“How much does an AI chatbot cost?” has two answers, and confusing them is where budgets go wrong. There is the one-time cost to build it, and the ongoing monthly cost to run it. A simple FAQ widget and a chatbot that answers accurately from your own data, captures leads, and hands off to a human are very different projects. Here is how to scope both without guesswork, US-first.
What drives the build cost
The build is where most of the spend lands, and a short list of decisions moves the number more than anything else:
- Scripted vs. grounded in your data. A button-and-flow bot that follows a decision tree is cheap. A bot that answers free-form questions accurately about your pricing, policies, and products needs retrieval-augmented generation (RAG): your content chunked, embedded, and stored in a vector database like MongoDB Atlas Vector Search. That ingestion pipeline is the single biggest line on a serious chatbot.
- Where it lives. An embeddable widget on one marketing site is straightforward. The same brain wired into Slack, WhatsApp via Twilio, your help desk, and your app is several integrations, each with its own auth and error handling.
- What it can do. Answering is one tier. Taking action — booking a meeting, creating a ticket, looking up an order, issuing a refund — turns the chatbot into an AI agent, which is more valuable and more work to build safely.
- Human handoff and lead capture. Routing a stuck conversation to a person, capturing an email into your CRM, and logging the transcript are the unglamorous pieces that decide whether the bot is useful in a real business.
What it costs to run each month
The recurring bill has three parts, and for most sites it is far smaller than people expect:
- Model usage. You pay per token for the model that generates answers — Claude, an OpenAI model, or Gemini. Cost scales with how many conversations you have and how much context you feed each answer. Tight retrieval (sending only the few most relevant chunks, not your whole knowledge base) keeps this low.
- Embeddings and the vector store. Embedding your content is mostly a one-time cost, refreshed only when content changes. The vector store itself is inexpensive at typical knowledge-base sizes.
- Hosting. The widget and a small backend run cheaply — often a modest fixed bill on AWS for a marketing-site volume.
For a typical US marketing or support use case, the monthly running cost is usually small relative to the support hours it saves. The cost discipline is engineering, not luck: choosing the right model for the job, capping context size, and caching repeat questions.
Where teams overspend
The most common waste is reaching for the biggest, most expensive model for every answer when a smaller one handles 80% of questions perfectly. The second is stuffing the entire knowledge base into every prompt instead of retrieving precisely — that inflates the token bill on every single message. The third is building integrations nobody has asked for. Start with the one channel your customers actually use.
How to get an accurate quote
To get a real number rather than a range, write down four things: where the chatbot will live, whether it just answers or also takes action, what content it should answer from, and what happens when it cannot help. With that, a studio can map your bot to a concrete build, a monthly running estimate, and a plan. We build exactly this kind of grounded, streaming, lead-capturing assistant as part of our AI chatbots service — the same techniques behind the MCP-connected, AI-driven features in our own products like SocialPatra.
If you want a grounded estimate for your specific site and use case, send us the questions your customers ask most and where you want the bot to live at info@kodetra.com and we will map it to a plan, timeline, and price.