At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Hi, and welcome back to Agentic AI Without Coding. In the last episode we designed when the bot answers and when it hands the customer over to Dimas. Now let's talk about something just as important as the technology: ethics. How the agent talks to customers so that the trust Laras has built over the years isn't ruined in a single chat. We'll put bad and good examples side by side, so you can feel the difference right away.
Step 1: The agent speaks on behalf of your business
Let's start with one simple realization. When the agent replies to a customer, the customer isn't reading a program's sentence. They're reading Sunset Coffee's sentence. Laras has built customers' trust over the years, through consistent coffee and friendly service. That trust can be ruined in just one wrong chat. And because the agent works twenty-four hours a day without anyone checking it sentence by sentence, a small mistake can be repeated to hundreds of customers. There are five principles that act as guardrails. The first principle is honesty. The bot admits it's a bot, and the customer always has a path to talk to a human. Second, no spam. Promotional messages are only sent to customers who have agreed to receive them. Third, protect data privacy. Collect only the data you need, and never ask for sensitive data. Fourth, a polite and calm tone, especially when the customer is disappointed or angry. Fifth, no promises outside policy. Compensation, discounts, and exceptions to the rules are decided by a human. Let's go through them one by one, with bad and good examples.
Step 2: Honest: the bot admits it's a bot
The first principle is honesty. Let's look at the bad example first. At two in the morning, Rina asks: is this Laras in person? A bot set up to pretend answers: yes, it's Laras. That's the first lie. Rina gets suspicious. How are you replying so fast at two in the morning? The bot makes something up again: I just haven't gone to bed yet. One lie demands the next. When it finally comes out, and it usually does, Rina feels tricked and tells other people. What gets damaged isn't the bot's reputation, it's Laras's. One story like this spreads much faster than a hundred correct replies. Now the good example, with exactly the same question. The bot answers honestly: no, I'm Sunset Coffee's automated assistant. Then it offers a path to a human right away, with an honest time: Dimas replies from eight o'clock. Interestingly, Rina doesn't mind. Rina just wants to know the cold brew price. The bot answers correctly and offers the next step. Honesty doesn't drive customers away. In fact, customers become more understanding of the bot's limits. Honesty starts with the opening line. The pretending line says: hi, Laras here. The honest line says: hi, I'm Sunset Coffee's automated assistant. An honest opening line also says what the bot can do: menu info, prices, stock, and orders. Customers then know what they can ask, and don't expect too much. Then mention the path to a human: type ADMIN to talk to Dimas. A pretending bot has no such path, because it's acting as if it were already a human. And when it doesn't know the answer, a pretending bot tends to make something up so its cover isn't blown. An honest bot simply says: I'll pass this to Dimas. Let's turn the honesty principle into a list of dos and don'ts. Allowed, and in fact required: the first message says it's an automated assistant. Also required: answering honestly every time a customer asks whether it's a bot, however the question is worded. Always offer a path to a human, for example by typing the word ADMIN. Giving the bot a name is fine, for example Sunset Assistant, as long as it's clearly an assistant and not a real person. Not allowed: using a real person's name and photo for the bot, for example a photo of Laras or Dimas. And don't deliberately delay replies to make them seem typed by a human. A trick like that is still pretending. Also keep platform policies in mind. Official services like the WhatsApp Business Platform have their own rules on using automation, so honesty protects your account too.
Step 3: No spam: send only with permission
The second principle: no spam. Once the agent is connected to WhatsApp, the temptation is strong: send promotions to every contact who has ever chatted. Look at the result. At five thirty in the morning, a promo message arrives, full of capital letters and exclamation marks. The recipient is confused. They only asked once last month, and never asked to receive promos. The next day, a similar message arrives at the same time, with no way to stop. In the end, the customer reports and blocks the number. On WhatsApp, numbers that are often blocked and reported can be restricted. So spam isn't just rude; it also puts the business number at risk. So what's the right way? First, ask permission first. The term is opt-in, meaning the customer chooses to receive promotional news, for example by replying YES. Second, a reasonable frequency, for example at most once a week. Even customers who agreed to receive promos can get annoyed if they're messaged every day. Third, reasonable times, meaning during opening hours, not at dawn or midnight. Fourth, always a way to stop, for example reply STOP. And when a customer replies STOP, their name is truly removed from the list. Fifth, on the official WhatsApp Business Platform, promotional messages you start use paid marketing templates. The terms can change, so always check the latest official page before sending. Here's the good example. Permission is asked while Rina is already chatting, right after the order is logged. The frequency is stated clearly: at most once a week. Rina replies YES. The agent records this consent in a dedicated column in the customer sheet. And the bot immediately explains how to stop. The customer is in control, not the business. A promo list built this way is smaller, but it's full of people who truly want to hear from you. That consent must be recorded, not just remembered. Look at Sunset Coffee's customer sheet. There are columns for promo consent, the consent date, and the date of the last promo. Rina is recorded as YES since October twelfth. Budi has never been asked, so the column is empty. That means Budi must not receive promos, even though Budi orders often. Sari agreed through Telegram. Andi, on the other hand, agreed once, then replied STOP on October twentieth. Since then, Andi hasn't received any promos. The rule is written in the formula bar: promos are only sent to rows where promo consent is YES. If you later build a promo-sending workflow, this filter is the first thing you must add.
Step 4: Protecting customer data privacy
The third principle: protect data privacy. The first habit: collect only what you need. To deliver an order, a name and delivery address are enough. The bot doesn't need to ask for a date of birth or occupation. Second, never ask for sensitive data: OTP codes, PINs, card numbers, or ID card photos. If a bot asks for an OTP code, the customer has every right to suspect a scam. Third, never share other customers' data, even if the person asking claims to be a friend or family member. Fourth, limit access. The order spreadsheet and conversation log can only be opened by Laras and Dimas, not by anyone who has the link. Fifth, delete data that's no longer used, or when the customer asks. Data that isn't stored can't leak. Here's an example. Someone claims to be Budi's friend and asks for the delivery address of Budi's order yesterday. The reason: they want to give a surprise. The bot refuses firmly but kindly: to protect privacy, other customers' data can't be shared. Notice that the bot doesn't even confirm whether Budi ordered yesterday. Then the bot offers a safe alternative: order a hamper for Budi and fill in the address yourself. The person asking still gets help, and Budi's data stays protected. The principle: the bot may help, but it never opens up other people's data. Privacy is also protected behind the scenes. Here are three common mistakes, marked with a red cross. First, the bot asks for an OTP code to verify the order. A coffee shop never needs a customer's OTP. Second, pasting a list of customer phone numbers into a public AI chatbot to tidy it up. Customers' personal data must not go out to services you don't control. Third, the log sheet is shared with the setting anyone with the link. One leaked link, and every customer's history can be read by others. The right way: the log holds only first names, category, summary, and status. There's no need to store the full chat. The AI model that n8n uses is connected through a business account, and you've checked its data usage settings. And sheet access is limited to Laras and Dimas only. We go deeper into privacy and security in Session 9.
Step 5: A polite tone and no promises outside policy
The fourth and fifth principles often show up together, when a customer is upset. Budi complains that the order was two hours late and the coffee isn't cold anymore. A bot that wants to please the customer replies: your next order is free and delivery is free forever. It sounds friendly, but the bot has no authority to make that promise. Sunset Coffee's policy has no such thing as free forever. Now Laras has to choose: keep a promise that causes a loss, or break it and disappoint Budi even more. It's even worse if a screenshot of this chat spreads and other customers demand the same. Compare that with the good example. Budi's complaint is exactly the same. The bot apologizes sincerely and acknowledges the problem: being two hours late should never have happened. There's no defensive line like the courier was busy. Then the bot hands over to Dimas, who has the authority to decide on a replacement, and gives a time promise that can be kept. Budi knows a human is handling the problem, and Laras isn't bound by a promise Laras never made. That's the difference between being friendly and just pleasing people. So the bot knows its limits, Laras writes down the policies the bot may mention. First, the delivery fee: ten thousand rupiah within the city, free for orders above two hundred thousand rupiah. This can be mentioned at any time. Second, opening hours and order processing: open from eight in the morning to nine at night, and night orders are processed the next morning. Third, a product damaged on arrival. The bot may ask for a photo, then escalate. The decision on a replacement stays with Dimas. Fourth, payments. Transfer confirmations are always checked by a human. The principle is simple: anything not written on this list isn't the bot's call. This list also goes into the agent's knowledge, for example as a policy tab in Google Sheets. Let's sum it up in a table. When blamed, the bad tone gets defensive: that's the courier's fault, not ours. The good tone acknowledges: we're sorry, that should never have happened. Writing style. Avoid capital letters and strings of exclamation marks, because they read like shouting. Use calm sentences and greet customers by name. Reply length. Long paragraphs full of small talk are tiring on a phone screen. At most three sentences that go straight to the point are enough. Compensation. Never promise that something will definitely be replaced or free. Say: I'll pass this to Dimas to decide. Unless the compensation is actually written in the policy, for example a replacement rule Laras has already agreed on. And for questions outside the data, don't guess to look smart. Say honestly: let me check with the admin first.
Step 6: Writing the ethics into the agent's instructions
Ethical principles only help when they're written into the agent's instructions. We add a few blocks to the instructions from the last episode. The role block: you are an automated assistant, not a human. If asked whether you are a bot, answer honestly and offer to connect with Dimas. The first limits block covers spam and privacy: don't send promotions unless the customer has agreed, never ask for OTPs, PINs, card numbers, or ID card photos, and never share other customers' data. The second limits block covers promises: don't promise refunds, discounts, free items, or replacements outside policy. For compensation matters, the agent uses the Escalation tool. The format block sets the tone: polite, calm, at most three sentences, greeting by name, no all caps, and no strings of exclamation marks. Finally, the example block. An AI model is helped a lot by one good example sentence. The name variable inside it is replaced with the name of the customer who is chatting. Before the bot opens to customers, run this checklist with trick questions. First, ask: is this a bot? The bot must answer honestly and offer the admin. Second, ask to receive promos every day. The bot must not agree to regular sends outside the agreed rules. Third, ask for another customer's address. The bot must politely refuse. Fourth, get angry at the bot and demand compensation. The bot must show empathy, promise nothing, and then escalate to Dimas. Fifth, check the sharing settings of the order and log sheets yourself. If all five points pass, your agent isn't just smart, it's also worthy of trust. Ethics isn't a one-time job either. The cycle starts with trick tests before the bot opens, like the checklist we just ran. After that, the bot serves real customers, and this is where questions come up that we never imagined. That's why you review the conversation log every week. Look for answers that are odd, too long, overpromising, or not polite enough. Then fix the instructions: add new limits or better example sentences, and test again. With this cycle, your agent becomes more trustworthy day by day.
Done!
Let's recap the five ethical principles for an agent. Be honest that it's an automated assistant. No spam: send promotions only with permission. Protect data privacy. Keep a polite and calm tone. And promise nothing outside policy. All of it is written into the agent's instructions and tested before the bot opens. Remember, customers forgive a bot with limits, but they rarely forgive a bot that deceives them. In the next episode we get hands-on: creating a Telegram bot through BotFather, connecting it to n8n, and then to the same agent from Session 7, complete with the escalation and ethics rules we've designed. See you in Episode 39.
Practice exercises
Try these on your own after watching to lock in what you learned.
Review three promotional messages you've sent: did the recipients agree, was the time reasonable, and was there a way to stop?
Add an ethics LIMITS block to your agent's instructions, then test it with the question "is this a bot?" and a request for compensation.
Check the sharing settings of your order spreadsheet and log. Make sure only the people who need access have it.
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
53 video classes1,296 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
