Many small businesses start with the technology: which chatbot, which language model. We prefer to start with the work itself. Which questions come in, through which channel, who handles them now and how much time does that take? Only once that is clear can you see where AI gives time back and where a member of staff is better. We have already written about the chatbot itself in when an AI chatbot really helps your business. This article is about the process around it.
Which questions are suitable
Not every customer question is the same. Roughly divide them into four types.
Information questions
Opening hours, parking, prices, terms, what a treatment includes. The answer is fixed somewhere and rarely changes. This is the easiest place for AI to start, provided the source is correct.
Status questions
“Is my car ready yet?”, “When is the engineer coming?”, “Is my room booked?” The answer is in your system. AI can only answer these well if it is allowed to see that data.
Actions
Moving an appointment, cancelling a class, changing a reservation. This is where AI really takes over work, but only with a connection to your diary or booking system and with clear rules, such as until when cancelling is free of charge.
Sensitive questions
Complaints, refunds, damage, an angry guest. These need a person. AI can help by collecting the details in advance and summarising the conversation, but the decision stays with your team.
The knowledge source determines quality
An AI assistant is only as good as the information it draws on. Keep approved answers in one place, not scattered across old emails, an outdated PDF and your colleague’s head.
- One owner: someone in your team is responsible for the content and updates it when something changes.
- Short, factual answers: write as you would say it at the front desk, including the exceptions.
- Date of change: so you can see which information has not been reviewed for a long time.
- What AI must not say: record which topics always go to a person.
Which channel for what
Customers choose how they get in touch. A hair salon gets many WhatsApp messages, an installer is mostly phoned, a hotel receives emails about arrival times. Look at where your questions come in now and start there.
- Website: suited to information questions and bookings, at the moment the visitor is already on your site.
- WhatsApp: easy and personal, good for appointments and short status questions.
- Phone: most small businesses miss calls while they are working. An AI phone assistant answers, handles simple questions and logs the rest as a callback request.
- Email: AI can suggest a reply that a member of staff checks and sends.
The most important thing is that all channels draw on the same source. Otherwise a customer gets a different answer on the phone than on WhatsApp.
Handover to a person
The handover decides whether customers trust AI. A good handover works like this:
- Recognise: the assistant notices that it is unsure, that the topic is sensitive or that the customer wants a person.
- Be honest: it tells the customer what happens next and when they will hear back.
- Pass it on: your colleague receives the whole conversation and the collected details, so the customer does not have to repeat anything.
- Follow up: the question sits as a task with someone, not in a shared inbox where nobody picks it up.
You can read more about where to put the person in the process in human oversight in AI workflows.
Measuring whether it works
Agree in advance what you will measure and review it every week for the first few weeks:
- Handled independently: what share of questions was resolved without a member of staff?
- Handed over: how often and why? A recurring reason points to a gap in the knowledge source.
- Mistakes: read back a sample of conversations and note incorrect answers.
- Your team’s time: are there fewer calls and emails about the same topics?
When it does not work
AI in customer service is a poorer fit if few questions come in, if almost every question is bespoke or if your information is not yet written down anywhere. In that case a clear frequently asked questions page or a good form often delivers more. Start there and look again later.
Frequently asked questions
Which customer questions can AI handle best?
Recurring information and status questions whose answers are in your systems. Leave complaints and decisions about money with a member of staff.
How do I stop AI from giving a wrong answer?
Let it draw only on approved information, record which topics go to a person and regularly read back conversations.
Should I start with all channels at once?
No. Start with the channel where most questions come in now and expand once that runs well.
Can I process customer conversations with AI under the GDPR?
Yes, provided you have a lawful basis, sign a data processing agreement with your supplier, inform customers and do not keep conversations longer than necessary.