Almost nobody uses e-commerce chat assistants.
I build them, and that is still the observation to start from. The chat bubble sits in the corner of half the stores on the internet, and on most of them it is furniture. Shoppers scroll past it the way they scroll past cookie banners.
This is the second time the industry has been here. In 2016, conversational commerce was declared the future: messenger bots, bot stores, “chatbots are the new apps.” By 2018 the wave was over, and the large platform assistants of that era had quietly shut down or scaled back. Now LLMs have restarted the claim. If we don’t understand why the first wave failed, and why adoption is still low in the second, we will fail the same way with better technology.
The mechanism of the first failure
Here is a 2017 shopping conversation, reconstructed from a thousand real ones like it:
Customer: Will this tent handle a Scottish autumn?
Bot: Sorry, I didn’t understand. Please choose: Shipping Returns Track my order Customer: waterproof?
Bot: Sorry, I didn’t understand. Please choose: Shipping Returns Track my order Customer closes the tab.
The 2016 bots were decision trees wearing a conversational costume. The surface accepted any sentence; the machinery matched it against a handful of scripted branches. The gap between those two things is where trust died.
Researchers had documented the dynamic before the commerce wave even crested. Luger and Sellen’s CHI 2016 study of conversational agent users, titled “Like Having a Really Bad PA”, found user expectations “dramatically out of step” with what the systems could do. A conversational interface implicitly promises understanding. When the machinery cannot honor the promise, the user does not recalibrate; they leave, and they remember.
That memory is the industry’s real inheritance. Every chat bubble today pays an expectation tax levied by the bots of 2017.
There was a second, quieter failure: the bots that worked were measuring the wrong thing. “Deflection” (tickets not reaching a human) became the metric, because it was easy to measure and easy to sell to a support budget. But deflection is a cost metric. Nobody was measuring whether the conversation sold anything, because nothing about a scripted tree could sell. Conversational commerce failed partly because almost none of it was commerce.
What each tool actually is
Strip away the category marketing and the three tools are different machines.
Live chat is a staffing model, not software. It sells well because a competent human runs the full discovery loop: asks, retrieves from memory, explains, adapts. Its constraint is arithmetic. One agent, one shift, one language, a few conversations at a time, while shoppers browse at 11pm across every market you ship to. Live chat does not fail at conversation; it fails at coverage.
Scripted bots are lookup tables with a text field. Wired to an order system, they answer “where is my order?” cheaply and always will. That is honest, useful deflection work. It has nothing to do with selling.
LLM-based assistants are the first machines that can actually run the discovery loop: understand an unformulated need, retrieve matching products, explain fit, refine. What they add is coverage of the giant middle ground live chat cannot staff: open-ended product questions at any hour. What they risk is a new failure mode the scripted bots never had. A decision tree fails visibly (“I didn’t understand”). An ungrounded LLM fails fluently, inventing a spec or a discount with perfect confidence. The old failure lost a conversation. The new one can lose a customer.
Why “which converts best” is the wrong question
Every comparison post, including one I used to have at this URL, eventually ranks the three tools against each other. The ranking is mostly noise, because the tools do not compete for the same conversations.
Complaints and judgment calls need a human, full stop. Software apologizing for a damaged parcel reads as an insult. Order status needs a database lookup, and dressing it in an LLM adds cost and failure surface to a solved problem. Discovery (“help me choose,” “will this fit my setup”) needs the loop, and the loop at 2am needs a machine.
The failure pattern across both waves of conversational commerce is category confusion: deploying one tool and grading it on all three conversation types. A scripted bot graded on discovery fails. A human team graded on 24/7 coverage fails. An LLM graded on complaint handling fails, and should never have been given the complaint.
A better model: a trust budget
The model I use instead of feature comparisons is a budget.
Every store has a small budget of customer trust in its conversational surface, and the budget is spent by failures, not by absences. A store with no chat bubble loses nothing from its budget. A store whose bubble fails a real question pays twice: the lost sale and a lowered expectation for the next visit.
This explains the low adoption better than any technology argument. Shoppers ignore chat bubbles because chat bubbles have been, on average, a bad bet for a decade. It also dictates the deployment order: the assistant earns budget on questions it can answer flawlessly from grounded data, and hands off everything else with context attached. Handing off is what protects the budget, not a weakness of the system.
The tools, layered that way, stop competing. The assistant takes every conversation first. It sells to the exploring majority, resolves the routine questions instantly, and routes the complaints and the four-figure hesitations to humans who now spend their hours only on conversations that need judgment.
Practical implications
If you run a store, three things follow.
Read a week of your own chats and support emails before buying anything. The mix of order status, discovery, and complaints in your inbox is the only comparison table that matters.
Grade each tool only on its conversation type. And instrument the assistant’s failures specifically: the transcripts where it was wrong or invented something are your trust budget leaking, and they matter more than the average rating. This is measurable; I wrote about how we measure it.
Never redeploy a scripted bot for discovery because it demos well on the three questions in the demo script. That is the 2016 mistake with a fresh coat of paint.
The shift
Conversational commerce did not fail because customers dislike conversation. Customers conduct billions of shopping conversations a year; they just have them with humans, in stores, because that is where the conversations were honest.
The category failed because it kept shipping the surface of a conversation without the substance of one. Adoption is low because shoppers are, collectively, running an accurate evaluation. The way to change the number is not a better bubble. It is being worth talking to.