There is a large and recent cohort of ecommerce and customer-experience leaders in market for an AI chatbot. Read their interviews expecting first-time adoption and the picture does not fit: most of them already run one. It came bundled with a helpdesk they bought years ago, or it is a live-chat vendor from before large language models existed. They are in market because that bot no longer clears the bar their customers hold it to — which makes this a replacement market wearing the clothes of an emerging one, and changes what the buying question actually is.
Do you need an AI chatbot for ecommerce?
Most ecommerce teams do not need to buy a separate AI chatbot, because they already have support automation bundled into an existing helpdesk or contact-centre platform. The real buying decision is whether that tool is good enough — or whether the business needs a separate AI shopping assistant.
Those are the two jobs the phrase covers: deflecting support tickets and helping someone choose a product. They share a name and almost nothing else. Treating them as one category obscures the differences that matter in evaluation — owner, budget, vendor set and measure of success.
What are the two types of AI chatbots for ecommerce?
The clearest way to see the split is to lay the two out as if they were what they really are: two different products that happen to share a name.
| Job one Support chatbot | Job two AI shopping assistant | |
|---|---|---|
| Job | Deflect support tickets | Help someone buy |
| Stage | Post-purchase — order status, returns, sizing complaints. | Pre-purchase — which of these is right for me, will it work with what I own, what size. |
| Owner | Customer experience or support. | Ecommerce, site search or merchandising. |
| Usually bought as | Bundled. The AI features inside the helpdesk or contact-centre platform already in place. | Standalone. A separate purchase from a young vendor, usually on a trial. |
| Measured on | Deflection and resolution — with the more mature buyers deliberately shifting toward sentiment. | Conversion. Rarely measured rigorously; A/B tests are the exception, not the norm. |
| Requires | Documented procedures and decision trees the AI can follow. | Product content good enough to answer a real question. |
| If you get it wrong | A confidently wrong answer to a customer who already has a problem. | A widget nobody uses, or one that answers questions your product page should have. |
The support job is the better-understood of the two, and it has its own hard-won lessons about what these tools do and do not deliver — the recurring one being that the deflection pitch runs ahead of the reality and has to be validated on your own ticket history before you price a plan around it. That evidence lives on our CS-platform lessons page, alongside why brands switch helpdesks in the first place. The rest of this page is about the half those pages do not cover.
What AI chatbots are ecommerce buyers deploying?
The two jobs produce two completely different procurement pictures, and the contrast is sharper than anything either one shows on its own.
On the support side, most buyers in these interviews use the AI capabilities bundled into their existing helpdesk or contact-centre platform rather than buying a separate chatbot. The large majority already run one of those platforms — Gorgias, Zendesk, Gladly, Kustomer, Freshdesk, Intercom — whose vendor now ships AI features in the product. The chatbot is a line in a renewal conversation, not a category with its own RFP.
On the shopping side, the vendor set has not settled. Buyers in these interviews name Rep AI, Tolstoy, Inbenta, Cavalry, Session AI and Crescendo among others, and the striking thing is how little repetition there is: no vendor recurs often enough across these interviews to look like a default choice. For a buyer, that scattering is the most useful fact about the category. It means these interviews do not reveal a consensus vendor for buyers to inherit, so the shortlist has to be built from first principles.
It also explains the shape of the deployments. Where buyers are moving on shopping assistants, they are running short trials on a limited product range rather than committing — one specialty retailer has a trial queued rather than a rollout, another brand is in conversations with two vendors without having reached an RFP. That trial-first behaviour is consistent with a market that has no default vendor.
Why are ecommerce buyers replacing chatbots instead of adopting them?
This is the finding that reframes the category. The in-market demand is not mostly first-time adoption; it is buyers whose existing bot has fallen behind a standard that moved. The same mechanism shows up in different brands with different vendors.
The bundled assistant is a decision tree with an AI label
A luxury apparel brand has run its helpdesk's virtual assistant for years and describes it plainly as more of a basic chatbot than an AI agent — adequate for routing, short of the personalization the brand needs. The tool works. The bar moved.
The pre-LLM chat vendor cannot reach the new models
A beauty brand is offboarding its conversational vendor because it does not integrate effectively with the language models now available, and is separately discontinuing another chatbot tool over dissatisfaction. Neither is a story about a bad product; both are about a product built before the current generation.
The tool that worked was discontinued
An apparel brand ran generative question-and-answer on its product pages for a year and A/B tested it to a slight conversion lift — a genuine, measured result. It is re-shopping anyway, because the vendor is going away. In a category this young, vendor durability is part of the product.
Read together, these say something specific about timing. The buyers in market here are rarely choosing between having a chatbot and not having one. It is choosing between living with a bot built to an older standard and paying the switching cost to reach the current one — and switching costs on the support side are substantial, because the assistant is entangled with the helpdesk, the macros and the team that lives in it all day.
What has to be true before an AI chatbot works?
An ecommerce AI chatbot needs an accurate knowledge source before it can work reliably. Support bots need documented procedures, policies and decision trees; shopping assistants need complete, accurate product content.
The prerequisite buyers name is not technical, in other words. It is that the business has written down what it knows.
A customer-experience director running one of the more mature deployments in this corpus puts it directly: the challenge is ensuring the AI has the proper context and workflows, and without comprehensive procedures and decision trees the AI struggles to resolve queries. That is a striking thing to hear from a team that is otherwise happy with its platform. The constraint is not the model. It is whether the operational knowledge exists in a form the model can follow.
The shopping-assistant equivalent is product content. A pre-purchase assistant can only answer what the catalogue, the specifications and the existing questions already contain; where that content is thin the assistant either declines to answer or invents. This is why the teams furthest along on shopping assistants tend to be the ones who already did product-data work for other reasons — the same groundwork that answer engine optimization depends on.
The practical test is unglamorous: before evaluating vendors, look at whether a competent new hire could answer your twenty most common customer questions using only what is currently written down. If they could not, you have a knowledge problem to solve before you have a chatbot problem.
How do buyers measure an ecommerce AI chatbot?
Support chatbots are typically measured on deflection and resolution; AI shopping assistants are intended to be measured on conversion, but few buyers in these interviews describe rigorous A/B testing.
Less rigorously than they intend to, then, and differently for each job.
Support deployments are measured on deflection and resolution rates, and the more mature buyers describe deliberately moving away from that. One team frames its goal as driving sentiment rather than resolution rates — an acknowledgement that a deflected ticket and a satisfied customer are not the same thing, and that optimizing hard for the first can quietly damage the second.
Shopping assistants are measured on conversion, and this is where the evidence is thinnest. Only a small number of buyers in these interviews describe having actually A/B tested one, and the single clearly reported result was a slight lift — real, positive, and not enough on its own to support a broader performance claim. That is not evidence the tools do not work. It is evidence that few buyers in these interviews have tested the category rigorously enough to support a broader performance claim — which is a reason to structure your own deployment as a test with a holdout rather than a rollout.
Want this read against your own stack?
Get my read →Should you buy a standalone AI shopping assistant?
Buy it as a test, not as infrastructure — which is what the buyers moving on it are doing.
The case is strongest when four conditions hold: your products genuinely need explanation, your product content is complete enough to answer the questions, you can measure conversion against a holdout, and you can tolerate the vendor disappearing or changing direction. Each of them is something these interviews point to. A pre-sale question has to be a real barrier rather than a rare one. Thin product content produces an assistant that either declines to answer or invents. Measuring against last month rather than a holdout will not tell you whether the tool did anything. And in a market with no default and young companies, a vendor going away is a live possibility rather than a tail risk — one brand here has already had it happen.
Where those conditions do not hold, fix the prerequisite before buying the tool: write down what the business knows, and fix the product content. That work makes the assistant viable later, improves the product pages now, and is the same groundwork the AI-discovery surfaces need anyway.
One boundary is worth drawing while you are here, because these three ideas are easy to conflate. An AI shopping assistant is distinct from agentic commerce: the assistant helps the customer choose on your property, while agentic commerce lets an AI agent act in the transaction on the customer's behalf, usually somewhere you do not own. The assistant is a tool you buy. Agentic commerce is a shift you prepare for.
The framing that survives all of this is simple. The question is almost never whether to have an AI chatbot — it is which job you are buying for, and whether you have given it anything true to say.
Common questions
Do you need an AI chatbot for ecommerce?
Most ecommerce teams asking this question already have one, so the practical decision is rarely whether to adopt. It is whether the bot already bundled into your helpdesk clears the bar your customers now expect, and which of two different jobs you actually need done: deflecting support tickets, or helping someone choose a product before they buy. Those are separate purchases with separate owners, budgets and success measures, and treating them as one category obscures the differences that matter in evaluation.
What are the two types of AI chatbots for ecommerce?
Support deflection and shopping assistance. Deflection resolves post-purchase questions — order status, returns, sizing complaints — is owned by customer experience, and almost always arrives bundled in the helpdesk or contact-centre platform a brand already runs. Shopping assistance answers pre-purchase questions on product pages to move someone toward a decision, is owned by ecommerce or site search, and is usually a standalone purchase from a much younger vendor. They look like the same product in a demo and behave nothing alike in production.
What is the difference between an AI chatbot and an AI shopping assistant?
An AI support chatbot primarily handles post-purchase customer-service questions such as order status and returns, while an AI shopping assistant handles pre-purchase questions intended to help a customer choose a product. Support chatbots are usually bundled into helpdesk platforms and measured on deflection; shopping assistants are more often standalone ecommerce tools measured on conversion. Both are distinct from agentic commerce, where an AI agent acts in the transaction on the customer's behalf.
What AI chatbots are ecommerce buyers actually deploying?
On the support side, most buyers in these interviews use what their existing platform ships — the AI features inside Gorgias, Zendesk, Gladly, Kustomer, Freshdesk or Intercom rather than a separate purchase. On the shopping side the vendor set is young and highly scattered, with buyers in these interviews naming Rep AI, Tolstoy, Inbenta, Cavalry, Session AI and Crescendo among others. No vendor recurs often enough across these interviews to look like a default, which is itself the most useful thing to know about the category.
Why are ecommerce buyers replacing chatbots instead of adopting them?
Because the bot they have predates the standard their customers now hold it to. Buyers describe bundled virtual assistants as decision-tree tools with an AI label, and standalone chat vendors from the pre-LLM era as unable to integrate with current language models. One luxury brand describes its helpdesk's assistant as a basic chatbot rather than an AI agent; another brand is offboarding a chat vendor for exactly that reason. Much of the in-market activity in these interviews is replacement demand rather than first-time adoption.
What has to be true before an AI chatbot works?
It needs something accurate to read. On the support side that means documented procedures and decision trees — one customer-experience director is explicit that without comprehensive SOPs the AI struggles to resolve queries at all. On the shopping side it means product content good enough to answer a real question, which is why the teams furthest along on shopping assistants are usually the ones who already invested in product data. A chatbot does not create knowledge the business has not written down.
How do buyers measure an ecommerce AI chatbot?
Differently for each job, and less rigorously than they intend to. Support deployments are measured on deflection and resolution, though the more mature buyers describe deliberately shifting toward customer sentiment because a deflected ticket is not the same as a satisfied customer. Shopping assistants are measured on conversion, and this is where evidence is thinnest: only a small number of buyers in these interviews describe having actually A/B tested one, and the result reported was a slight lift rather than a step change.
Should you buy a standalone AI shopping assistant?
Buy it as a test, not as infrastructure. The buyers moving on shopping assistants are running short trials on a limited product range rather than committing, which fits a category this young: the vendor set has not settled, one brand in these interviews has already had its generative question-and-answer tool discontinued underneath it, and measured results are scarce. The case is strongest where products genuinely need explaining before purchase and where the product content is already good enough to answer the questions.
Get this research made for your stack
Whether you need a chatbot at all — and which of the two jobs you are actually buying for — depends on your ticket mix, your product complexity, and what your helpdesk already ships. Do a 15-minute interview and get the version of this that applies to you: what peers at your size deployed, what they replaced, and what they measured.
Get my personalized read — 15-min interviewNo password · your interview is anonymized before it ever informs a page like this one.