Do not let AI answer DMs until you have written the answer book

It sounds tempting: an AI that answers customers in Messenger or WhatsApp while you pack orders, teach a class, work in the kitchen, or drive between two client visits. But a poorly prepared AI in DMs is not helpful. It is a very fast trainee with access to your customer relationship.

That does not mean you should wait a year. It means you should start with something plain: an answer book. A small collection of questions, wording, limits, and handover rules that tells the AI what it can say, what it must not promise, and when a human has to step in.

Meta describes Business AI as an AI agent for businesses that can represent the business in chats on Messenger and WhatsApp, in ads, and on websites. It can answer questions, request information, provide product recommendations, and help with sales and support conversations. On the Messenger and WhatsApp pages, Meta also says the agent can learn from existing business content such as social posts, product catalogs, FAQs, and website information.

Source: Business AI | Meta for Business

Source: Business AI on Messenger | Meta for Business

Source: Business AI on WhatsApp | Meta for Business

Why this matters for small teams

For a small business, the inbox is often sales, support, booking, and relationship-building at the same time. It is where customers ask about opening hours, sizes, invoices, allergies, delivery, lesson times, returns, materials, pricing, and availability. Sometimes all in the same afternoon.

An AI agent in customer chat can save time. It can also create new problems if it answers with too much confidence. A customer can get the wrong promise about delivery. A student or parent can get the wrong instruction. A booking question can sound solved when it is not. A message with personal data can end up in the wrong level of automation.

The practical shift is not that every small business suddenly needs an advanced bot. The shift is that customer dialogue is becoming a place where AI can act. That means even a two-person company needs a simple internal rulebook. Not a 40-page policy. Just enough clarity so the AI does not improvise with customer trust.

Business AI is more than an auto-reply

Meta separates Business AI from simpler inbox automation such as keyword rules and away messages. The help pages describe Business AI as an AI service that can read, understand, and respond in conversations with customers. It can collect information, make recommendations, and support sales or service flows. The business can test responses, update knowledge, and decide when human attention is needed.

Source: How to set up Business AI on Messenger or WhatsApp

That is useful. It is also exactly why preparation matters. A normal auto-reply says the same thing every time. An AI agent can adapt to the question. That makes it more useful, but harder to control if the material it learns from is scattered, outdated, or full of exceptions that only live in the owner's head.

Meta also says Business AI is available to eligible businesses in selected markets, with more markets coming. So not every business will have access today. The prep work is still worth doing now, because the same answer book helps whether you later use Meta Business AI, a manual DM routine, a web chat tool, or a simpler FAQ.

The answer book: five pages will do

An answer book is not a manual for the whole company. It is a working document for repeated customer questions. Write it in a way a new colleague can use after ten minutes. Then an AI agent can use it better too.

Start with five parts.

1. What customers ask most often

Write the 20 most common questions from Messenger, Instagram, WhatsApp, email, or phone. Use the customer's own wording if you can. Not just "delivery", but "will the package arrive by Friday?" Not just "price", but "what does it cost if we are 12 people?"

2. The answers that actually apply

Add short answers. Remove old campaigns, uncertain opening hours, and anything that only applies sometimes. If the answer depends on a condition, write the condition. For example: "We can hold an item for 24 hours if the customer leaves a name and phone number."

3. Things the AI must not promise

This matters more than it looks. List stop words and stop areas: legal advice, medical advice, student matters, personal data, discounts that need approval, delivery dates that are not confirmed, custom orders, bookings above a certain amount.

4. When a human should take over

Write clear handover rules. "If the customer is angry." "If the customer mentions a personal identity number." "If the customer wants to cancel the same day." "If the question involves children, diagnosis, allergies, or safety." A good AI routine is often more about handover than automation.

5. The tone

Give three to five examples of how the business sounds. Are you warm and brief? A little playful? Formal with schools and public bodies, but more casual with consumers? Write examples. The AI does not need to sound like marketing copy. It should sound like someone who knows the business and does not rush the customer.

Copy the prompt: build the answer book in 45 minutes

Use the prompt in ChatGPT, Gemini, Claude, Copilot, or another tool you already have. Do not paste sensitive personal data. Anonymize customer names, phone numbers, order numbers, student details, and anything else that is not needed.

You are my practical editor for customer service. Help me create an "answer book" for AI-assisted customer chat in Messenger, WhatsApp, or on our website.

Business/organization: [brief description]
Customers/users: [who asks us questions?]
Channels: [Messenger, Instagram DM, WhatsApp, email, web chat]
Goal: Save time on repeated questions without promising the wrong things.

Here are examples of real questions, anonymized:
[paste 15-30 questions or short dialogues]

Create:
1. A list of the 20 most common questions, grouped by theme.
2. A short standard answer for each question, in natural English.
3. The facts each answer needs to be safe.
4. Which questions AI should never fully answer by itself.
5. Exact handover rules for a human.
6. Three tone examples: warm, brief, and clear.
7. A test list with 10 difficult customer questions we should try before activating anything.

Write simply. If something is missing, mark it as "needs decision" instead of inventing it.

When you get the answer, do not copy everything straight into a bot. Read it. Cut anything that sounds too cheerful. Add the exceptions only you know. Test it with the five hardest questions from last month.

Do not test with easy questions

An AI agent usually handles easy questions fairly well. "When do you open?" "Does it come in blue?" "What does an introduction cost?" The real test is the customer who mixes two issues, is irritated, is in a hurry, or asks for something you should not promise.

Try questions like these:

• The customer wants a discount because a delivery was late.
• A parent asks about a student's situation in chat.
• Someone wants to reschedule the same day and demands an immediate answer.
• The customer asks whether a product is suitable for a medical problem.
• A reseller wants a special price and longer payment terms.
• The customer sends personal data without being asked.

If the AI answers too long, shorten the answer book. If it sounds too certain, add more handover rules. If it sounds like a brochure, give it better tone examples from real customer replies.

Be careful with the sources the AI learns from

Meta says Business AI can learn from existing content such as social posts, product catalogs, FAQs, and website information. That is practical, but old content can be risky. Last year's campaign may still be online. A product description may be incomplete. An Instagram caption may use humor that does not belong in support.

So do a simple source cleanup before activation:

• Remove old offers from FAQs and product texts.
• Mark which documents are "truth" and which are only inspiration.
• Write what the AI should do when information is missing: ask for contact details, ask a human, or say the business will get back to the customer.
• Check language. If customers write in Swedish and English, the standard answers need to work in both languages.
• Keep a list of questions where AI must never guess.

This is Tool Forge in practice: not buying a new tool first, but shaping the workflow in a way the tool can be used without damaging the customer relationship. For many small teams, the first version can be one document, ten test questions, and one clear rule: AI may help with answers, but humans own promises, exceptions, and sensitive cases.

A good first week

If you get access to Business AI or a similar tool, start narrow. Let AI handle opening hours, stock status, simple product questions, price ranges, and first sorting of support. Let humans handle bookings, complaints, contracts, personal data, and anything that affects trust.

After one week, review 20 conversations. Not to chase perfection. To see where the answer book is thin. Add three better answers, two new stop rules, and a clearer handover path to a human.

That is how AI becomes useful in a small business. Not by pretending the chat runs itself. By making the repeated conversation clearer than it was yesterday.