How it works
What happens to an abandoned cart
Every abandoned cart goes through the same ten steps. Your rules and the maths run in plain code. The AI does two jobs only: pick the offer and write the message. If anything is uncertain, the shopper simply gets a reminder.
The ten steps
In order, every time
Follow one cart from the moment it’s left to the money in your report.
Step 1
Catch the abandoned cart
Your store tells us a cart was left. We check the message is genuine and save the cart.
Nothing is decided from a message we cannot prove came from your store.
Step 2
Split it into a group
One cart in five is set aside and never messaged, so the extra sales can be proven.
Assignment happens before anything is sent, so nobody chooses which carts look good.
20% leave here: saved and watched, never messaged.
Step 3
Check before acting
Already bought? No consent to email? We skip the cart and note why.
The worst message is one sent to someone who already bought.
Skipped carts leave here, with the reason logged.
Step 4
Work out what this cart allows
Your discount cap and free-shipping rules are applied first, in plain code.
The decision model never does arithmetic. It only sees options that are already allowed.
Step 5
Decide the offer
The AI reads the cart and picks the offer most likely to bring the shopper back.
A shopper who was coming back anyway does not need a discount.
Step 6
Gate on confidence
Not confident? The shopper simply gets a friendly reminder instead.
When the model is unsure, it falls back to a reminder, never to a bigger discount.
Step 7
Create the offer
A discount becomes a single-use code that expires, so it can’t be passed around.
The code shows which sales came from the message, and it cannot be shared around.
Step 8
Write the message
A short, personal message in the shopper’s language, with its link and code checked.
A message that is missing its link or code does not get sent.
Step 9
Send and watch
One email goes out. Then we watch for an order for 72 hours.
Both groups are measured the same way, or the comparison means nothing.
Step 10
Report
Every week you see messaged against held-back carts, and the money in between.
You see the same numbers we do, in a format you can check yourself.
Step 5, up close
One call, four questions
The decision model gets the cart and the options your rules allow, and answers together. Each answer comes back with a probability, and every one is logged.
Offer
One choice
Reminder only, free shipping, a small discount, or a related product. Only the options your rules allow for this cart are on the list.
Price sensitive
Yes or no, with a probability
Does price look like the reason they stopped?
Coming back anyway
A score
How likely are they to return and buy without hearing from us?
Related product
One choice from your catalogue
If suggesting something else makes sense, which of your related products fits this cart?
Below the confidence gate, the answer is not used: the shopper gets a plain reminder. Before a pilot starts, the agent’s choices are compared with a store owner’s on fifty sample carts: cheap and expensive, new and returning, one item and many. If it disagrees too often, the questions are rewritten before anything goes live.
The decision step sits behind one function, so it can be swapped for a different model, or for plain rules, without touching anything else.
Safeguards
Tested before any shopper hears from us
Each one has to pass on a test store before a pilot starts.
Buyers are never chased
Someone who completes the order before the send time gets no message. We check again right before every send.
One message per cart
If the same webhook arrives twice, the shopper still hears from us once.
The cap holds
No discount ever goes above the cap you set, even if the model picks the discount option. The limit lives in code.
No consent, no message
A shopper with no email address or no consent is skipped, and the reason is logged.
Held-back carts stay untouched
Held-back carts never get a message from us, so the comparison stays honest.
Every link works
Each message carries a working recovery link, and its code applies at checkout.
Every language renders
Messages are checked in the shopper’s language across Gmail, Outlook and Apple Mail, on desktop and phone.
Outages fail safe
If the decision model or the language model is down, the cart gets a template reminder instead of nothing.
Detection depends on the store platform. Platform details
What are your abandoned carts really worth?
Find out in 4 weeks, free, with a holdout group so the answer is yours to check.