You keep the human touch by letting AI do the preparation while a person makes the decisions and sends the reply. Use it to draft, summarise and tidy. Keep people in charge of anything involving money, a complaint, a customer's legal rights or somebody who is upset. And if you later add a chatbot, give it a narrow job and an obvious way out to a human.
Start with AI on your side of the conversation
The safest first step is one the customer never sees. A person still reads every message and presses send. The assistant only helps you get there faster:
- Draft a reply. Paste in the customer's question and the relevant part of your own policy, and ask for a reply in your tone, under a set length, ending with the next step.
- Summarise a long thread. Ask for what the customer wants, what has been promised so far, and what is still open.
- Cool a reply down. If you wrote something while irritated, ask for a calmer version that keeps the facts.
Before you paste, take out the customer's name, address, email, phone number and order number where you can. The National Cyber Security Centre (NCSC) points out that what you type into a public AI chatbot is visible to its provider and stored, and advises against including sensitive information. The guide to using AI tools safely with customer data goes through this properly.
Check every reply about refunds and rights
This is the part most likely to hurt you. An AI assistant does not know your terms or UK consumer law unless you give it the wording, and the NCSC warns these tools can present incorrect statements as facts. A confident, friendly, wrong answer about a refund is still wrong, and it went out in your name.
GOV.UK's guidance for businesses is plain on the basics:
- You must offer a full refund if an item is faulty, not as described or does not do what it is supposed to.
- Customers who buy online, by mail or by phone generally have 14 days from receiving the item to tell you they want to cancel, and a further 14 days to send it back. You must refund within 14 days of getting it back. Some items, such as personalised or perishable goods, are treated differently.
- It is illegal to mislead customers about their rights or to try to take those rights away.
So a draft that says "we do not offer refunds on sale items", when the item was faulty, is not just poor service. Read the full page on accepting returns and giving refunds, check the current rules, and make it a habit that a person compares any reply about returns, cancellations or faults with your written policy before it is sent. This is general information, not legal advice.
Build a bank of approved answers
Good AI help depends on good source material. An hour spent here pays back every week.
- Go through the last month of emails and messages and list the ten questions you are asked most.
- Write one clear answer to each, or draft it with the assistant and correct it. Include the real details: delivery times, how to return something, who to contact.
- Have the person who knows your terms best check each answer against them.
- Save them where the whole team can find them, with the date they were last checked.
These become your saved replies, the basis of a help page on your website, and the only material a chatbot should be allowed to answer from. If drafting replies is new to you, the guide on writing prompts that get useful answers will help you brief the assistant well.
If you add a chatbot, give it a narrow job
A website chat assistant can answer routine questions when you are closed. It can also give a wrong answer to a hundred people before you notice. If you try one, set it up carefully:
- Say it is automated. Tell people at the start that they are chatting with an automated assistant. There is more on this in the guide on telling customers when you use AI.
- Limit what it answers. Give it your approved answers and set it to hand over when a question falls outside them, instead of guessing.
- Do not let it decide. Refunds, compensation, exceptions and complaints go to a person. The chatbot can collect the details and say when someone will reply.
- Make a human easy to reach. Show an email address, phone number or "talk to a person" option at all times, not only after three failed answers.
- Expect people to try to trick it. The NCSC notes that these tools can be manipulated by cleverly worded input into saying things they should not. Do not connect a chatbot to anything that lets it issue refunds or change orders by itself.
- Ask the supplier about data. Find out where conversations are stored, who can read them and how long they are kept, and mention the chat tool in your privacy notice.
- Read the transcripts. Weekly at first. You will find wrong answers, gaps in your help pages and questions you never thought to cover.
Keep people for the moments that matter
Some conversations decide how a customer talks about you afterwards. Handle these personally, from start to finish:
- A complaint, or a second message about the same problem.
- A faulty or unsafe product.
- A late order for something that mattered, such as a gift or an event.
- A bereavement, illness or money trouble mentioned in passing.
- Anyone who seems confused, distressed or vulnerable.
You can still ask the assistant for a first draft in private. But read the customer's message yourself, use their name, refer to what actually happened, and say what you will do and by when. That is the human touch, and no tool supplies it for you.
Measure solved problems, not speed
Fast replies are easy to count and easy to be proud of. What you want is problems that stay solved. Once a month, look at three things:
- Repeat contacts. How many customers had to write again about the same issue?
- Handovers. If you run a chatbot, how often did it pass someone to a person, and how long did they then wait?
- Corrections. How many AI drafts needed a factual fix before sending? If the number is rising, your approved answers need updating.
Now and then, ask a customer directly whether their problem was sorted. Their answer tells you more than any response-time figure.




Comments