What to Consider Before Building an AI Chatbot

Before building an AI chatbot, be clear about the specific job it will do, whether you have accurate and current content for it to draw on, how it will handle questions it cannot answer, and what it will cost to run and maintain. A chatbot is most useful when questions are frequent, answers are already documented somewhere, and a person can step in when the conversation needs judgment.

Write one sentence describing what the chatbot should achieve, such as 'answer common questions about admissions and capture interested parents as leads' or 'help staff find the correct process in our internal policies'. If you cannot write that sentence, the project is not ready. The job decides everything that follows: the content it needs, where it lives, what it can do and how success is measured.

• Good fit: repeated questions with documented answers, such as fees, timings, eligibility, policies, order status and how-to steps.
• Good fit: helping staff search long manuals, product specifications or internal procedures.
• Good fit: collecting details for an enquiry or booking request before handing over to a person.
• Poor fit: decisions with legal, medical or financial consequences for the user.
• Poor fit: questions where the answer depends on judgment, negotiation or information that is not written down anywhere.
• Poor fit: very low volumes, where a clear FAQ page and a contact form would do the job.

Not every chat needs a large language model. A rule-based bot with buttons and fixed answers is predictable and cheap to run, and it works well for a short menu of options, such as checking store hours or choosing a department. An AI chatbot understands free-text questions and can answer in natural language, which helps when questions vary widely. Sometimes the honest answer is that a better-organized FAQ page and a fast WhatsApp reply from a person would serve customers more reliably. Many useful systems combine all three: buttons for common paths, AI for open questions and a person for everything else.

Most business chatbots do not rely on what the AI model already knows. Instead, they search your own documents for relevant passages and ask the model to answer using only those passages. This approach is often called retrieval-augmented generation. Behind the scenes, documents are split into sections and converted into embeddings, which are numeric representations of meaning that allow vector search to find passages related to a question even when the wording differs. The quality of the answers depends heavily on the quality of these documents.

• Collect the sources: FAQs, policies, brochures, price guides, manuals and past support replies.
• Remove outdated documents, or the chatbot will quote last year's fees.
• Resolve contradictions between documents before the chatbot finds them for you.
• Assign an owner to each document who is responsible for keeping it current.
• Structure content with clear headings and one topic per section, which improves retrieval.
• List topics the chatbot must not answer and where it should redirect people instead.

AI models can produce answers that sound confident but are wrong. This is often called hallucination. It can happen when the right document was not found, when the documents are unclear, or when the model fills a gap with a plausible guess. No setup removes this risk completely, so the design should reduce it and limit the damage when it happens.

• Instruct the chatbot to answer only from retrieved sources and to say so when it cannot find an answer.
• Show the source document or page link with each answer, so users and staff can check.
• Keep the scope narrow, especially at launch.
• Never let the chatbot commit to prices, discounts, delivery dates or refunds unless the figure comes directly from a live system.
• For actions such as booking or cancelling, confirm details with the user and let the business system, not the model, carry out the change.
• Add a clear note that answers are automated and that a person is available.

A good chatbot knows when to step aside. Decide what triggers a handoff, who receives it, during which hours, and what the user sees outside those hours. When a conversation is handed over, the person should see the full transcript and any details already collected, so the customer does not have to repeat themselves.

• The user asks for a person, in any wording.
• The chatbot cannot find a relevant answer, or fails twice in a row.
• The message suggests frustration, a complaint or an urgent problem.
• The topic is on the restricted list, such as legal or medical questions.
• The lead looks high-value, such as a bulk order or a large project enquiry.

People type personal details into chat windows, sometimes more than you asked for. Decide what data the chatbot collects, where conversations are stored, who can read them and for how long they are kept. Check what the AI model provider does with data sent through its API and which retention settings are available. Mask or avoid sending sensitive identifiers where possible, and restrict internal chatbots so staff only retrieve documents they are allowed to see. Requirements under India's Digital Personal Data Protection Act, 2023 and any sector rules should be assessed with your legal advisers. Sectors such as healthcare need particular care about what the chatbot may say and store.

A chatbot can sit on your website, inside WhatsApp, in a customer portal or mobile app, or in an internal tool for staff. Each channel has different limits on formatting, message length and how users are identified. A chatbot becomes more useful when it is connected to your systems, for example to log leads in your CRM, check appointment slots or look up order status. Any feature that shows personal information must first confirm who the user is. Our pages on AI agents and chatbots and AI assistants and automation describe common setups.

• Build a test set of real questions from emails, call notes and WhatsApp chats, along with the correct answers.
• Include awkward cases: spelling mistakes, vague questions, mixed languages such as Hinglish, and questions outside the scope.
• Try to make it misbehave, for example by asking it to ignore its instructions or reveal internal information.
• Have people who know the subject review answers for accuracy and tone.
• Re-run the same test set after every change to content, prompts or model.
• Start with a soft launch to a small group or for limited hours before a full rollout.

Most AI models are billed by usage, usually based on the amount of text processed in each request and response. Costs rise with the number of conversations, the length of each conversation, how much document text is sent with every question and the choice of model. Practical ways to keep costs under control include using smaller models for simple tasks, limiting how much context is sent, caching answers to very common questions and setting usage alerts. These usage charges are paid to the model provider and are separate from development and hosting costs, so estimate them based on expected volume before launch and review them monthly afterwards.

• Update the knowledge base whenever prices, policies or products change.
• Review a sample of conversations regularly, especially those with low ratings or handoffs.
• Add answers for questions the chatbot could not handle.
• Re-test when the model provider releases a new version or retires an old one.
• Monitor costs, response times and errors from connected systems.

Avoid judging a chatbot only by the number of conversations. A busy chatbot that frustrates people is not helping. Look at a combination of signals: how many conversations ended with a useful answer without needing a person, how often handoffs happened and whether they were appropriate, thumbs up and thumbs down ratings, the list of unanswered questions, leads or bookings captured, and whether your team spends less time on repeated questions. Reading real transcripts each week often tells you more than any dashboard. For a wider view of where AI can help beyond chat, see practical ways to add AI to an existing business.

• You can describe the chatbot's job in one sentence.
• You have current, owned documents covering the main questions.
• You have listed topics it must not answer.
• Handoff triggers, owners and out-of-hours behavior are defined.
• Data collection, storage, retention and provider settings have been reviewed.
• A test set of real questions with correct answers is ready.
• Expected usage costs are estimated and someone is responsible for monitoring them.
• Someone owns ongoing content updates and transcript reviews.