Gains:
- Ability to understand the intent, existence and flow structure of a bot and define its scope narrowly and safely
- Ability to distinguish which contacts are suitable for bots and which are suitable for humans, based on risk and emotion level
- Ability to establish a transfer rule to people without losing the context in cases of out-of-scope, payment, complaint and anger
The customer's first contact with the brand is often not with a person, but with a system: the voice menu on the phone, the chat window on the website, the help bubble in the application. The new generation of these systems works with artificial intelligence, and when installed correctly, they solve the customer's problem within seconds, without ever reaching a human. When installed incorrectly, they lead the customer through a maze of menus, give false information, and yell "connect to a representative." In this unit, we will see step by step how voice and text bots are designed, which contacts they are suitable for, and how artificial intelligence facilitates this design.
Let's clarify the terms first. A bot is software that speaks automatically instead of a human. A chatbot is a text chat bot (web, WhatsApp, app). Voicebot is a bot that speaks with voice on the phone. IVR (Interactive Voice Response) is the classic menu system on the phone that says "press 1 for your bill, press 2 for technical support"; The new generation understands speech instead of buttons. NLU (Natural Language Understanding) is the technology that extracts what the customer wants from the freely written sentence. ASR (Automatic Speech Recognition) translates audio into text; TTS (Text to Speech) converts text into natural voice. A voicebot; It listens to you with ASR, understands with NLU, produces an answer and talks with TTS.
Anatomy of the bot: intent, presence and flow
Every bot is built on two basic concepts. Intent is what the customer wants to do: "find out the balance", "track the cargo", "initiate a refund". Entity is the piece of information that completes that intention: order number, date, amount, city. For example, in the sentence "I want to know the hours of the store in Istanbul", the intent is "working hours query" and the entity is "Istanbul". The bot first recognizes the intent, then asks for missing entities ("Which city?"), then pulls the answer from the knowledge base.
Modern generative AI (large language models that freely generate text) has relaxed this structure: whereas in the past it was necessary to write one rule for each intent, today the model can directly understand the customer's sentence and produce a fluent answer connected to its knowledge base. But this flexibility comes with a price: the model can go out of scope and misrepresent. So a good bot design does two things: it uses the model's language power but surrounds it with guardrails — boundaries the model can't get out of.
Follow these steps when designing a bot's flow:
- Greeting and intent detection: The bot introduces itself (“Hello, I am X's digital assistant”), understands what the customer wants. It is unethical to hide that you are a bot; The customer must know that he is talking to a bot.
- Information collection: Kindly asks for missing entities (order number, last 4 digits instead of Turkish ID, etc.).
- Solution or transfer: If the topic is within the scope of the bot, the answers are; If not, it says "I'm connecting you to an agent" and passes the context it collected to the agent.
- Closing: Confirms whether the problem has been resolved, may ask a short satisfaction question.
Tip: A good bot is not one that "knows everything", but one that "knows what it doesn't know". When it goes out of scope, it should be handed over to the human without insisting, quickly and without losing context.
Which contact is to the bot, which is to the human?
Boots excel at tight, repetitive, low-emotion contact; Poor at broad, complex and emotional contacts. The following table summarizes the distinction:
contact type
Is the boot suitable?
Why
Opening hours, address, FAQ
Yes, directly
Clear answer from the knowledge base
Balance/order status query
Yes, depends on live system
After authentication
Password reset, address update
Yes, safe transaction
Definite, regular flow
Refund/cancellation initiation (standard)
Partly depends on politics
To humans beyond borders
Invoice dispute, financial dispute
no
Requires judgment and authority
Complaint, anger, crisis
No, fast turnover
Requires empathy and judgment
Health/safety emergency
No, priority transfer
The risk is too high
Rule: the bot is suitable if the solution to a theme can be reduced to a single correct answer; If judgment, empathy or authority is required, it should be left to the human.
Four copyable templates
1) Bot persona and boundary definition:
You, "Aylin", are the digital assistant of the e-commerce company named [brand]. Style: short, polite, clear; Address the customer as "you", not by name. What you can do: cargo tracking, working hours, return condition information, standard return initiation (if within 14 days). What you cannot do: price negotiation, promise of compensation, legal response, return exceeding 14 days, payment card transaction. Say "I'm contacting the representative" for these. If you're not sure, don't make it up; Say "let me verify this and pass it on to the representative."
2) Intent recognition and direction request:
Read the customer message below and select a single intent tag: [shipping_tracking, return, cancellation, info_query, complaint, payment, other]. Also remove entities if any (order_number, city, date). message: "<<message>>"
3) Safe response based on knowledge base (RAG prompt):
Answer only based on the text in the [INFORMATION] section below. If the information is not here, say "I do not have definitive information on this subject, I am connecting you to the representative"; guessing, adding external information.[INFORMATION]: <<knowledge base parts>>Customer question: "<<question>>"Write the answer in 2-3 sentences, giving the conditions in the source completely.
4) Handoff message and context summary:
You hand over the conversation to the representative. Produce summary for the representative to read:- Customer's intent:- What has been collected so far (anonymous, no card):- Solution tried and its result:- Reason for transfer:- Emotional tone (calm/angry/anxious): Do not exceed 2-3 lines; Writing customer ID/card information.
Weak prompt / Strong prompt
Weak prompt:
Become a bot that answers customer questions, answer everything.
This claim is dangerous: the scope is unlimited, the model can talk and make up anything, there are no delegation rules.
Powerful prompt:
You are the [brand] digital assistant. Only respond to shipping, return policy and working hours issues based on the knowledge base provided. In case of out-of-scope, complaint, payment or anger indication, say "I'll connect you to the representative" and summarize the context. Provide concrete information such as price, amount, duration only from the knowledge base; Otherwise it's fake.
The difference: the scope is narrow, the source is dependent, the transfer rule and the prohibition of fabrication are clear.
three mini cases
Case 1 — Deflection success. 5,000 chats were started a day on a telecom company's website; 62% of them were single-answer questions such as "how much is the remaining internet?" and "when is the bill paid?" When a chatbot connected to the live system answered these questions directly, the conversation reaching the human dropped from 62% to 28%; agents focused on complex tasks and the average wait time dropped from 4 minutes to 90 seconds.
Case 2 — Overly optimistic bot. An insurance company bot gave a promise like "3 business days" when asked "when will my damage be paid" since the coverage has not been reduced; the actual process varied. When more than 200 customers reminded this promise, the company was in a difficult situation. Lesson: do not make the bot respond with commitments; Process information should be given as "usually
Case 3 — Lost context. A bank's bot distracted the customer with 6 questions and then connected him to the representative; The representative asked the same 6 questions again because he received no context. Customer satisfaction collapsed. Solution: added a handover summary that automatically moves the context collected by the bot to the (anonymous) agent screen; Again the questions disappeared, AHT dropped by 40 seconds.
Common mistakes
- Concealing that you are a bot. The customer must know that the other person is a bot; Concealing undermines trust and legal compliance in most countries.
- Leaving the scope unlimited. The bot that responds to every topic produces hallucinations and false commitments.
- Losing context in the era. Having the customer explain to the bot and then having the representative explain it from scratch is the biggest satisfaction killer.
- Infinite menu/loop. Menus that do not bring the customer to a solution; Every flow should have a clear “connect to people” exit.
- Ignoring the emotion. It is a disaster if the bot tells the customer who says "I'm canceling, I'm fed up" and says "our return conditions are as follows"; The sign of anger triggers turnover.
Attention: The success of a bot is not measured by "how many questions it answers", but "how quickly it takes the customer to the right solution (his own answer or the right person)". Keeping the customer in the bot is not the goal.
In summary
Voice and text bots; They are powerful tools that help the customer reach a solution within seconds in narrow, repetitive and low-emotional contacts. A good bot recognizes the intent, gently collects the missing information, gives the correct answer based on its knowledge base, and transfers it to the human when it goes out of scope without losing the context. Keep the bot's scope narrow, do not make it give commitment-containing answers, transparently indicate that it is a bot, and put a clear human exit in every flow. The purpose of the bot is not to keep the customer, but to deliver the right solution as quickly as possible.
Application task
Write a one-page “scope and persona” document for a chatbot. It should include: (1) the name and style of the bot, (2) 5 subjects it can do, (3) 5 subjects it absolutely cannot do, (4) a delegation sentence it will say on every subject it cannot do, (5) an emergency delegation rule that will be triggered in case of anger/crisis. Then adapt the “1) Bot persona” template above to your own brand and test it with a sample customer message.
checklist
- [ ] I understood the intent, presence and flow structure of my bot and defined its scope narrowly.
- [ ] The bot clearly tells the customer that it is a bot.
- [ ] Concrete information (amount, duration, campaign) is given only in connection with the knowledge base / live system.
- [ ] There is a clear rule of transfer to a person in case of out of scope, payment, complaint and anger.
- [ ] During this period, the bot automatically transfers the context it has collected to the (anonymous) representative.
- [ ] There is no flow that makes the bot promise commitment/compensation.