How AI helps recover abandoned carts in Telegram bots and on websites
An abandoned cart is not a lost customer, it is a customer who hit an obstacle. The job of automation is to find out which one, without becoming a nuisance.
Why carts get abandoned here specifically
The reasons on this market differ from the universal lists. Frequently it is unexpected regional delivery cost, the absence of the payment method they use, doubt about stock availability, and plain mobile connection loss during checkout.
Connection loss is an underrated cause. On 3G in the regions a user can lose the session at the payment step and simply never return because they do not know whether the order went through. That customer needs a status update, not persuasion.
Before building recovery flows, examine your funnel and find the step where the bulk of drop-off happens. If it is the delivery step, no follow-up campaign will help — the delivery terms need fixing.
The channel is Telegram, not email
Abandoned cart emails perform poorly in Uzbekistan: mail is used mostly for registrations and work correspondence, and many people check a personal inbox once a week.
Telegram delivers an order of magnitude better open rates, but the irritation threshold is lower too. One message gets read; three in a row get the bot blocked — and that is losing the channel permanently, unlike an email unsubscribe.
If the user came from the website and is not linked to your bot, the only workable options are a web push or a message on their next visit. Demanding a phone number before checkout just to enable follow-ups is a bad trade: cart conversion drops further than you gain.
Where the LLM fits
First scenario: the model composes a personal message based on cart contents rather than a 'you left an item behind' template. Naming the specific product with a relevant detail works noticeably better than an anonymous reminder.
Second and more valuable scenario: the bot asks an open question about the reason and parses the free-form reply. 'Delivery is expensive', 'my size is gone', 'changed my mind' — the model classifies and routes to the right branch: a delivery discount, a back-in-stock alert, or a polite close.
Third scenario: answering objections in real time. If the customer asks about warranty, the bot answers from the knowledge base and returns them to checkout with one button. That is a conversation rather than a campaign, and its conversion is on another level.
Timing and limits
The working pattern is no more than two touches. The first 30-60 minutes after drop-off while intent is alive. The second a day later, addressing the most common objection in your niche, usually delivery.
A third message almost always produces more blocks than orders. If two touches did not work, bring the person back through retargeting rather than direct messages.
Respect the time of day. A message at one in the morning local time reads as spam regardless of content. Set a sending window and queue overnight messages for the morning.
What to measure
Metrics: share of recovered carts, share of opt-outs and blocks, and average order value of recovered orders. The last is usually lower than the original because some customers return with part of the cart.
Always keep a control group that receives nothing. Without it you will credit automation with orders the customer would have placed anyway, which is the most common way to fool yourself in reporting.
Look separately at the abandonment reasons the bot collected. That is the most valuable output of the whole exercise: it shows what to fix in the store rather than how to persuade better.