Better customer service · Lesson 4 of 6

When to use AI in customer service, and when not to

AI can handle recurring questions well, but it can also be wrong and sound convincing while it is. In this lesson you decide what to leave to AI, what stays with people and how to set it up safely.

8 min

What you'll learn

  • Sort questions into suitable and unsuitable for AI
  • Make AI work only from information that's correct
  • Make sure customers can always reach a person
  • Test answers and keep checking them

Step by step

  1. 1

    Look at the questions you really get

    Read back through a few hundred conversations, or a sample if you have fewer, and sort them by topic. Questions that keep coming back and have a fixed answer you can look up are the best candidates: opening hours, delivery times, how something works. Questions where you have to weigh things up or make an exception usually aren't.

  2. 2

    Decide what AI never answers

    A wrong answer about opening hours is annoying, but a wrong answer about health, money or rights can really harm someone. A dental practice can let AI answer questions about parking, but pain after treatment goes straight to a person. Put complaints, disputed cancellations, vulnerable situations and personal details on this list too. Write the list before you start and add to it as things come up.

  3. 3

    Only give AI information that's correct

    An AI agent is only as good as its sources: an old price list leads to old prices. Gather your information in one place, remove outdated pages and name someone to keep it current. Also tell the AI what to do when it isn't sure, for example: “If the answer isn't in the information, say you don't know and forward to a team member.” This lowers the risk of mistakes but doesn't remove it.

  4. 4

    Make sure a person can always be reached

    A customer should never get stuck in a conversation with AI. Make sure “I want to talk to a person” always works, and be honest about what happens next, for example that outside opening hours it will be the next working day. In Bugalou you set this up in the AI Agent under Forward, with triggers and the option Auto-forward after unanswered messages (the AI Agent comes with Pro and Enterprise).

  5. 5

    Be open that it's AI

    Let customers know from the first message that they're talking to an AI assistant, for example: “Hi, I'm the digital assistant of [company]. I can answer questions about appointments and opening hours, and I'll pass you to a colleague when needed.” That's honest and helps people adjust their expectations. Also check which rules apply in your country and sector to AI and to handling customer data, and ask an adviser when in doubt.

  6. 6

    Test with your hardest questions

    Before launch, collect real questions, including vague, angry and wrong ones, and ask about things your AI shouldn't know or promise. Check whether the answer is correct, whether it makes things up and whether forwarding works. Repeat this after every big change. In Bugalou you can do this in your AI agent's Playground, before any customer sees it.

  7. 7

    Start small and keep reading along

    Turn on AI for one channel or a few topics first, and read along daily in the first weeks. After that you can move to a regular sample, but don't stop, because AI can keep making mistakes. Note that a conversation with no team member involved isn't necessarily a customer who was helped: they may have given up. So ask customers themselves how it went.

Common mistakes

Using AI to make people unreachable

If AI is mainly a wall between customer and company, the frustration still ends up with your team, only bigger. Make the way to a person easy to find.

Not checking what the AI claims

AI sounds confident even when it's wrong, and may promise something you don't offer. Read samples and treat mistakes as a signal to change your sources and instructions.

Starting with the hardest questions

Complaints and exceptions look tempting to hand off because they take so much time, but that's where a mistake costs the most. Start with simple, recurring questions.

Tips

  • Pass on the context at a handover, so the customer doesn't have to ask their question again.
  • Keep AI replies short. Long texts hide mistakes and get read less carefully.
  • Every question AI can't answer also points to something missing from your website or sources.

You're done when

  • You have a list of topics AI does and doesn't handle
  • The information AI uses is up to date and has an owner
  • A customer can always ask for a person and knows what happens then
  • You've tested with difficult questions and read along regularly

Frequently asked questions

Can AI replace customer service agents?

For many businesses that's not the right goal. AI can take over recurring questions, so people have more time for conversations that need understanding or a decision. A person has to stay reachable, also because AI can make mistakes.

How many questions can AI solve?

That depends on your type of questions and the quality of your information, so there's no honest percentage up front. Test on your own conversations, start small and check whether customers were really helped.

What if the AI says something wrong?

Count on it happening at some point. Put it right with the customer, change the source or instruction and test again. Agree in advance who picks this up, and let a person decide what you offer the customer.

Can I process customer data with AI?

That depends on the rules that apply to you, your supplier and what your customers have been told. Find out where the data is processed, share no more than necessary and ask a legal adviser when in doubt.