What should you automate in a shop's customer service?
Do not guess from memory how many simple questions you get. Take the last 50 customer messages and for each one write down two things: where you got the answer and whether you had to decide anything. “How much is delivery to a parcel locker?” is settled by a price list. “Why hasn't my order arrived yet?” is also about a parcel, but here you need to check a specific order and sometimes the carrier too.
A bot suits cases where the answer sits in the shop's data and nobody has to rule on anything: delivery cost and method, return rules, dimensions or materials from the product page. A disputed complaint, a request for an exception to your terms or a price negotiation stays with you. Watch out for short questions. “Will this part fit my bike?” is only a few words, but without compatibility data in the catalog neither a bot nor you can answer it reliably.
To start, pick the one topic that comes back most often and prepare a proper source for it, for example a description of how you pack or a price list of delivery options. Decide who fixes that description when something changes. Then ask the bot the same question five ways, the way customers write it, typos included.
Test shopping advice on your own products
A different case is the customer who does not know what they want. They write “looking for a gift for someone who is just starting to run, up to PLN 200” and hope somebody will suggest something. A bot with access to your catalog should ask for the missing details and show a few proposals from your range as cards with a photo and a price, from which the customer can go to the basket.
The explanation tells you the most. If the bot says the shoes “will do well in a marathon” and the description says nothing of the kind, you have a mistake to fix, even if the product itself fits. Also ask for something you do not sell. A good bot will admit it has no such offer instead of pushing anything from the catalog.
How should a bot hand a case over to a human?
A customer with a case for a human must not get stuck in a loop of “I don't understand, try again.” Write down in advance the situations that should reach you, such as a disputed complaint or a request for terms outside your offer, and note who takes them and when the customer will get a reply. Leave a way out too. Someone who asks for a human should not have to convince the bot first.
In CartBot you take over the conversation in the panel and see its whole history, so the customer does not have to tell the story again. On handover the system tries to prepare a short summary of the case, though it does not always manage. When nobody is available, the bot asks for contact details.
A handed-over case waits in the panel until someone looks at it. A promise of “I'll connect you with an agent in a moment” made while you are standing at the label printer ends in silence on the other side, so set the message for the customer according to the hours when someone really answers. Contacts left in the evening need someone to read and close them in the morning.
Automation in a one-person shop
When you are alone, every break from packing costs you. Before you automate anything, record for a few days the time from opening an order to putting the finished parcel aside, including finding the goods on the shelf and correcting addresses. The time spent taping the box alone will look too good. Say you have 4 hours a day for shipping and one order takes 8 minutes on average. That is 30 orders, but only if nothing interrupts you for the whole 240 minutes. Leave a margin.
Automating customer service processes also covers what happens without any conversation, meaning order status emails and printing labels from a carrier integration. Place a test order and follow the parcel all the way. If the “your parcel has been shipped” email goes out right after the label is printed, while the box is still on your table, fix that condition before customers start asking where their shipment is.
A bot will take part of the questions off you, as long as it has the data. If it checks order status, it should do so only after getting the order number and the email address used at purchase. It cannot always give a delivery date, because the shop itself does not always know it. Once every few days read a sample of its conversations, because some of the mistakes nobody but you will catch.
Customer service on Black Friday and at the peak of the season
In 2026 Black Friday falls on 27 November and Cyber Monday on 30 November. Before the ads start, count the load in minutes, because the number of visits alone will not tell you how many people will need help. Say you normally answer 20 questions a day and at the peak 60 will arrive. At 5 minutes per case that is 300 minutes instead of 100. More than three extra hours, every day. Do the calculation on the cases from your last promotion, separately for pre-purchase questions and for cases about orders already placed, because if delayed parcels fill the queue, recommending products will change nothing.
Before the promotion, write down the answers to questions about its rules: until when the code works, whether it combines with a markdown, what happens after the basket changes, when a parcel ordered today will arrive. Put them in the bot's knowledge base and in a cheat sheet for whoever is on duty. For a discount on a selected category, add the exclusions so that nobody has to work out again why a code works on one jacket and not on another. Test a refusal too: put an excluded product in the basket and enter the code. A bare “invalid code” message leaves the customer with one more question.
A bot can suggest a configured discount code, but the code has to work in the shop itself. A well described promotion in the chat does not prove that the discount is applied in the basket, so go through every code all the way to checkout. When the campaign ends, check that the bot has stopped suggesting it.
At the peak, set an order of priorities. Payment failures and errors that touch many orders at once come first. Do not put off a complaint just because the basket was small. Leave the navigation redesign for after the campaign, but a wrong price or a payment that does not work needs a reaction on Friday evening too, so prepare in advance a way to switch off a faulty code and decide who has the say on stopping the promotion. If you change the conditions midway, update the bot's knowledge base as well.
For a backlog, set a threshold ahead of time, based on your daily capacity and the shipping time you promised. When there are more unshipped parcels than that, slow the ads and extend the shipping time on the site. The new deadline applies only to new orders, though. Write separately to customers who bought earlier and tell them about the delay. And if you cannot confirm that a parcel will arrive before Christmas, promise nothing. “We will check the shipment status with the carrier and get back to you tomorrow by noon” is enough, provided you keep to that deadline.
How to check whether automation pays off
Count the cases the bot resolved correctly without you and those that came back to your desk anyway. Read the conversations that broke off right after the bot's answer too, because no further message does not mean the customer got what they were looking for.
Count time first on an example, then on your own data. Say that in a week the bot correctly handled 30 questions, each of which used to take you 4 minutes. That is 2 hours. If reading the conversations and making fixes took you 45 minutes, you are left with 75 minutes saved.
With advice and promotions, sales come in as well. Once measurement is connected to the shop, CartBot shows the orders linked to conversations. Their value does not prove extra profit, because you do not know how many of those people would have bought without the chat. If you compare the Black Friday weekend with an ordinary one, write down the discounts and the ad traffic next to the revenue, because they change the result too.
After every bigger campaign, go through the conversations where a customer came back with the same problem. Log answer errors separately from handover problems, since they are fixed in different places. For each gap note one concrete fix and the person who will make it. That list is where you start preparing for the next campaign.