Customer support automation: where a small shop starts

Start with the requests that come up all the time and cost little when answered wrongly. Leave the rest with a person, and move the line only when the bot has earned it.

A common first move is the opposite one: switch on a chat widget, let it answer anything, and learn which requests were a bad idea from the first angry message. The about page says what this site is for; the sorting below is the practical part.

How do you sort requests?

Take a week of messages from your inbox or chat and put each request into one of four boxes, using two questions. How often does this come up? And if the answer is wrong, what does it cost?

Here is the sorting on made-up examples. None of these are real customers.

• Frequent and cheap to get wrong: "Where is my order?", "What are your opening hours?", "Do you ship to my country?" A wrong answer costs a follow-up message.
• Frequent and expensive to get wrong: "Can I return this after the window?", "I was charged twice." The volume tempts you to automate, but a wrong promise costs money or trust.
• Rare and cheap: "Do you have a gift wrap option?" Fine for the bot, though not worth your first afternoon.
• Rare and expensive: "My parcel arrived broken and I want compensation." Give this to a person.

Start with the first box. It is also the one where the answer is a fact you already know, not a judgment call.

Why is the second box the risky one?

A language model writes confident sentences whether or not it has the rule behind them. Whatever your bot tells a customer about refunds, delivery times or plan changes, your shop has said. So the bot should answer only from rules written down in advance, and if they are not written yet, that comes before any automation.

What does "automate" mean for one request?

It is not one switch. Each request can sit at one of three levels:

1. The bot answers fully (order status when it can see live order data, opening hours).
2. The bot gathers details and passes the chat on (order number, photo of the damage, what the customer wants).
3. The bot does not touch it and says so.

Level two gets overlooked. A bot that collects the order number and the problem before a person joins removes the opening round of questions without risking a wrong promise.

Where should a human stay in charge?

Anything involving an exception, a dispute about who is at fault, an angry customer, or a request the written rules do not cover. If you catch yourself thinking "it depends on the customer", that request belongs to a person. The same goes for anything where the bot cannot see the facts: without access to live order data, it should not comment on where a parcel is.

A first week, step by step

1. Export a week of messages and sort them into the four boxes. A hand tally is enough.
2. Pick one frequent, cheap request and write its answer in full, with conditions. Automate only that one.
3. For every request you keep with people, decide what the bot says instead, for example "I'll pass this to a colleague; please have your order number ready."