AI customer service works best when it answers common questions from your own information and hands the conversation to a person as soon as it is unsure. Used that way it takes repetitive enquiries off your team, including after hours. Used loosely, it gives confident wrong answers to your customers.
This guide covers the jobs it suits, the ones it does not, how to set it up so it stays accurate, and the checks a person should keep doing.
What AI customer service can handle
The jobs that suit AI have an answer you could write down and a low cost if the answer is occasionally imperfect. Typical examples:
- Opening hours, location, parking and contact details.
- Service area and availability. “Do you cover Penrith?” “Can you come Saturday?”
- Booking and rescheduling, when it connects to your real calendar.
- Order or job status, when it can look the record up in your system.
- Policy questions. Returns, deposits, cancellations, warranty terms as you have written them.
- Collecting details for a person to follow up: name, number, what the problem is, a photo.
Think of a new receptionist on day one with a good folder of answers. They can handle a lot, and they know when to say “let me get someone to call you back”. That is the standard to aim for.
What it should not decide on its own
Some conversations need a person’s judgement or authority, however good the tool is:
- Complaints, and customers who are upset or distressed.
- Refunds, discounts, price changes and exceptions to policy.
- Anything about safety, health or legal rights.
- Quotes that depend on a site visit or a detail the tool cannot see.
- Any question where the right answer is “we would need to check”.
A good setup recognises these and hands over, with the conversation so far attached, so the customer does not repeat themselves.
Examples from ordinary businesses
An electrician. Most enquiries ask whether the business services a suburb, whether it does a certain job, and when someone can come. An AI customer service assistant answers from the service list and area, then offers real appointment times from the calendar. A request for an emergency callout or a job the assistant has not been given information about goes straight to the owner’s phone.
A dental or physio practice. Patients ask about fees, parking, referrals and rebooking. The assistant can answer from the practice’s own information and book into free slots. A question about symptoms, medication or treatment advice is not for the tool. It should say a clinician will call and pass it on.
A childcare centre. Families ask about waitlists, tours and fees. The assistant can take their details and offer tour times. Anything about a particular child’s care goes to the director.
How to set up AI customer service so it stays accurate
The quality of the answers depends far more on your material than on the AI. Plenty of poor results come from a tool that was never given your prices, or was given them once and never updated.
- Write the answers down first. Take your 30 most common questions and write the correct answer to each in plain words. If your own team gives different answers, settle that first.
- Give the tool only what you want it to say. Your approved information, not your whole website and every old document.
- Set the handover rules. List what must go to a person, who receives it, and how fast. Include a rule for what to do if nobody is available.
- Connect it to the real systems. A tool that books into a calendar nobody else uses will double-book you. It should write to your CRM so every conversation is on record.
- Name an owner. One person is responsible for updating the information when prices, hours or policies change.
This is the same approach we describe for the website chat, and for voice in our guide on the AI receptionist.
Testing it before customers see it
Do not trust the demo. Use your own questions.
- Ask the 30 most common questions your team answers. Mark each right, partly right or wrong.
- Ask ten awkward ones: a complaint, a question outside your services, a rude message, a vague one, a question in broken English.
- Ask for something it should refuse, such as a discount or a price it does not have.
- Try to get it to invent something. See what it does when you ask about a service you do not offer.
Count the wrong answers and, separately, the times it should have handed over and did not. The second number matters more. A tool that says “I’m not sure, let me get someone” is working. One that guesses is not.
What a person must keep checking
Once it is live, review a sample of real conversations every week at the start, and monthly after that. Look for:
- Answers that were wrong or out of date.
- Questions it could not answer, which tell you what to add to its information.
- Handovers that were slow or lost.
- Any customer who asked for a person and did not get one.
Treat these reviews as part of running the service, in the same way you would review how a new staff member handled their first month. Our guide on reducing AI mistakes explains why wrong answers happen and how to cut them.
Be clear with customers, and protect their information
Tell customers when they are talking to an automated assistant, and make it easy to reach a person. Some places have rules about disclosing automated systems, and rules about calls and messages differ by country, so check the regulator’s guidance where you operate. Customers who message you on WhatsApp are a special case, covered in our guide to a WhatsApp AI chatbot for business.
Customers also share personal information in these conversations. Check where it is stored, who can see it, and whether the provider uses it for anything beyond running your service. In Australia, the Office of the Australian Information Commissioner publishes guidance on privacy and commercial AI products at oaic.gov.au, and the Privacy Act does not apply to every small business, so check whether it applies to yours.
How to tell whether it is paying for itself
Measure the same few things before and after. As an example, suppose your office answers 40 routine questions a week at about three minutes each, and misses a handful of after-hours enquiries that never call back. Those numbers are made up, so use your own. The questions to ask are:
- How many hours a week did routine questions take before, and how many now, including review time?
- How many enquiries arrived outside hours, and how many of those became bookings?
- How many conversations were handed to a person, and were they the right ones?
- Have complaints about being ignored or misinformed gone up or down?
If hours fall and bookings hold or rise, it is earning its keep. If staff are spending more time fixing the assistant’s mistakes than answering the questions, stop and fix the information it works from before expanding it. Our guide on measuring the return on AI goes through this step by step, and the same approach applies in local services businesses, where many enquiries are repeats.
When to build it, and when a basic tool is enough
If you get a few enquiries a week, a simple auto-reply, a well written FAQ page and a person who answers within the day may be all you need. An AI tool is worth considering when the same questions arrive repeatedly, when enquiries come outside hours, or when slow replies are costing you work.
When it is worth building, the tool should answer from your information, book into the real calendar, write to your CRM and hand over to a person when it should. CentriWeb builds that kind of AI customer service and looks after it afterwards. We say so when an off-the-shelf tool does the job, and for many small teams it does.