How much does it cost to build an AI chatbot in 2026?

"AI chatbot" covers very different products. Name the type you need before you ask for quotes, since each adds work to the last.

• FAQ bot over your documents. Uses retrieval-augmented generation (RAG): the bot searches your help center, policies or product docs for the relevant passages, and a language model answers from them, ideally with links to the source.
• Support assistant with human handoff. Answers common questions, looks up order or account status, and passes the conversation to a human agent, with its history, when it is unsure or the customer asks.
• Internal knowledge assistant. Helps your team search wikis, SOPs, contracts and past tickets. It needs company login and permission-aware search, so people only get answers from documents they are allowed to open.
• Agent that takes actions. Calls your systems through defined tools to create tickets, update records, book appointments or start refunds. It needs confirmation steps, audit logs and strict limits on what it may do.

A demo that answers a few questions takes days. A chatbot you can put in front of customers takes much longer, and these six factors explain the gap.

• Data preparation. Collecting documents, removing outdated versions, extracting text from PDFs, tables and scans, splitting content into sensible chunks and keeping the index in sync when documents change. Messy source material is a common reason AI budgets grow.
• Retrieval quality. Many wrong answers come from fetching the wrong passages, not from the model. Good retrieval combines semantic and keyword search, filters by product or region, and is tuned against real questions.
• Guardrails and evaluation. A test set of real questions with expected answers, scored after every change, plus rules for refusing out-of-scope questions, resisting prompt injection and keeping personal data out of logs.
• Integrations. Your website, WhatsApp, Slack, a helpdesk, a CRM or your own APIs. Each needs authentication, error handling and a plan for when the other side is down.
• Admin and review tools. A screen to read conversations, flag bad answers, upload documents and see usage. Without it, nobody can improve the bot after launch.
• Usage-based model costs. Model providers charge per unit of text processed, so every conversation costs money every month. Agents cost more because one task can take several model calls.

The ranges below are general patterns for custom builds by Indian studios and agencies, not quotes. One lakh is ₹1,00,000, about US$1,100. Studios in the US, UK or Australia usually charge more for the same scope.

Every row also carries a monthly model and hosting bill. If your needs are simple and your data is public, an off-the-shelf chatbot tool on a monthly plan may be enough; custom software vs off-the-shelf covers that trade-off.

The build is paid once. Model usage, search hosting, monitoring and maintenance recur every month. Model prices and versions also change, and older models get retired, so plan for occasional re-testing.

• Cache what repeats. Serve answers to common questions from a cache, and embed each document once, re-processing it only when it changes.
• Use smaller models where possible. Route simple steps, such as classifying a question or rewriting a search query, to a smaller, cheaper model and save the larger model for the final answer.
• Send less text. Pass the few most relevant passages, not whole documents, and trim long chat histories.
• Set usage limits. Cap messages per user and per account each day, add rate limiting, and set budget alerts and spending caps with your model provider.
• Measure cost per conversation. Log the text processed per request so you know what each conversation, customer and feature costs.

If you charge your own customers for AI features, make sure your plans cover that cost. SaaS pricing models explains when usage-based pricing fits.

StackWise builds LLM features, RAG over private documents and agents with leading commercial and open-source models. AI development is quoted after discovery, where goals, users and constraints are agreed along with scope, timeline and a written quote. NDAs are signed on request before you share sensitive documents.

To make discovery useful, bring sample documents, 30 to 50 questions your customers or team actually ask, the systems the bot must connect to, and any rules about where data may go. During the build you get weekly demos and a live staging URL, and at handover you own the code and every account. For an example of AI in a shipped product, see StartupDeckAI, a full-stack AI SaaS with an automated test suite.