lets the node reason over the features and decide on a segment, and it supports a Response Schema so the output comes back as structured JSON. Set the Response Schema to:

{
  "segment": "vip | loyal | new | at-risk | bargain-hunter | lapsed",
  "lifetimeValue": "number",
  "nextPurchaseLikelihood": "high | medium | low",
  "reasoning": "string"
}

Prompt: Classify this customer using the taxonomy in the prompt. Features: {{ features }}. Be conservative - prefer 'loyal' over 'vip' unless the customer is clearly in the top 5 percent of lifetime value.

Step 5: Update Klaviyo Profile

Add a Connector step on klaviyo using update-profile. Set:

ai_segment - the chosen segment string
ai_ltv - lifetime value
ai_next_purchase - high, medium, or low
ai_segmented_at - timestamp

You can also call add-profiles-to-list to push the profile into a Klaviyo list named after the segment, so your flows can listen on list membership.

Step 6: Log Movement for Reporting

Add a final Connector node that writes the (profile id, old segment, new segment, timestamp) tuple into a mongodb collection. Use insert-documents. This makes "who moved from at-risk to lapsed this week" trivially queryable, which marketing teams love.

Tips

Re-score the full base monthly even if you run daily incrementals - definitions drift and a periodic full pass keeps the segments honest.
Cache feature computation results in mongodb so you do not recompute lifetime value from scratch every run.
Start with 3 segments, not 8 - it is easier for the AI and for your campaign team.

Common Pitfalls

Segment churn - if customers flip between VIP and at-risk every run, your prompt or features are too sensitive. Add hysteresis (require two consecutive runs in the new segment before flipping).
Rate limits - Klaviyo has per-endpoint quotas. Batch profile updates and stagger if you are scoring tens of thousands.
Missing data - profiles with zero orders confuse the classifier. Pre-filter those into a fixed new segment before the Agent mode step.

Testing

Run the workflow manually against a list of 5-10 hand-picked profiles whose segment you already know. Inspect the reasoning field in the Agent mode output - it should match your expectations. Adjust the prompt and re-run until it does, then turn the schedule on. You can also ask Miraxa to review a failed run or refine the classification prompt in context.