Abandoned carts are a people problem

We started LiveRecover this year on a simple observation. Someone puts items in a cart, gets to the point of buying, and stops. The entire industry's response to that is a sequence of automated emails, and the sequence performs about as well as you'd expect from a message that everyone can tell is automatic.

The thing nobody does is ask them what happened.

Why the automation reflex is wrong here#

Abandonment gets treated as a funnel leak, and funnel leaks get patched with automation. That framing is right for most of the funnel and it's specifically wrong here, because abandonment isn't one behavior. It's a dozen unrelated situations that happen to produce the same event.

Shipping cost surprised them. Wrong size and they weren't sure. Genuinely just browsing. Got distracted. Wanted to check whether there's a code. Payment failed and they didn't understand the error. Comparing against another store. Wanted to ask a question and there was nobody to ask.

An automated sequence responds to all of these identically, because it doesn't know which one it's talking to. So it does the only thing available to it: reminds them the cart exists, and eventually offers a discount. Which means the tool's answer to "I was worried about the return policy" and "I was going to buy anyway" is the same, and one of those is money you gave away for nothing.

A person can find out which situation it is by asking, and then do the thing that situation needs. Half the time the thing that situation needs costs nothing.

Why SMS#

Email is where recovery lives and it's the wrong channel for this. Open rates are what they are, the message lands in a pile of other commercial mail, and the reply path is theoretical — nobody replies to a marketing email and expects a human.

Text messages get read almost immediately, and more importantly the format carries a social expectation of a reply. When you get a text you either respond or you deliberately don't. It's a conversational medium, and we're trying to have a conversation.

The obvious risk is that it's intrusive. Getting this wrong is worse in SMS than anywhere else, because the channel is personal enough that abusing it makes people genuinely angry rather than mildly annoyed. That constraint is real and I think it's healthy — it forces the message to be worth sending.

The part everyone tells us is a mistake#

We use real people. Not a chatbot with a friendly name. Human beings reading the message and writing back.

Every technical person who hears this tells us it won't scale, and they're describing a real problem rather than a hypothetical one. Humans cost money per conversation. Software costs money once. The unit economics of our approach are structurally worse than the competitor who automates, and that gap doesn't close with volume — it widens.

Two reasons we're doing it anyway.

It works better, and by enough to pay for itself. A conversation that resolves an actual objection converts at a rate that a reminder sequence doesn't approach. If the difference is large enough, the labor is affordable.

It's a research position. Every conversation is a merchant learning why their customers don't buy, in the customer's own words. That information doesn't exist anywhere else in ecommerce. Right now it's a byproduct. I suspect it's worth more than the recovery.

What I don't know yet#

Whether the labor model survives growth, or whether we end up automating the easy half and keeping humans for the rest. I'd guess the second, and I'd rather arrive there by finding out which conversations actually need a person than by assuming from the start that none of them do.

Everyone in this category is starting from the assumption that none of them do. That's the whole bet.