Chatbot Development Company

Rule-based or LLM chatbot - which do you actually need?

A rule-based bot follows flows you design: buttons, decision trees and intent classifiers of the kind Dialogflow and Rasa are built around. It is predictable, cheap per message and easy to audit, which suits order tracking, bookings and lead capture. It breaks when customers phrase things the flow never expected.

An LLM chatbot reads free text and writes its own replies, so it copes with messy questions and can answer from long documents. The trade-offs are a cost for every message, and replies that must be grounded and checked. Most of the bots we are asked for work best as a hybrid: fixed flows for anything that touches money or accounts, an LLM for open questions.

Which channels should the chatbot run on - web widget, WhatsApp or in-app?

Put the bot where your customers already ask questions. The channel changes the build more than most buyers expect:

• Web widget: the fastest to launch and fully under your control; it can show links, tables and forms, and it sees which page the visitor is on.
• WhatsApp: needs a WhatsApp Business Platform number and a verified Meta business. Messages you start outside the 24-hour customer service window must use templates Meta has approved, and Meta charges per template message by category and country.
• In-app chat: an SDK or screen inside your iOS and Android apps, where the user is already signed in, so the bot can act on their account safely.
• One conversation core behind all three keeps answers identical everywhere and means a new channel is an adapter, not a second bot.

How do you stop a chatbot from making things up?

By never letting it answer from memory. The bot first searches your own content - help articles, policies, product sheets, past tickets - and the model is instructed to answer only from the passages it found, and to say so when nothing relevant turns up. This is retrieval-augmented generation (RAG), and it is how the bot quotes your refund policy instead of an average of every refund policy on the internet.

Two rules matter as much as the retrieval. Data that changes by the minute, such as order status, stock and account balances, is fetched live from your APIs rather than copied into the index. And every answer is tested against a set of real customer questions before launch and after each prompt or model change.

When should the chatbot hand the conversation to a human?

Handoff is designed, not bolted on. Good triggers are specific: the search finds no relevant source, the customer asks for a person, the same question is rephrased twice, or the topic is one you never want a bot to settle alone - complaints, cancellations, medical or legal questions, anything involving a payment dispute.

When a person takes over, they should get the full transcript, the customer's identity and what the bot already looked up, inside the helpdesk they already use, such as Zendesk or Freshdesk. Outside working hours the bot should say when a person will reply and collect the details, rather than pretend it can solve the problem.

Can one chatbot handle several languages?

Usually yes, and the model is rarely the hard part. The work is in three places. Retrieval must find an English help article when the question arrives in Hindi or Arabic, which needs multilingual embeddings and testing per language. Replies must follow your tone and terminology in each language, which means a review by someone who speaks it. And handoff needs agents who can take over in that language, or a clear message when none are available.

On WhatsApp every template is approved per language, so each language adds a small approval step before launch. We scope languages one by one in the quote rather than promising all of them on day one.

What does a chatbot cost to run per conversation?

An LLM chatbot is billed by tokens, so the cost of a conversation is its number of turns multiplied by what each turn sends: instructions, retrieved passages, recent history and the reply. A six-turn conversation that sends 3,000 tokens of context per turn uses around 18,000 input tokens before any output. Multiply by your monthly conversations and your provider's current per-token price, then add hosting, the vector index and, on WhatsApp, Meta's template fees.

The levers are practical: answer the top intents with fixed flows that cost almost nothing, route easy questions to a smaller model, trim retrieved passages, summarize long histories instead of resending them, and cache frequent answers. We log token spend per conversation from day one so these choices are made on your numbers.

When is a chatbot the wrong answer - and when is an AI agent the better fit?

A chatbot is the wrong buy when customers ask a small, stable set of questions; a clear FAQ page and good site search will serve them better. It is also the wrong tool when the job is doing things rather than answering: issuing refunds, rescheduling deliveries, updating CRM records across several systems with decisions along the way. That is AI agent development - software that plans steps and calls your tools with permissions and an audit trail.

If you need a model trained or fine-tuned on your own data, LLM development is the right service; for classification or extraction from text with no conversation at all, NLP development. And if the chat assistant is itself your product, our ready-made ChatGPT clone starts from $2,799 and ships in 6 working days.