AI chatbot development: what a production chatbot actually needs.

01From FAQ bots to AI chatbots

The chatbots most people remember were decision trees: press 1 for billing, press 2 for delivery. They never surprised you, and they rarely helped. Modern AI chatbots understand free text, follow a conversation, and write a real answer.

That flexibility is also the risk. A language model on its own will answer confidently whether or not it knows. For a business, a confident wrong answer about price, policy or eligibility is worse than no answer. So the question is not whether to use AI, but how to keep it honest.

02How a RAG chatbot works

Retrieval-augmented generation keeps the model tied to your facts. It works in four steps:

1. Index. Your help articles, policies, product sheets and past answers are split into passages and stored in a search index, often a vector database such as Postgres with pgvector.
2. Retrieve. When a question arrives, the system finds the few passages most relevant to it.
3. Answer. The model writes a reply using only those passages, in your tone.
4. Cite or decline. The reply links to its source. If nothing relevant was found, the chatbot says so and offers a person.

Done well, this gives you answers that are both natural and checkable. Done badly, with poor retrieval or no refusal rule, it gives you the same confident mistakes with extra steps.

03What separates a production chatbot from a demo

• A content pipeline. When a policy changes, the index updates the same day. Stale answers are the most common failure after launch.
• An evaluation set. A few hundred real questions with the right answers, run before every change, so quality is measured, not guessed.
• Guardrails. Topics it must not handle, promises it must never make, and wording it must use for regulated subjects.
• Hand-off to a person with the full transcript, so the customer never repeats themselves.
• Access control. Internal documents stay internal; a customer only sees what they are allowed to see.
• Analytics on what it couldn't answer. That list is your best guide to which content to write next.

04Which channel to start on

Start where your customers already talk to you. For most businesses selling in India that is WhatsApp, where replies are fast and conversations continue over days. A website widget suits pre-sales questions from new visitors. An internal assistant in your team's chat tool suits policies and procedures. The same knowledge base can serve all three, so start with one and add channels later. If WhatsApp is your first channel, read our guide to WhatsApp automation before you build.

05What an AI chatbot costs

Build cost depends on how many sources you index, how many systems the chatbot needs to act in (booking, order status, CRM), and which channels you launch on. A chatbot that only answers questions is small. One that also looks up orders and books appointments is a product. At Michi Studios a discovery sprint starts from $1,500 and a production build from $8,000, with a fixed quote before work starts.

Running cost is mostly model usage per conversation plus hosting for the index. With sensible model choices it is usually modest per conversation, and we estimate it at your real volumes during discovery. See the full breakdown in AI app development cost.

06Questions to ask an AI chatbot development company

1. How will it answer only from our content, and what happens when the answer isn't there?
2. How do we update the content, and how fast does the chatbot reflect it?
3. Can we see the evaluation set and the accuracy on it before launch?
4. How does hand-off to a person work, on each channel?
5. Which models will you use, and can we switch later?
6. Where is our data stored, and is it used to train anything?
7. Who owns the code, the index and the prompts?
8. Who watches quality after launch, and how?

07Questions people ask