Most churn prediction articles talk about saving customers who are already gone. They describe win-back emails and discount campaigns targeting people who have not bought in 90 days. By the time you send those emails, the customer has already found a replacement.
The real value of churn prediction is catching customers before their first reorder window closes. If a customer usually buys every 45 days and day 50 passes with no purchase, that is the moment to act. Not day 90.
AI can build a churn risk score for every customer that updates automatically. Here is how to set one up using data you already collect.
What a Churn Risk Score Actually Is
A churn risk score is a number from 0 to 100 assigned to each customer. A score of 85 means the customer is highly likely to stop buying. A score of 15 means they are engaged and likely to reorder.
The score is not a guess. It is calculated from behavioural signals that correlate with churn:
- Days since last purchase compared to their average purchase interval
- Order frequency trend whether purchases are speeding up or slowing down
- Average order value trend whether spend is shrinking
- Email engagement whether opens and clicks are declining
- Site visit pattern whether browse sessions are getting shorter
- Support contacts whether complaints or questions are increasing
Each signal contributes points to the score. The AI model weights them based on which signals actually predict churn in your store, not a generic benchmark.
Step 1: Export Your Customer Data
You need a spreadsheet with one row per customer and columns for:
- Customer ID
- First order date
- Most recent order date
- Total number of orders
- Average days between orders
- Average order value
- Last 3 order values (to spot declining spend)
- Email open rate (last 30 days)
- Email click rate (last 30 days)
- Number of support tickets (last 90 days)
- Days since last site visit
Most e-commerce platforms export this directly. Shopify gives you most of it through customer reports. WooCommerce has customer export plugins. If you use Klaviyo or Mailchimp, pull email engagement from there.
Step 2: Feed the Data to an AI Model
You do not need a data science team for this. You need a spreadsheet and an AI tool that can run a simple classification model.
Upload your customer data to a tool like Google Sheets with a prediction add-on, or use a no-code AI platform. The model needs to learn one thing: given these signals, did similar customers in the past go on to reorder or not?
To train it, label your historical customers:
- Churned if they have not ordered in 2x their average interval
- Active if they ordered within their average interval
- At risk if they are between 1x and 2x their average interval
The model learns which signal patterns precede churn. For example, it might find that customers whose email open rate dropped 40% and whose last order was 20% smaller than usual are 3x more likely to churn.
Step 3: Generate Risk Scores for Every Customer
Once the model is trained, run it on your current customer base. Each customer gets a score from 0 to 100.
Sort your customer list by score, highest first. The top 20% are your churn risks. These are the customers who need attention this week, not next month.
Export the scored list and segment it into three buckets:
- Red (70-100): High risk. Act now. Personal email from the founder. Not a discount.
- Yellow (40-69): Moderate risk. Trigger a re-engagement email sequence. Offer a small incentive.
- Green (0-39): Low risk. Leave them alone. Do not spam engaged customers with win-back emails.
Step 4: Automate the Score Updates
Manual scoring works once. To catch churn before it happens, the score needs to update weekly. Set up an automation that:
- Pulls fresh customer data every Sunday night
- Runs the model to recalculate scores
- Flags any customer whose score jumped more than 20 points since last week
- Sends you a summary email with the new red-flag customers
This way you are not checking scores manually. You get a Monday morning email that says: 3 customers moved into high-risk this week. Here are their names and what changed.
Step 5: Act on the Scores
The score is useless without action. For each risk bucket, have a specific response:
Red customers: Send a personal email from the founder. Not a template. Reference their last purchase. Ask if something went wrong. Offer to help, not to discount. A customer who feels heard is more likely to return than one who gets a 10% off code.
Yellow customers: Trigger a 3-part email sequence. Email 1 shares a relevant product tip. Email 2 shows a customer story. Email 3 offers a small incentive if they have not reordered by then.
Green customers: Do nothing different. Continue your normal email cadence. These customers are healthy. Over-contacting them can push them toward churn.
What the AI Model Actually Finds
Every store is different, but common patterns emerge:
- Customers who go from opening 80% of emails to opening 20% are 4x more likely to churn within 60 days
- Customers whose third order was 30%+ smaller than their first two are signalling declining interest
- Customers who submitted a support ticket in the last 30 days and have not reordered are at high risk if the ticket was about a product quality issue
- Customers who used to visit the site weekly and now visit monthly are disengaging even if they have not stopped buying yet
The AI surfaces these patterns automatically. You do not need to guess which signals matter. The model tells you.
How Long This Takes
The first build takes a weekend. Exporting and cleaning customer data takes 2 hours. Training the model takes 1 hour if you use a no-code tool. Setting up the automation takes 2 hours. Designing the email responses for each risk bucket takes another 2 hours.
After that, the system runs itself. You check the Monday email. You send personal notes to red customers. The yellow sequence sends automatically. Your churn rate drops because you are catching people before they leave, not after.
Want more like this? Join MOKU Club for free. Weekly resources, early access to new guides, and occasional templates you can actually use. Join below.
