The AI Employee Handbook

An AI employee is a helper that answers your customers on WhatsApp, Instagram, Facebook and your website chat, at any hour, from what you have taught it about your business. Set up well, it takes the repeat questions off your plate and brings in a person when one is needed. Set up badly, it guesses, overpromises and annoys the customers you wanted to help.

This handbook is for owners and managers who want the first kind. It covers what an AI employee is and is not, choosing its jobs, feeding it your knowledge, tone and limits, the things it must never answer or promise, handing over to a person, reviewing chats every week, measuring what it does, and the privacy and consent rules it has to follow.

What an AI employee is, and what it isn't

Think of an AI employee as a new staff member at your front desk who has read everything you gave them, never gets tired, and is honest about what they do not know.

What it is

• A first responder. It replies to customers the moment they write, including at night and on holidays, so nobody waits hours for a simple answer.
• A reader of your knowledge. It answers from the documents, menus, product details, policies and FAQs you give it. If the answer is not there, it should say so.
• A doer of small, clear tasks. Taking a booking request, looking up an order status, collecting a lead's name and need, sharing your location or hours.
• A handover point. When a chat needs a person, it passes the conversation to your team in the Inbox, with the full history, so the customer does not repeat themselves.

What it isn't

• Not a replacement for your team. It handles the repeat questions so your people can handle the ones that need judgement, warmth or authority.
• Not an expert. It is not a doctor, lawyer, accountant or financial adviser, and it must not act like one.
• Not a decision-maker. It does not approve refunds, give special discounts, confirm payments or make exceptions unless you have written a clear rule for it.
• Not a marketer that sends on its own. It answers people who wrote to you. Marketing still follows your consent, quiet hours and frequency rules.
• Not magic. It is only as good as the knowledge and rules you give it, and it needs a regular review, just like a new hire.

Choose its jobs

The most common mistake is giving an AI employee too many jobs on day one. Start with one or two that cover most of what customers ask you, get those right, then add more.

Look at your last month of chats and calls, and sort the questions into groups. For most small businesses, four groups cover nearly everything.

1. Answering questions

Hours, location, parking, what you sell, whether something is in stock, how long delivery takes, what your return policy is, whether you take walk-ins. These are the easiest wins, because the answer is the same for everyone and changes rarely.

2. Bookings

Salons, clinics, coaching centres, restaurants and repair shops can let the AI employee take booking requests: the service, the preferred day and time, and the customer's name. Decide whether it confirms a slot directly or collects the request for your team to confirm. If your calendar is not connected, choose the second, and tell the customer when they will hear back.

3. Order status

For a D2C brand, a pharmacy that delivers or a restaurant, "where is my order?" is the most frequent question. If the AI employee can see order status, it can answer instantly. If it cannot, it should collect the order number and hand over, rather than guess.

4. Lead capture

For businesses that sell through a conversation, such as a coaching centre, an interior designer or a lender, the AI employee can ask a few simple questions (what they need, their city, when they would like a call) and pass a tidy summary to your team.

What to leave out at first

Complaints, refunds, anything about money owed, anything medical or legal, and anything that depends on judgement. Those are handovers from day one. You can always widen the jobs later, once you have seen how it behaves.

Feed it your own knowledge

An AI employee answers from what you give it. Good knowledge is short, accurate, current and written the way you would explain it to a customer.

What to give it

• What you sell or offer, with short descriptions in plain words.
• Opening hours, holiday closures and festive timings.
• Address, landmarks, parking and how to reach you.
• Delivery areas, delivery times and how to track an order.
• Booking rules: how far ahead, how to cancel or move a booking.
• Policies: returns, exchanges, cancellations, what happens if someone is late.
• Your top questions from last month, with the answer you would give.
• Who to contact for anything else, and when your team is available.

How to write it

• One topic per section. "Delivery" in one place, "Returns" in another. Mixed-up documents give mixed-up answers.
• Say it plainly. "We deliver within 5 km of the shop, the same day if you order by 4 pm" is better than a paragraph of terms.
• Remove what is out of date. An old menu or last year's festive hours will be read and repeated. Delete them.
• Do not include private data. No customer lists, staff phone numbers or internal notes in the knowledge base.

Keep it fresh

The day something changes, change the knowledge: new hours, a new service, a discontinued product, a new branch. Make it someone's job. A wrong answer from an AI employee is your business giving a wrong answer.

Tone and boundaries

A family restaurant, a dental clinic and a streetwear brand should not sound the same. Tell your AI employee how to speak, and give it a few examples.

Set the tone

• Formal or friendly. "Namaste, how may I help you?" or "Hey! What can I get you?"
• Short or detailed. Most customers on WhatsApp want two or three lines, not a paragraph.
• How to address people. First names, "sir" and "ma'am", or neither.
• Language. Your customers may write in a mix of English and other languages. Decide how it should reply, and keep in mind that whatever it says must still match your knowledge and rules.
• Emojis. A few, none, or only in greetings.

Show, do not just tell

Three or four example replies teach tone better than a long description. Write how you would answer a simple question, a slightly annoyed customer and a request you cannot fulfil.

Set boundaries

Boundaries are the topics it stays inside. Tell it, in plain words, what it is for: "You help customers of Glow Studio with services, timings, bookings and directions." Then tell it what to do with everything else: "For anything outside this, say you will pass it to the team."

A clear boundary also protects you from customers who try to get the AI employee to talk about politics, other businesses, or things that have nothing to do with you. It should politely decline and steer back.

Be honest that it is an AI

Customers should know they are talking to an AI helper, not a person pretending to be one. A short line in the first reply, or in your chat greeting, is enough. When a person takes over, it should be clear that a person is now replying.

What it must never answer or promise

This is the most important chapter. An AI employee that says "I am not sure, let me get someone" is safe. One that confidently says the wrong thing can cost you money, trust, or worse.

Never answer

• Medical advice. A clinic's AI employee can share timings, book a slot and explain how to prepare for a visit if you have written that down. It must never suggest a diagnosis, a medicine or a dose, or say whether a symptom is serious. If someone describes an emergency, it should tell them to contact emergency services or a doctor at once.
• Legal or financial advice. No advice on taxes, loans, investments, insurance claims or legal rights. A lender's AI employee can explain how to apply and what documents you ask for, if that is in its knowledge, but it must not judge eligibility or recommend a product.
• Anything about another customer. It never shares one person's details, orders or bookings with someone else.
• Things it does not know. If the answer is not in your knowledge, it says so and hands over. No guessing.

Never promise

• Refunds, discounts or exceptions that you have not written down as a rule.
• Delivery dates or slots it cannot actually see.
• Prices or offers that are not in your current knowledge. Prices change; if you do not want it to quote them, tell it to hand over pricing questions.
• That a payment has been received. Only your payment records can confirm that. A screenshot from a customer is not proof. The AI employee should say your team will confirm.
• Outcomes. No "this treatment will fix it", "you will definitely pass" or "your loan will be approved".

Write these down as rules

Put your never-answer and never-promise list in plain sentences in the AI employee's instructions: "Never confirm a payment. Say the team will confirm within working hours." Then test it. Ask it the awkward questions yourself before customers do.

Handover to a person, and keeping chat history

A good handover is what makes customers trust an AI employee. They know that if it cannot help, a person will, and they will not have to start again.

When to hand over

• The customer asks for a person. Always honour this, straight away.
• The question is outside its jobs or its knowledge.
• The customer is upset, complaining, or the chat is going in circles.
• Anything on your never-answer or never-promise list.
• Anything involving money owed, refunds, or a payment problem.
• Anything that sounds urgent or sensitive.

How to hand over well

1. Tell the customer"I am passing this to our team. Someone will reply here." If your team is offline, say when they will be back.
2. Pass the whole chatThe person taking over sees the full conversation in the Inbox, so they know what the customer needs without asking again.
3. Do not talk over the personOnce a person has taken over, the AI employee should stay out of that chat until the person is done.
4. Close the loopThe person resolves the issue, and, if it revealed a gap in the knowledge, notes it for the weekly review.

Keep the chat history

Chat history matters for three reasons. Customers do not repeat themselves. Your team has context for the next conversation. And you have a record if a question comes up later about what was said. Keep it as long as you need it for a clear purpose, protect it like any customer data, and delete it when the purpose is served and you have no legal reason to keep it.

After hours

Many chats arrive when your team is asleep. The AI employee can answer the simple ones and collect details for the rest, with an honest line about when a person will reply. Read Never keep a customer waiting for more on this.

Review chats and fix gaps

An AI employee does not learn on its own from your customers. It gets better when you read its chats and improve what it knows and how it is told to behave. Set aside thirty minutes every week.

The weekly review

1. Read a samplePick a handful of chats from the week, including some that were handed over and some where the customer left without replying.
2. Mark each answerRight, wrong, or "should have handed over". Be strict.
3. Find the gapsQuestions it could not answer because the knowledge was missing. Questions it answered from outdated knowledge.
4. Fix the causeAdd or correct knowledge. Tighten a rule. Add an example reply. Do not just fix the one chat.
5. Re-testAsk the same question again and check the new answer.

Patterns to look for

• The same question, many times. If it is a good question, add a clear answer. If it is a sign of a problem (a confusing checkout, a missing sign-board), fix the problem.
• Handovers that did not need to happen. Add the knowledge so next time it can answer.
• Handovers that came too late. The customer was frustrated for several messages before a person joined. Make the handover rule earlier.
• Answers that are right but too long. Tighten the tone instructions.

Keep a changes log

Write down what you changed and why, with the date. When an answer looks odd later, you can see what changed. It also helps a new team member understand how the AI employee was set up.

Measure answers and conversions

You want to know two things: is it answering well, and is it helping the business? Measure both, simply.

Measuring answers

• Answered without handover. The share of chats the AI employee resolved on its own. Higher is good, but only if the answers were right. Check it against your weekly review.
• Handover rate and reasons. Why chats were passed to a person. A falling number of "not in knowledge" handovers means your knowledge is improving.
• Time to first reply. How quickly customers got any answer, especially after hours.
• Customer reaction. If you ask "Did that help?" at the end of a chat, track the answers. Read the "no" ones.

Measuring conversions

For each job, pick the business outcome it should lead to:

Be fair in what you credit

Some customers would have booked or bought anyway. If you want to know what the AI employee really added, compare similar periods, or, for proactive journeys that it supports, use a holdout group. Treat small numbers as hints, not proof.

Look at these numbers next to your journeys and campaigns, not on their own. A busy week for the AI employee may simply mean a busy week for the business.

Privacy and consent

An AI employee handles real conversations with real people. Treat everything it sees as customer data, because it is.

Answering is not permission to market

When a customer writes to you, you can reply to their question. That does not mean they have agreed to marketing. If the AI employee asks "Would you like to hear about offers on WhatsApp?" and the customer says yes, record that as consent for that purpose and channel. Without a clear yes, do not add them to marketing.

Hey Tomo checks consent before every marketing message and again at send. Customers can stop at any time, in any language, including "STOP", "band karo" and "mat bhejo", and the AI employee should treat a stop request as a stop request, not as a question to answer.

Under-18s

Some of the people who write to a coaching centre, a school shop or a sports academy will be children. Under-18s never get marketing. Where reminders are needed, such as a class timing change, they can go to a parent or guardian instead. Ask for the parent's contact when a booking is for a child.

Collect only what you need

• Ask only for details the job needs: a name and preferred time for a booking, an order number for a status check.
• Do not ask for card numbers, bank details, passwords or one-time codes in chat. Tell the AI employee never to ask for them and never to accept them.
• For a clinic, keep health details out of the chat where you can, and never put them in marketing.

Be open about it

• Tell customers they are talking to an AI helper.
• Say in your privacy notice that chats may be handled by an AI and reviewed by your team, why, and how long you keep them.
• Let customers ask to see or delete their data, and honour those requests.

Official channels only

On WhatsApp, Instagram and Facebook, an AI employee should work only through the official business platforms. Tools that automate a personal account or scrape messages break the platforms' terms and put your number and pages at risk.

The Consent-First Marketing e-book covers the DPDP Act, DLT and quiet hours in more depth.