Customer service automation: how to sort your enquiries before you connect an AI agent

Contents

1. Why you start from the messages, not the website
2. Step 1: collect your last 100 enquiries
3. Step 2: sort into three piles
4. Step 3: what you prepare for each pile
5. Step 4: what to connect first
6. Three mistakes that keep coming back
7. What to count after a month
8. How this looks at WhaleBiz
9. Frequently Asked Questions

Most businesses looking for customer service automation start in the wrong place: the FAQ page on their website. That page says what the owner thought, two years ago, customers would ask. The business WhatsApp says what they actually ask: "Is there parking?", "Can I move it to Thursday?", "Did you send me the invoice?", "I was charged twice". Build a customer service bot from that old page and you get a bot that beautifully answers questions nobody asks, and gets stuck on the first one somebody does ask. The exercise below takes one evening: take your last 100 enquiries, sort them into three piles, and only then decide what an AI agent closes alone, what it does in a system and what goes to a person.

Why you start from the messages, not the website

An FAQ page is written once and forgotten. The messages on WhatsApp, Instagram and the website chat are new every day, and they tell you three things the page does not know:

• The words customers use. A customer does not write "what is your cancellation policy", they write "I can't make it tomorrow, what now".
• What repeats most. A question that arrives twice a day is worth more than one that arrives once a month, even if the second looks more important.
• When they ask. If a large share of enquiries arrives in the evening and at weekends, that is where automation pays first.

One more note on words: what people search on Google as a "customer service bot" is today usually an AI agent. It does not lead the customer through a menu of buttons. It reads what the customer wrote, or hears what they said in a voice message, answers from the business's own information, and knows when to stop and hand over to a person. The differences between the approaches are covered in the guide to AI customer support.

Step 1: collect your last 100 enquiries

You need no system for this. A simple spreadsheet is enough, one row per enquiry, with five columns:

1. What the customer wrote, in their own words. Copy, do not summarise.
2. The channel: WhatsApp, Instagram, Facebook, website chat, phone.
3. The time and the day.
4. What was actually done: answered by message, booked an appointment, sent a document, passed it to someone.
5. Who handled it and how fast.

A few things worth knowing before you start. One enquiry can hold two requests. "I wanted to book for next week, and by the way I was charged twice last time" is two rows, because each half goes to a different pile. Voice messages and photos count like any enquiry: a customer who sends a photo of a broken product is asking a question, even without writing a word. And a small business that gets few messages can go a month back instead of counting 100 enquiries. The goal is to see what repeats, not to build statistics.

Step 2: sort into three piles

Every row in the sheet goes into one of three piles. The deciding question is not "how hard is this enquiry" but "what does it take to answer it".

Pile A: an answer from information

Here are the questions whose answer you already have, in someone's head or in a document: opening hours, address and parking, what to bring to the first treatment, how to prepare for a test, what the service includes, the business's price list, the cancellation policy, delivery times. The customer does not need anyone to do anything, they need to know something.

This is the easiest pile to hand to an AI agent, on one condition: the information is written down. An agent that answers from the business's documents answers only what is written in them. Whatever lives in the secretary's head has to move onto a page. How to do that without writing a book is explained in the guide to a knowledge base for an AI agent.

Pile B: an action in a system

Here the customer wants something to happen: book an appointment, move it, cancel it, get a copy of an invoice, get a payment link, find out where the order is. The answer is not in a document, it is in the calendar, in the clinic management software or in the shop's system.

An AI agent can do most of these actions, but only if it is connected to the place where the information lives. A calendar synced both ways, so that an appointment the secretary moved also shows up for the agent. Clinic management software that connects over an API. And for systems with no ready connection, a custom integration project. For every action in this pile you ask two questions: where does the information live, and what may the agent do alone, without approval.

Pile C: judgement

Here is what must not be closed automatically: a complaint, a refund request, "I was charged twice", a medical or legal question that needs a professional, an angry customer, an unusual request, a large deal that someone needs to manage.

The agent does not disappear from this pile. It receives the enquiry, collects the details (name, what happened, when, a photo if there is one), tells the customer exactly what happens next, and hands over to a person with a summary. The decision stays with a person. What changes is that the person gets an orderly enquiry, not a "hi, can we talk?" with no context at all.

Step 3: what you prepare for each pile

For pile A: one page per topic

Go through the questions in the pile, group them by topic, and write one clear answer per topic, in the words you would use with a customer on the phone. Check three things: that the information is current (winter hours, this year's price list), that there are no two versions that contradict each other, and that there is an answer even when the answer is "no". For example, "there is no private parking, there is a public car park two minutes' walk away" beats silence.

For pile B: a list of actions and permissions

Write one line per action: what the customer asks for, which system it happens in, and where the limit is. Examples of limits businesses set: moving an appointment is allowed up to 24 hours before, a late cancellation goes to a person; an invoice copy is sent only to the number on the customer's card; a payment link is sent only after the appointment is booked. That way the agent knows where it acts alone and where it stops.

For pile C: a handoff rule

Four things must be written down from day one: who gets the enquiry (a name, not "the office"), in which channel they get it, what the summary includes, and what the agent promises the customer. A line like "someone will get back to you tomorrow by nine" is the business's promise, and someone has to keep it. An enquiry that was handed over and nobody opened is worse than one that was never answered, because the customer has already been given a promise. How to connect the handoff to the CRM so that it really reaches a person is explained in the article on the hot lead handed to nobody.

Step 4: what to connect first

The order that works for most businesses:

1. Pile A, on all channels at once. If the agent answers about hours and parking on WhatsApp but not on Instagram, whoever writes on Instagram waits for exactly the same answer.
2. The handoff rule for pile C, on the same day. Even an agent that answers only information questions will get a complaint on day one, and it needs to know what to do with it.
3. From pile B, the action that repeats most in the sheet. In a clinic it is almost always booking and moving appointments. In a shop it is usually "where is my order". Connect it, check it for a month, and only then add the next one.

Three mistakes that keep coming back

• Building from the FAQ page instead of the messages. The result is familiar to anyone who has already tried a simple bot: excellent answers to questions nobody asks, and "I didn't understand, try again" for everything else.
• Putting complaints in pile A. A complaint sometimes looks like a question ("why was I charged twice?"), but it needs a check and a person, not a general explanation of the billing policy.
• Forgetting the evening and the weekend. If the sheet shows that many enquiries arrive after working hours, that is exactly where an AI agent gives the first answer, instead of the customer waiting for morning. The handoff to a person can wait for morning, the first answer cannot.

What to count after a month

After a month, go back to the sheet, this time with the data from the CRM, and count four things:

• How many enquiries the agent closed on its own, without a handoff, and out of how many.
• How long it took to the first answer, during the day and in the evening.
• How many of the enquiries handed to a person were answered by the time the agent promised.
• Which questions the agent could not answer. Each one is a new line on the pile A page, or a new action for pile B.

The first number is not supposed to reach one hundred percent. A business where every enquiry closes without a handoff is probably missing the enquiries where a person would have closed a sale or saved a customer.

How this looks at WhaleBiz

WhaleBiz is a technology company that wrote and runs an autonomous sales system: AI agents for WhatsApp, the website, Instagram, Telegram and Facebook, with a built-in CRM. In customer service this means the agent answers from the business's documents in any language, understands voice messages and photos, books appointments in a calendar synced both ways, and saves every conversation on the customer's card. Whatever needs judgement goes to a person with a summary. The details are on the AI agent for customer support page, and what each plan includes is on the pricing and plans page.

You can do the exercise itself without any system at all. A spreadsheet, 100 enquiries and three piles will tell you more about the service in your business than any demo.

Frequently Asked Questions

How many enquiries do I need for the sorting to give a real picture?

Usually the last 100 enquiries are enough to see what repeats. A business that gets few messages can go a month back. Include every channel, voice messages and photos too, and split an enquiry that holds two requests into two rows.

Can an AI agent handle complaints?

It can receive the complaint, collect the details, tell the customer what will happen and when, and hand over to a person with a summary. The decision, for example a refund or compensation, is made by a person. The customer gets an answer at once, and the person gets an orderly enquiry.

Do I need to replace my calendar or management software to connect an agent?

No. The agent connects to calendars with two-way sync and to clinic management software over an API. For a system with no ready connection, a connection can be built as a custom project.

What happens to enquiries that arrive at night and at weekends?

The agent closes pile A questions and the pile B actions you connected at any hour. An enquiry that needs a person gets a first answer at once, with a clear promise of when someone will get back, and goes to a person at the start of the next working day.

How much does customer service automation cost?

It depends on the channels and on what you connect. What each plan includes is on the pricing page of the site, and connections to special systems are priced as a separate project.