Building an AI chatbot for your business: a complete guide

Start with the problem, not the technology

The most common mistake in chatbot projects is starting with the question “which AI model should we use?” The right questions are: what will the chatbot do, who will it serve, and how will success be measured? A furniture store chain that needs a chatbot to advise on designs and give preliminary quotes has very different requirements from a logistics company that needs shipment tracking, or an HR department that wants to answer questions about leave policy.

List the questions customers or employees have asked most often over the past three months, pulled from your Facebook Page inbox, Zalo, email and call center. Usually, a small group of topics such as shipping fees, returns or order status accounts for the bulk of the questions. That is the sensible scope for your first version, rather than trying to make the chatbot a “know-it-all” from launch day.

Choosing an architecture: scripted, LLM or RAG

There are currently three main approaches, each suited to a different level of complexity and budget. You don't necessarily need the “smartest” option; choose the one that is good enough for the job and easy to control.

In practice, many effective systems combine approaches: tasks with clear rules, such as booking appointments or looking up orders, are handled by fixed flows or API calls, while open-ended questions go to RAG. For sensitive data, businesses can consider open-source models such as Llama or Qwen deployed on their own infrastructure instead of sending data to external services.

• Scripted chatbots: use buttons and predefined conversation flows. Low cost and accurate within the designed scope, but rigid when customers ask something off-script.
• Large language model (LLM) chatbots: call models such as GPT, Claude or Gemini directly to understand natural-language questions. Flexible, but prone to wrong answers without your company's own knowledge.
• RAG (Retrieval-Augmented Generation): before answering, the system searches your company's knowledge base for relevant passages and hands them to the model as grounding. This is the balanced choice for most businesses.

Preparing data: the foundation of correct answers

A chatbot only answers well within the data it's given. If your warranty policy exists in three different versions scattered across Google Drive, the chatbot will give contradictory answers, just like an untrained new hire. That's why data preparation often takes more effort than the programming itself.

Don't forget to build a set of sample questions with approved answers, drawn from real conversation history. This dataset is used for testing before launch and again after every document update or model change.

• Gather and curate sources: FAQs, product catalogs, price lists, policies and the sales scripts of your best salespeople.
• Clean and standardize: remove outdated documents, unify terminology and clearly note the effective date of each policy.
• Chunk sensibly: split documents into topic-based passages so the search is more accurate, rather than feeding in an entire PDF dozens of pages long.
• Assign owners: each content group needs someone responsible for updating it when prices or policies change.

Integrating with communication channels and internal systems

The real value of a chatbot lies in its ability to connect. A chatbot that only gives generic answers will soon be ignored; one that can look up orders and inventory or book appointments genuinely saves time.

In the Vietnamese market, the usual priority channels are the website, Zalo Official Account and Facebook Messenger, since customers are already used to messaging there. Behind the scenes, the chatbot can connect to a CRM, sales software, ERP or ticketing system through APIs. Any action that writes data, such as creating an order or rescheduling an appointment, should include a clear confirmation step with the user before it's carried out.

Another must-have is handoff to a human. When customers are frustrated, want to file a complaint or ask about something beyond the chatbot's scope, it must transfer the conversation to a staff member along with the full context, so customers don't have to start over.

Managing quality, security and the risk of “hallucinations”

Language models sometimes produce answers that sound very convincing but are factually wrong, a phenomenon commonly called “hallucination.” A wrong answer about pricing or warranty terms can cause real damage, so multiple layers of control are needed right from the design stage.

Before launch, run the sample question set and have your customer care team try to trip the chatbot up. Rolling out in stages, for example only on the website or only during business hours when staff can supervise, helps you catch errors while the consequences are still small.

• Answer only from the documents provided, say clearly “I don't have that information yet” when unsure, and link to the relevant policy page.
• Don't let the chatbot make commitments on pricing, discounts or legal matters unless approval rules are in place.
• Limit the collection of sensitive information such as ID numbers or bank account numbers, and comply with current regulations on personal data protection.
• Segment data access: a customer-facing chatbot must not be able to access internal documents.

Measuring results and improving continuously

A chatbot isn't a build-it-and-forget-it project. After launch, regularly track metrics tied to your original goals instead of simply counting messages.

The questions your chatbot can't answer yet are your to-do list: add documents, fix contradictory content or design new handling flows. GREEN TECH typically works alongside businesses from data assessment and prototyping through to running and optimizing the chatbot after launch.

• Self-resolution rate: the percentage of conversations that end without being handed off to staff.
• Accuracy: the share of correct answers when a random sample of conversations is reviewed each week.
• Satisfaction: quick post-conversation ratings, combined with a close reading of negative feedback.
• Business impact: the number of leads, appointments and orders generated by the chatbot.
• Cost: operating cost per conversation compared with the cost of handling it manually.