Ecommerce support has a shape that is easy to see in the queue and hard to see in a strategy deck. A large share of contacts ask for something your site already says somewhere: delivery windows, sizing, whether a product works with another one, how the return process starts. Behind them sit a smaller number of contacts where a customer is upset, something went wrong with a specific order, or the answer depends on judgment. The first group wears the team out. The second group is where customers decide whether to come back.
AI can take real weight off the first group. It can also make the second group worse if it is allowed to improvise. This guide is about drawing the line between the two for your own store, deciding what the AI may answer, and putting it in front of customers in stages.
Start from your contact reasons, not from the tool
Before you look at any product, pull a month of tickets and chats and tag each one with a reason. If your help desk already has tags, use them, but read a sample anyway; tags drift. You are looking for the top ten or fifteen reasons, and for each one you want to know three things.
- Is the answer written down? Not "does an agent know it," but "is there a paragraph a new hire could read and answer correctly?"
- Does the answer depend on the customer's own data? "How long does standard shipping take" does not. "Where is the order" does.
- What happens if the answer is wrong? A wrong sizing note costs a return. A wrong statement about a warranty may cost far more.
The third question matters most. It gives you a risk ranking that you can defend to your finance and legal colleagues, and it keeps the conversation about customers instead of about technology.
The first things to automate: stable, written, low-stakes
The best early candidates share a profile: the answer is the same for every customer, it already exists in writing, and a mistake is cheap and easy to correct.
- Policy questions. Delivery areas, typical delivery times by method, how the return window works, what the exchange process looks like, which payment methods the store accepts. These are the same question in a hundred phrasings.
- Product facts. Materials, dimensions, compatibility, care instructions, what is in the box. If your product pages hold this, the AI can answer from them. If your product pages do not, that is the first finding of your project.
- How-to and setup. Assembly, first use, connecting a product to an app, descaling a machine. Customers who have just received a parcel ask these in bursts.
- Store information. Opening hours for a pickup point, holiday closures, how to reach a person and when they are available.
Notice what is missing: anything the customer has to be identified for. You can get a long way on the list above without touching an order system at all, and that is a good place to start because there is nothing to go wrong with someone else's data.
What to keep with people, at least at first
Some contacts should reach a person quickly, and the AI's job is to recognize them and hand off with the details already collected.
- Anything emotional. A gift that arrived broken the day before a birthday, a repeated failed delivery, a customer who says "this is the third message about this." Tone matters more than speed here.
- Anything involving a customer's payment details or account security. The AI should not ask for them and should say where to go instead.
- Disputes and exceptions. Late claims, outside-the-window requests, goodwill decisions. Someone has to own the decision, and the customer should know who.
- Complaints about a person or a partner. A courier, a seller, one of your own agents.
- Vague or unusual product problems. If the customer is describing something your knowledge does not cover, a confident guess is the worst outcome.
Do not treat this list as permanent. It is where you start. As you read conversations and see which of these the AI handles well, you can move items across. Moving them the other way, after a customer has been burned, is much more expensive.
Fix the knowledge before you blame the AI
Teams are often surprised that the hard part is not the AI. It is the source material. An assistant that answers only from what you give it is exactly as good as three things: whether the facts exist, whether they agree with each other, and whether they are current.
Run a quick audit on your top ten reasons:
- Find every place the answer lives: the help center, the product page, the returns page, a macro, an internal doc, the checkout footer.
- Compare them. If the returns page says thirty days and a macro says twenty-eight, an AI will inherit that conflict, and so will your agents today.
- Name one owner per topic, and one place that wins when sources disagree.
- Write the missing answers. A good test is whether a colleague from another team could answer correctly using only the text.
A rule that helps in ecommerce specifically: when a price, a stock level or a promotion could differ between sources, the answer should not quote one at random. Decide in advance that the assistant says it cannot confirm and offers a person, rather than picking the source it likes.
Roll out in stages
Do not switch on every channel and every topic in the same week. A staged plan gives you something to measure and something to roll back.
- Test in private. Write twenty to thirty realistic questions for each topic you are automating, including the awkward versions: typos, two questions in one message, a customer who is already angry. Run them in a preview that sends nothing to customers and read every answer. The post on testing an AI support bot before launch has a fuller method.
- Go live on one topic and one channel. Website chat on shipping and returns policy is a common first slice because the answers are public and the cost of an error is small.
- Read every conversation for the first week. Not a sample. Every one. You are looking for wrong answers, but also for questions you did not predict, and for places where the handoff happened too late.
- Add the next topic only when the first is boring. Boring means you stop finding errors.
- Then add the next channel, and carry the same knowledge and rules across, so a customer on email hears what a customer on chat hears.
Design the handoff as carefully as the answer
Every ecommerce team knows the feeling of a customer who has to explain the same thing three times. When the AI hands a conversation to a person, the agent should receive the customer's question in their own words, the details already collected (order reference, what product, what they have tried), and the reason the AI stopped. The customer should be told what happens next and roughly when. Our post on chatbot to human handoff goes through what to include.
One detail to settle early: what the AI does outside staffed hours. Handing off to nobody is a failure. Taking a message that lands in the right queue, with the same summary an agent would have received, is a perfectly good outcome for a question that cannot be answered from the knowledge.
Measure what customers feel, not only what you saved
It is tempting to report deflection: the share of conversations the AI closed without a person. It is a useful number and an easy one to game. A better set for an ecommerce team:
- Repeat contact within a few days on the same topic. If customers come back, the answer did not work.
- Handoff quality. In a weekly sample, did the agent need to ask the customer anything the AI already had?
- Wrong or unsupported answers found in review, by topic. This is your list of knowledge to fix.
- Questions the AI could not answer. This list is free product research; it tells you what your site fails to explain.
- First response time for the contacts that do need a person. If the AI is working, this should improve because the queue is shorter, and you can check that it did.
Watch the peaks separately. Sale weeks and the days after a big dispatch behave differently from a normal Tuesday, and the AI's value is greatest when your team is least able to expand.
Where RyteDesk fits
RyteDesk is an AI front desk in early access, built for mid-market support teams. It answers on chat, email and phone from the knowledge and policies you give it, tells callers it is an AI, and hands off to your team or takes a message when it cannot help. You write the playbooks, test them in a preview that sends nothing to customers, and can read every conversation. If you want to try that approach on your own shipping and product questions, Join the waitlist; it is free to join.