How to Build an AI Chatbot for Your Website

The line between a helpful chatbot and a website liability sits in the "in-between" conversations. Basic FAQs about hours or pricing are trivial to automate. Complex integrations with legacy systems need a human. But most daily inquiries - the repetitive questions that drain support time without creating tickets - fall squarely in the middle. That's where chatbots either prove their value or fail spectacularly.

This guide covers five specific processes for building an AI chatbot for your website that handles those mid-tier conversations: selecting the right LLM beyond ChatGPT, designing conversation flows that don't frustrate users, connecting a knowledge base without triggering hallucinations, testing accuracy before launch, and deploying without disrupting the current site.

Three ways to build an AI chatbot

When building a conversational assistant for a website, there are three typical ways to implement it: no-code platforms, API integration, and fully custom development.

No-code platforms like Intercom, Drift, or Zapier enable creation of conversation flows in a drag-and-drop interface and easily connect a chatbot to a knowledge base. Intercom offers Essential, Advanced, and Expert plans with seat-based pricing, plus usage-based fees for AI resolutions. The chatbot tools in the automation platform Zapier make sense for simple Q&A's. However, multi-step conversations or CRM writes that go beyond the pre-built connectors are handled by no-code platforms very poorly. A bot won't be able to update a Salesforce opportunity in the middle of a conversation, for example, or route tickets based on custom account data. And then there's the usage-based pricing that compounds as traffic increases.

API integration: In terms of implementing conversational AI for a website, API integration of services such as OpenAI's API or Anthropic's Claude API is common. Conversational AI solutions that have been integrated into a website or web application usually are supported by a lightweight back end. Such a back end can consist of a serverless function or an HTTP webhook that is addressed by the chat UI. According to Anthropic's API pricing page (2026), Claude Haiku 4.5 costs $1 per million tokens for input and $5 per million tokens for output. Thus, conversational AI solutions are typically priced in "per million tokens" and therefore can be cost-efficient, particularly for large volumes of dialog. API chatbot implementations are usually supported by a few hundred to a few thousand lines of code. The major barrier to implementing an API-based chatbot is getting started and finding a fitting solution in due time.

Fully custom build - Train your own model or fine-tune a pre-trained model to build a conversational AI chatbot for a website. With a fully custom build, the chat interface and the backend of the chatbot can be created from scratch and hosted on your own servers. While this offers the greatest degree of control over the chatbot and its behavior, it requires significant technical expertise and the necessary infrastructure to host the chatbot. While a fully custom build can be implemented in a relatively small amount of code, the technical complexity is substantial.

Automating FAQ responses

For most teams, the way to get started with chatbots is to use a no-code chatbot platform (Intercom's Fin, Drift, Zapier Interfaces, Voiceflow). These platforms let teams set up a working chatbot in an afternoon. Implementation involves wiring up a script tag, creating conversation flows with a visual builder, and connecting a knowledge base (e.g. URL crawling, or doc uploads). Marketing can set up the whole deployment without ever having to talk to engineers.

With no-code platforms, teams rent simplicity. And as needs for more complex features (beyond basic automation) emerge, the tradeoff is steep.

Key considerations include control over data (where it is stored and used) and prompts; shallow prompt configuration; and scalable and cost-effective features. Common features and workflows for support FAQ automation, lead qualification, appointment booking and more exist on no-code platforms (e.g. Intercom's Fin, Drift, Voiceflow). Support features enable teams and companies to handle the inevitable increase in volume of conversations. But limitations of no-code platforms include being restricted to the platform's workflow for complex processes. This means even using intent templates and pre-built connectors to CRMs, marketing automation systems, etc. limits true customization.

The biggest limitation of using a no-code platform for chatbots is the limitation of customization. Most platforms provide intent templates and allow connection to CRMs such as Salesforce or HubSpot, or any other number of CRM platforms. Fine-tuning of the model and the prompts used by the chatbot is possible, but there is only so much that can be done on a surface level to customize prompts. Also, the workflow of a chatbot created with a no-code platform is usually limited in terms of complexity and depth.

A number of the platforms have unique strengths that fit into particular use cases. Intercom's Fin is a support automation platform and therefore is best used for automatically answering frequently asked questions as part of creating a ticket. In contrast, Drift is a sales qualification platform optimized for qualifying hot leads and routing them to the correct sales representative as quickly as possible. All of Drift's features are specifically designed to ensure as many qualified leads are speaking with a representative as quickly as possible, and being routed to the correct rep in real time. Zapier Interfaces is a platform for connecting a chatbot to the rest of a tech stack and is best suited to teams who are primarily already building out a tech stack on top of a core of Zapier integrations. Voiceflow, on the other hand, has the most design flexibility of the platforms listed here, while still being built on top of a model stack and being hosted on their infrastructure. The design of conversational flows can be much more finely tuned than on the other platforms.

Teams can set up a chatbot in hours, not weeks. But as needs outgrow the simple setup and configuration of the visual builder, the more advanced features of the platform will often be charged as enterprise features at higher cost.

Automating custom integrations

Five core components exist to build a conversational AI assistant for a website - full control without another SaaS subscription:

The five core components of a conversational interface for a website to provide help and support to customers can be tricky for technical people to wire together to build a full conversational interface without having to pay for another SaaS product. Building against the OpenAI or Anthropic API directly is generally preferred because it allows addition of a thin custom layer on top to deal with a specific chat interface and context, and then deploy to the same infrastructure used for the website.

OpenAI API keys and Cost monitoring. To connect to the OpenAI API, teams will first need to sign up for an OpenAI account and retrieve the API key. The API key should be stored as environment variables and never hardcoded. Additionally, cost monitoring should also be implemented in order to alert for any increased usage in order to prevent any unexpected spend.

Create a chat interface. Most chat interfaces are created as open-source React components, or companies use a SaaS product and SDK to integrate the AI. Vercel, for example, provides a free AI SDK for their customers to create conversational interfaces without requiring additional SaaS subscriptions.

Design your system prompt - this is more important than you think. The system prompt is the core of a chatbot. It defines the chatbot's character. It describes what the chatbot knows and what it does not know. And it describes how the chatbot will behave under pressure. In other words, it will describe how the chatbot will behave in situations where it is not sure about the answer to a user's question. For example: "You are a support assistant for Acme Corp. You only answer questions regarding pricing and features of products of Acme Corp. If a user asks a question regarding competitors of Acme then you will answer by referring the user to the page on Acme's website that differentiates Acme from its competitors." This prompt will run through every single conversation that the chatbot has with a user.

Connect your prompt to a knowledge base. A prompt will only work well for a conversational AI assistant if it is connected to a knowledge base that contains all of the information that the assistant will need to answer any questions that may be asked by a user. If there is a very small knowledge base of say a few FAQ pages and some product documentation then these can be included in the prompt. However, for most use cases a vector database such as Pinecone will be used for knowledge retrieval-augmented generation (RAG) of responses. In this case, the conversational AI assistant will query the knowledge base in real time for the most relevant information and then include the relevant parts of the documentation in the response context.

Deploy your Assistant and Monitor Performance. Track the API's usage of tokens and the performance of the chat assistant. Log the interactions that have taken place. Track the quality of the responses that have been given and also track any failures that have occurred with the chat assistant. Deploy the chat assistant as an Express or Next.js API on AWS or Vercel.

Choosing between OpenAI and Anthropic

Both GPT-4 and Claude are capable models for powering chatbots. GPT-4 handles creative tasks well and draws on a broad general knowledge base. Claude excels at following detailed multi-step instructions and generating structured outputs such as JSON, and can handle longer context windows when passing in documentation or other lengthy text.

Automating real-time data queries

A chatbot that's helpful to customers vs a basic FAQ reader requires up-to-date information for customers. Questions like 'Where is my order?' or 'How many API calls do I have left this month?' require the chatbot to query your database to arrive at the correct answer.

Following the model's determination that external information is required to service a user query, function calling in the LLM (supported by OpenAI and Anthropic) can then follow the pattern outlined above to query servers that execute database calls or API requests. The returned results from these external calls are then used by the LLM to generate human natural language responses to queries received from users of the conversational interface.

As the assistant processes questions, it will need to pull information from various data sources to answer the questions posed.

For the integrations, typical data sources that would be wired up to a conversational assistant are: Salesforce (e.g. account data, case data, contact data), HubSpot (e.g. deal data with stage, contact data with properties, pipeline data), Stripe (e.g. subscription status, invoice data, payment methods), internal databases (e.g. order data, inventory data, user settings). The hardest part is to build out a secure API layer around these integrations to validate and execute the requests correctly.

Security is where most teams underestimate the work.

Always query data within the context of an authenticated user's session. The bot should only be able to see the data that the logged-in user is permitted to see. NEVER allow the LLM to generate raw SQL. The LLM will inevitably hallucinate table and field names or generate creative joins that will expose other users' data. Always validate ALL function calls before executing them. Reject any prompt that indicates the LLM believes it can generate results for a request such as "all orders" as opposed to "orders for user X". Prompt engineering can be used here to inform the LLM as to the functions that exist and what parameters those functions take. As with all untrusted input - treat prompt engineering as such.

Chatbot cost structure and pricing

Many teams get caught off guard by the true cost of a chatbot. At first glance, an entry-level chatbot plan may seem affordable, but the reality is that usage-based pricing can quickly escalate as conversation volume grows.

Breaking down the conversational AI pricing to the numbers:

No-code platforms: variable monthly costs that scale with usage

The no-code platforms (like Intercom, Zendesk, etc.) fall into this category. Intercom offers Essential, Advanced, and Expert plans with seat-based pricing, plus usage-based charges for AI resolutions. If a support team is handling substantial conversation volume - and that's realistic for many support teams - usage-based charges can add up on top of the base platform license.

The biggest shock to pricing for most teams is when they hit higher conversation volumes. This is where most of the platforms will trigger to the next tier or scalability plan that often are charged based on conversations, resolutions per month, etc. The irony here is that a large part of what a good chatbot does is handle the more trivial support questions before staff have even had to respond. Thus, the very best performing chatbots may need to be scaled back to avoid escalating costs.

API-based custom build: upfront development, then token-based usage

Here's a breakdown of the costs for conversational AI approaches that use an API to wire up services like Claude or GPT-4 to an application, including one-time development costs:

• Simple chat interface with a static knowledge base: Building a simple chat interface with a static knowledge base (i.e. no machine learning or dynamic updates to the knowledge base) that can interact with customers requires developer time. This would include building a simple front-end for the chat interface (e.g. a simple React interface), connecting the front-end to the Anthropic or OpenAI API (e.g. Claude, GPT-4) for the conversational AI, and then deploying the application to a hosting platform (e.g. Vercel, AWS).

• CRM integration with function calling: This type of functionality would enable the conversational AI to pull data from CRM systems such as Salesforce. The conversational AI would then use that information to create new tickets in CRM systems such as HubSpot. The development of this type of AI functionality would require a developer with proper error handling experience and familiarity with CRM systems that support function calling. The developer would need to create validation logic for the conversational AI as well as IAM controls to ensure that the chatbot does not cause any harm to the CRM system.

In addition to the costs of the development of an AI interface, ongoing API subscription costs must be considered. According to Anthropic's API pricing page (2026), Claude Haiku 4.5 costs $1 per million tokens for input and $5 per million tokens for output. These costs are based on the number of conversations in length and the AI models being used.

When to build vs. buy

Use no-code first for FAQs, lead collection, and simple triage to get immediate value. Businesses with modest conversation volumes can have this up and running within a week without needing developers or complex design and maintenance.

Use an API if there is a developer or technical founder on the team, a need to control costs as the solution scales, or requirements to integrate with other systems (e.g. pull in CRM data, trigger off of other backend actions, put information into internal databases). No-code platforms are not designed to handle complex workflows and will fall apart under the strain of trying to work around limitations. API-based implementations offer visibility into usage and resulting costs, as opposed to getting hit with additional usage fees while stuck in a subscription tier for a no-code platform.

Build from scratch when specific requirements are not met by the platforms available. This would mean hosting on-premise, needing to fine-tune a model on internal data, or processing sensitive information that should never leave your infrastructure. The bot is the product; therefore build the product. In all other cases, businesses would end up solving a problem that does not exist.

For a low-traffic site (under 1,000 visitors per month), just use a contact form. The effort of managing the bot would be greater than the effort of handling the emails in such cases.

Do not use a bot for complex questions for which AI will not be able to give correct answers. Typically, this would be for legal advice, for medical advice, and for any situation in which giving wrong information could cause serious problems. Do not use a bot if there are not the resources to run it. A bot that gives out bad information is worse than no bot at all. It will erode trust in a company faster than a slow response time would.

The biggest mistake any company makes with conversational AI is to build a chatbot because it looks like new, shiny, innovative technology. Most companies end up overcomplicating the very simple use case they're trying to solve with conversational AI, and in the end, what they build doesn't solve real problems for real users. Start with the problem and work back to the technology, not the other way around.

Frequently Asked Questions

Can I build my own chatbot for free?

Businesses can build a chatbot for free, provided they know how to code in Python and use a number of APIs. For example, OpenAI's API pricing (2026) is usage-based, and teams can host their chatbot on the free tier of a cloud provider. However, they will need to create all of the prompts for the chatbot, handle all of the errors for the chatbot, and keep the chatbot up and running.

How much does it cost to build conversational AI for a website?

The cost for building conversational AI for a website can vary depending on the development route taken. No-code platforms offer various pricing tiers - Intercom offers Essential, Advanced, and Expert plans with seat-based pricing and usage-based AI fees, while Zapier offers Free, Professional, Team, and Enterprise tiers. Custom development including integration of an LLM (Large Language Model) and a knowledge base requires developer time and project scoping. After this there will be additional monthly costs for the API usage depending on the amount of traffic to the website.

What's the difference between a rule-based chatbot and an AI chatbot?

Rule-based chatbots are based on a predetermined decision tree, previously set up by a human administrator. Each rule within the decision tree states something along the lines of: "If the user asks this, then respond with that." Often, the response is a pre-written message that was written out beforehand to handle a specific question or scenario. The major issue with these types of chatbots, however, is that they don't perform very well when a user asks something that the administrator hadn't previously anticipated.

AI-powered chatbots, on the other hand, use a large language model (LLM) to process and generate human-like responses to a user's question or prompt. These types of chatbots are not restricted by how they are programmed in advance, can deal with a large volume of conversations and unexpected questions that had not been anticipated when programming a rule-based system. However, the cost of running an AI-powered chatbot is usually higher than a rule-based system as it is dependent on the cost of the API to power the chatbot as well as the time it takes to get the AI up to speed on a large amount of information or questions and be able to accurately answer them.

Do I need a developer to build an AI chatbot?

For simple, FAQ-style chatbots (such as those used to direct customers to FAQs for common issues), no coding skills are required. For these types of chatbots, no-code platforms such as Voiceflow or Chatbase can be used. However, for more complex conversational solutions, such as those that require CRM integration, custom workflows etc., it is highly recommended to bring in a developer to pattern the solution.

How do I connect my AI chatbot to my CRM?

To integrate an AI chatbot with a CRM, first set up an API connection between the chatbot platform and the CRM. Then set up the chatbot to collect all of the lead data during a user's conversation with the chatbot. Finally, set up the chatbot to automatically add this new lead information to the CRM via a webhook or API call. The hardest part is figuring out what data the chatbot collects in the conversations needs to be mapped to fields in the CRM in order to avoid adding duplicate records or missing out on key information.

What should I include in my chatbot's knowledge base?

Support questions and answers, product documentation, pricing information, FAQs (which are probably already in written form from previous phone or in-person conversations with sales teams) etc. Keep the KB information concise - typically a few paragraphs at most per entry. Make sure the LLM can read the information on the KB pages as if they were reading written material. Update KB information on a quarterly basis in order to incorporate the various questions from website conversations that the current KB doesn't cover.

Gable Innovation helps growing businesses find, implement and integrate AI technology such as chatbots. That could be helping determine if conversational AI makes sense for a website, and then building and integrating a custom chatbot for the business. Gable offers free 30 minute discovery calls to discuss further.

Book a call at gableinnovation.com