- AI replies work best on repeat questions with clear answers: hours, prices, availability, directions and how to book.
- Build an answer library from the questions customers actually ask, written in your own words.
- Set clear rules for handing over to a person, and make sure the customer knows when that happens.
- Check replies regularly, especially in the first few weeks, and update the library whenever something changes.
- Treat customer messages as personal information and know where they are stored and who can see them.
The same customer questions, every day
Picture the lunch rush at a busy café. There's a queue at the counter, the coffee machine is going flat out, and your phone buzzes again. "Are you open Saturday?" Then an email: "Do you have a table for six tonight?"
Look back at last week's messages and you'll probably see the same pattern. How much is a haircut? Where do I park? Do you deliver to Hamilton? Most messages are some version of the same few questions.
Each answer only takes a minute. The trouble is they arrive all day, by phone, email, website forms, social media messages and Google reviews, usually while you're serving someone else. Some sit unanswered until evening, and by then the customer may have booked somewhere else.
That's the work we mean by customer replies: opening hours, pricing, availability and directions, answered quickly and in your customer's language.
Worth knowing. For cafés and takeaways, Google review replies and customer questions are one of the three tasks we usually take off first, with a typical estimate of around one and a half hours a week. For other businesses the figure varies a lot, so count your own messages before deciding.
What AI replies can and can't do
AI is good at reading a message written in everyday language, working out what's being asked and writing a clear answer from information you've given it. It isn't a replacement for knowing your business.
| Usually a good fit | Better left with a person |
|---|---|
| Opening hours, holiday hours and location | Complaints and unhappy customers |
| Standard prices and what's included | Custom quotes and discounts |
| Checking availability and offering a booking link | Refunds, disputes and anything involving money owed |
| Parking, access, what to bring | Health, legal or financial advice |
| Order status, where your system can look it up | Anything sensitive or personal |
Here's a handy rule of thumb. If a new staff member could answer it correctly from a printed sheet on their first day, AI can probably answer it too. If it needs experience, judgement or authority, it should go to a person.
A typical question and reply might look like this:
hi are you open this saturday and is there parking nearby?
Kia ora, yes, we're open Saturday from 8am to 2pm. There's free parking behind the building, with the entrance off the side street. Anything else I can help with? – Kōwhai Café
Every detail in that reply, the hours, the parking, the sign-off, comes from information the café has written down. Which brings us to the most important part.
Build an answer library from your real questions
AI replies are only as good as what you give them to work from. Rather than letting the AI make things up from general knowledge, you give it an answer library: a short, plain document with your real questions and the answers you want given.
Building one is simpler than it sounds:
- Collect the questions. Scroll back through a month of emails, messages and reviews, and ask whoever answers the phone. Write down every question that comes up more than once.
- Group them. Hours, prices, bookings, location, products or services, policies. Many small businesses find their list fits into a handful of groups.
- Write the answers in your voice. Short, friendly and exact. Include the details customers need, such as the booking link or the parking entrance.
- Mark the no-go topics. List what the AI must never answer, and what to do instead.
- Give it one owner. One person updates the library when prices, hours or staff change.
Write down the answers you give out loud. Many businesses have never written their answers down. The exercise is useful on its own, because it often shows that different staff give different answers to the same question.
Where AI customer replies happen
Customers get in touch in different places. We set up AI replies on your website and by email, and draft replies to Google reviews for you to approve. Each one has its own quirks.
Website chat
A chat window on your website can answer questions while someone's still deciding whether to call. It works best with a clear way to book or leave contact details. Our guide to websites that turn visitors into calls covers the rest of the page.
For email, many businesses prefer the AI to draft replies that a person checks and sends, at least at first. It saves the typing while keeping a human eye on every message. Common questions can move to automatic replies once you trust them.
Google reviews
On Google, the conversation now happens mainly in reviews on your Google Business Profile, because Google closed its Business Profile chat feature in 2024. Review replies are public and stay visible for a long time, so they suit an AI draft that you approve rather than a reply that posts itself.
Social media messages
Messages sent through social media apps aren't covered by the replies we set up, so those conversations stay with you. Each platform has its own rules for business messaging. If customers often ask quick questions there, pointing them to your website or email, where the answer library can help, is a simple way to take some of the load.
Replying in your customer's language
New Zealand businesses serve customers who speak many languages. AI can usually recognise the language a message is written in and reply in the same one, using your English answer library as the source.
So a customer who writes in Chinese, Hindi or Punjabi can get an answer in that language, without you keeping a separate library for each. If it helps, we can also talk the setup through with you and your team in English, Hindi or Punjabi.
Helpful, but handle with care. Machine translation is generally good for everyday questions and less reliable for precise wording. If replies include prices, terms, health instructions or anything a customer might rely on, ask a fluent speaker to check a sample before you switch it on, and keep a way for customers to reach a person.
Handing over to a person
The most important part of any AI reply setup is what happens when the AI should stop. A good handover means the customer is never stuck in a loop, and your team member picks up with the full picture.
I was charged twice for my order yesterday, can someone sort this out?
Sorry about that. I'll pass this to the team now and someone will get back to you today. Could you reply with the name the order was under?
The AI didn't try to explain the charge or promise a refund. It recognised a money question, said what happens next and collected the detail the team will need. Four rules make that work:
- Clear triggers. Hand over when the question isn't in the library, when the customer asks for a person, when the message mentions a complaint or refund, or when the AI isn't confident it understood.
- Tell the customer. "I'll pass this to the team and someone will get back to you today" is better than a vague answer.
- Alert the right person. The message goes to whoever handles that topic, by email, text or a shared inbox, with the conversation so far.
- Cover after-hours. If nobody's working, the reply should say when a person will be back.
Google review replies you approve
Replying to Google reviews shows future customers that you care, but it's easy to fall behind. AI can draft replies using the details in each review, so a thank you for a five-star review mentions what the customer actually liked rather than a generic line.
★★★★★ Best flat white in town and the staff remembered my name. The cheese scones are a must.
Thanks so much, Priya. We're glad the flat whites and the cheese scones hit the spot, and the team will love hearing they made you feel welcome. See you next time. – Kōwhai Café
That's a draft, not a posted reply. We recommend a person approves every review reply before it's published, and that negative reviews are always handled by the owner or manager. A calm, personal response to a complaint is worth far more than a fast one.
Never let AI admit fault or offer compensation. A review reply is public. Anything that sounds like an admission, a refund or a promise should come from a person who has looked into what happened.
Keeping AI replies accurate and private
AI can sound confident even when it's wrong. A few habits keep replies accurate:
- Read a sample of replies every day for the first few weeks, then weekly.
- Update the library the same day anything changes, such as public holiday hours or a price rise.
- Tell the AI to answer only from the library and to hand over rather than guess.
- Keep a log of conversations so you can see what was said if a customer queries it.
Customer messages often contain names, phone numbers and sometimes sensitive details, so they're personal information under the Privacy Act 2020. Know where messages are stored, who can see them, how long they're kept and whether the tool provider uses them to train its models.
When we set up replies as part of our AI automation service, the data we store is hosted in New Zealand or Australia and never used to train anyone's AI model. Only Ambs works on your project and sees your data, and the AI tools we use run on business plans that don't train on it.
It's still worth asking where each connected tool, including any AI service, processes messages. Our guide to the Privacy Act 2020 and automation lists what to check with any provider.
Not sure where to start? List your ten most common questions this week. That list is the beginning of your answer library, and it soon shows whether AI replies are likely to save you real time.
Try it with your own numbers
Work out your hours back
Pick one repetitive task. Move the sliders to match a normal week.
An estimate from your own numbers, not a promise. In a free session we work out a proper estimate together. Book a free session →
Common questions
Will customers know they are talking to AI?
They should. Being upfront builds trust, and many customers are comfortable with an automated answer to a simple question as long as it is accurate and a person is easy to reach. A short line such as "I'm the automated assistant, a team member can help with anything else" is usually enough.
Can AI reply to customers in other languages?
Yes. Modern AI can usually understand a message in one language and reply in the same language, drawing on an answer library written in English. For important wording such as prices, policies or health information, have a fluent speaker check the replies before you rely on them.
What happens when the AI does not know the answer?
It should say it will pass the question to the team, collect any details needed, and alert a person. It should never guess, especially about prices, availability or anything that could be taken as a promise.
Is it safe to put customer messages through AI?
It can be, if you choose tools carefully. Check where messages are stored, who can access them, how long they are kept and whether the provider uses them to train its models. Ambs Solutions hosts the data it stores for clients in New Zealand or Australia and never uses it to train anyone's AI model. Only Ambs works on client projects and sees client data, and the AI tools we use run on business plans that do not train on that data. AI services may process messages in other countries, so ask where processing happens as well as where data is stored.