If you are building AI chat for a service business, the hardest part is not the model. It is the scope decision that comes before any code: which questions the chat should answer, which it should never touch, and what exactly happens when it hands over. Get those three decisions right, and a small chat can be genuinely useful. Get them wrong, and even a capable system will frustrate customers. This guide sets out a scope that works in practice.
Key takeaways
- Automate only stable answers that come from a source you control.
- Keep pricing, refunds, promises, complaints and anything personal with a human.
- Design the handover before building, and set a reply time the business can keep.
- Hand over the full conversation so the customer never repeats themselves.
- Test with real customer wording, and review failed chats every week.
Start with the questions customers actually ask
Before choosing a tool, list the questions customers already ask by email, phone and contact form. Sort them into three groups: stable answers, such as opening hours, delivery times and how to book; judgement answers, such as prices for unusual jobs, refunds and complaints; and personal answers, such as anything that needs account or order details. This list, not a feature list, decides what the chat should do. Automate only the first group at first. Keep the other two with a person until the business is confident the chat is working well.
Only automate answers that have a reliable source
An answer is only as good as its source. If opening hours, prices or service details change often, or live in several different documents, the chat will give wrong answers. Before automating anything, pick one source that is kept up to date, such as a single page or a short document, and let the chat use only that. If the answer is not in the source, the chat should say so and offer a human, not guess. A chat that invents a price or a policy is worse than no chat at all.
Design the handover before you build anything
Every chat needs a clear handover, and it should be designed first, not added later. Decide three things in advance: what the chat says when it hands over; how quickly a person will reply; and what the person receives. A message such as "I will pass this to our team, and they will reply within two working hours" is far better than silence. Set a reply time the business can actually keep, and say it plainly. When the chat hands over, it must pass the full conversation to the person, so the customer does not have to repeat everything. A handover that makes the customer start again is a failure, however good the chat itself was.
Keep people in the loop for judgement and emotion
Some conversations need a person, whatever the technology can do. Complaints, refunds, unusual requests, and anything where the customer is upset should go to a human quickly. The same applies to promises. A chat should never promise a delivery date, a price or a refund that only a person can approve. Write these rules down as a short list of topics the chat must hand over, and make sure everyone in the business knows them. Customers trust a chat that knows its limits more than one that pretends to know everything.
Test with real customer wording, then review every week
Before launch, test the chat with real customer questions, copied from actual emails and messages, including misspellings and short, unclear questions. If it handles those, it will handle most real chats. After launch, review the chats every week: which questions did it answer well, where did it fail, and where did customers ask for a human. Use those failures to update the source of answers and the handover rules. A small chat that improves every week will soon be more useful than a large one that nobody maintains.
FAQ
Should AI chat answer pricing questions?
Only for standard, published prices from a reliable source. Anything unusual, negotiable or personal should go to a human.
What should the chat do when it does not know?
Say so plainly, and hand over to a person with the full conversation attached. Never guess.
How fast should a human reply after handover?
Set a time the business can keep, such as two working hours, and state it in the handover message.
What should we review each week?
Failed answers, handover requests, and questions customers asked in their own words. Use them to update the answer source and the rules.
The goal is not to remove people from support. It is to let the chat handle the stable, repeated questions, so people have time for the conversations where judgement, care and a real decision matter. Build that boundary first, and the technology becomes simple.
Originally published on Pixel2Tech: https://pixel2tech.com/blog/ai-chat-support-automate-vs-keep-human
Top comments (0)