What an AI Chatbot for Ecommerce Should Actually Do (Beyond “Hi, how can I help?”)

Roughly seven out of ten online carts get abandoned. Baymard Institute puts the average at 70.19%, and the single biggest reason isn't price — it's unexpected costs at checkout, the shipping fee or tax that shows up on the last screen. Which means a huge chunk of your lost sales are really lost questions: "How much is shipping to Ireland?" "Will this arrive before Friday?" "Can I return it if it doesn't fit?" Nobody answered in time, so the tab closed.

That's the real job of an AI chatbot for ecommerce. Not to greet people. To answer the specific, boring, high-stakes questions that decide whether someone buys, and to take the repetitive post-purchase load off your inbox so your team isn't spending its day typing tracking numbers.

We build ChatterMate, an open-source AI support agent, so we spend a lot of time looking at what actually shows up in store support queues. The pattern is remarkably consistent, and most chatbots are aimed at the wrong part of it.

The three questions that eat an ecommerce support queue

Open almost any store's helpdesk and the tickets sort into three buckets.

The first, and by a wide margin the largest, is WISMO — "where is my order." Estimates vary by catalog and season, but order-status queries routinely make up somewhere between 20% and 40% of ecommerce support volume, climbing past 50% during peak periods like Black Friday and the December rush (Salesforce). These aren't hard questions. They're just relentless. A customer places an order, then checks on it four or five times before it lands. Multiply that by your order count and you've got a queue that never empties.

The second bucket is policy: shipping options and cutoffs, return windows, exchange rules, warranty terms, what "final sale" means on that discounted item. The answers already exist, sitting on your shipping page, your returns page, your FAQ. Customers just don't read them, or can't find them, so they ask a human instead.

The third is pre-sale: sizing, materials, compatibility, "does this work with the older model," "is this in stock in blue." This is the bucket that quietly costs you the most, because a pre-sale question that goes unanswered isn't a ticket — it's a sale that never happened.

A good chatbot has to handle all three differently. And here's where most of them fall over.

Why a generic chatbot makes ecommerce support worse, not better

Plenty of stores bolted a chatbot onto their site in the last two years and quietly regretted it. The reason is almost always the same: the bot made things up, or it dead-ended.

Making things up is the dangerous one. If a general-purpose AI chatbot guesses that your return window is 30 days when it's actually 14, you now have a customer waving a screenshot of your own bot promising something you don't offer. In ecommerce that's not a cute error. That's a chargeback, a bad review, and a support escalation that takes longer than if the bot had never existed. We wrote more about why this happens in our piece on what RAG is and why support bots need it. The short version is that a model with no grounding will always prefer a confident-sounding answer over "I don't know."

Dead-ending is the quieter failure. The bot can't help, doesn't say so clearly, loops the customer through the same three canned buttons, and never offers a way to reach a person. The customer rage-clicks, gives up, and emails you anyway, now annoyed. You've added a step, not removed one.

So the bar for an ecommerce bot isn't "can it chat." It's two things: does it only say true things about your store, and does it hand off gracefully when it's out of its depth.

Grounded answers: the part that actually matters

"Grounded" is the word we keep coming back to, and it's worth being concrete about what it means for a store.

A grounded chatbot doesn't answer from the general internet or from whatever the model absorbed in training. It answers from your content — your shipping policy, your returns page, your product descriptions, your help docs — and it can show where the answer came from. Ask it about international shipping and it pulls the real number off your real shipping page, with a link. Ask it something you've never documented and it says it doesn't know and offers a human, rather than inventing a plausible lie.

That single behavior changes the risk profile completely. You're no longer worried the bot will promise a free return on a final-sale item, because the bot can only repeat what your policy actually says. If your docs are good, your bot is good. If your docs have a gap, the bot surfaces that gap by handing off, which is useful feedback about what to write next. Your knowledge base becomes the product, and the bot is just the fastest way to read it.

This is also why the "train it on your docs" step isn't a chore to rush. The quality of an AI chatbot for ecommerce is capped by the quality of what you feed it. A store with a crisp, current returns page and clear shipping table will get a bot that sounds like it knows the business. A store with a two-year-old FAQ and three contradictory shipping promises will get a bot that sounds confused, because the business is confused, and the bot is just reading it back.

The honest math on deflection

Here's where I'll push back on the industry's own marketing, including some of ours.

Vendors love to quote deflection rates. And the numbers can be real. AI chatbots with solid retrieval commonly handle a meaningful share of routine questions before a human ever sees them, and Shopify's own research has pointed to AI handling a large chunk of inbound support. But "deflected" and "resolved" are not the same thing, and conflating them is how stores end up with a bot that looks great on a dashboard and terrible in real life.

Gartner found that only 14% of customer service issues are fully resolved in self-service, even though 73% of customers use self-service at some point in their journey (Gartner). Read those two numbers together. Almost everyone tries the self-serve route. Very few get all the way to done. The gap between "the customer engaged the bot" and "the customer's problem is actually solved" is where satisfaction lives or dies.

So the metric worth chasing isn't how many tickets you deflected. It's how many customers got a complete, correct answer and didn't need you afterward. WISMO and clear policy questions can hit that bar routinely. A damaged-item dispute or a custom-order change usually can't, and pretending otherwise just moves the frustration downstream. If you want to measure this properly rather than by vanity number, we broke down deflection rate versus real ROI separately.

The teams that win treat the bot as the front 60–70% of the conversation, not the whole thing.

Don't sleep on the pre-sale question

Most of the chatbot conversation is about cutting support cost. Fair. WISMO is expensive and dull. But the bigger prize for a lot of stores is on the way in.

Go back to that 70% cart abandonment number, and the fact that surprise costs are the top reason people bail. A shopper standing at your checkout with a sizing doubt or a "will it arrive in time" worry is a sale you can still save — if something answers in the ten seconds before they leave. That's not a support function. That's revenue.

A grounded chatbot that knows your shipping cutoffs, your size chart, and your stock language can settle those doubts on the spot: "Standard shipping to the UK is £3.95 and orders placed before 3pm today arrive Thursday." No waiting for an email reply that lands after they've bought from someone else. This is the same instinct behind why retail chatbots work for stores that live on quick answers — the buying decision happens in a narrow window, and whoever answers inside it wins.

Human handoff that doesn't feel like a wall

No bot should try to handle a "my order arrived smashed and I want a refund now" conversation to the bitter end. The move is a clean handoff — and the quality of that handoff is what separates a bot people tolerate from one they resent.

Done badly, handoff is a customer repeating everything they just typed to a human who has no context. Done well, the bot recognizes it's out of its depth, tells the customer plainly that it's bringing in a person, and passes the full conversation over so the agent picks up mid-thread. The customer feels escorted, not bounced. We went deep on the mechanics of handoff that doesn't frustrate customers, because getting this wrong undoes all the goodwill the bot earned answering the easy stuff.

The rule we use: the bot should be confident about facts from your docs and humble about everything else. Emotional situations, edge-case disputes, anything involving money moving in an unusual way. Hand it to a person early, with context attached.

Setting one up on your store without a six-month project

You don't need a platform migration to do this. The practical sequence looks like this.

Start by pointing the bot at your real content — shipping page, returns policy, FAQ, product docs. This is the step that determines everything downstream, so it's worth cleaning up contradictions before you feed them in rather than after. Next, decide the handoff rules: which topics go straight to a human, and what hours your team actually covers so the bot sets honest expectations. Then embed the widget on your store — most platforms take a snippet, and if you're on a custom React or Next.js front end, dropping in a chat widget is a small job. Finally, watch the transcripts for the first couple of weeks. Every question the bot punted on is a gap in your docs; fixing those is how the resolution rate climbs.

If data control matters to you, and for a store holding customer order data it should, an open-source, self-hostable option means the transcripts and customer questions stay on infrastructure you own, rather than sitting inside a vendor's black box. That's a large part of why we built ChatterMate the way we did, and why stores that care about privacy tend to end up on open-source support tools.

The short version

An AI chatbot for ecommerce earns its place when it does two unglamorous things well: it answers WISMO and policy questions correctly and instantly from your own store's content, and it hands the hard stuff to a human before the customer gets annoyed. Everything else — the greeting, the personality, the deflection dashboard — is secondary to those two. Get grounding and handoff right and the bot quietly removes your most repetitive work while catching sales you were losing at checkout. Get them wrong and you've just added a confident liar to your storefront.

Written by the ChatterMate team — we build an open-source, AI-first customer support agent that answers from your own docs, with citations.

If you want to try it on your store, ChatterMate is open source and free to start — your first 300 chats are free, and you can self-host it if you'd rather keep everything in-house. Point it at your shipping and returns pages, and see how many of those "where's my order" messages you never have to answer again.