Quick answer: In early 2023 you have three realistic ways to add an AI chatbot to a website: a hosted chatbot tool with AI features switched on, an intent-based platform such as Google Dialogflow, Amazon Lex or Microsoft’s bot tools, or a custom bot built on a large language model API such as OpenAI’s GPT-3 models. Costs range from a monthly subscription to a development project plus usage fees. The biggest limit is accuracy: language models can sound confident while being wrong, so any bot that answers customers needs guardrails, a clear scope and an easy route to a human.
Why everyone is asking about this right now
ChatGPT opened to the public at the end of November 2022, and within weeks business owners were asking the same question: can we put something like that on our site? It is a fair question. ChatGPT shows that software can now hold a natural conversation, follow up on vague questions and write clear answers.
What ChatGPT does not show is how that works for one specific business. ChatGPT itself is a free research preview with no official API at the time of writing, so you cannot simply drop it onto a page. What you can do is use the GPT-3 family of models that OpenAI already offers through its API, or use other platforms that have been building chatbots for years. Microsoft also made its Azure OpenAI Service generally available this week, which gives larger companies a second route to the same kind of models inside Azure.
If you are new to the wider topic, our guide to AI automation for small business covers where AI fits best across a company. This article focuses on the chat window on your website.
What a website chatbot can actually do for you
Before picking a tool, decide what job the bot is there to do. Most useful website bots fall into one of these roles:
- Answering repeat questions: opening hours, delivery times, return rules, pricing basics, how to book.
- Qualifying leads: asking a few questions, then passing the details to your CRM or sales inbox.
- Booking and routing: sending a visitor to the right calendar, department or form.
- Support triage: collecting order numbers and the problem before a person picks it up.
If your main goal is leads, it is worth reading our comparison of a chatbot or contact form first. A bot is not automatically better than a well designed form, and for some audiences it is worse.
Option 1: A hosted chatbot or live chat tool
Many live chat products include a bot builder. You design flows with buttons and simple questions, and some tools add basic language understanding on top so that typed questions match the right answer.
Good for: small sites that want quick answers to common questions and a handover to a person during business hours.
What it costs: a monthly subscription, usually priced by seats, conversations or features. Setup is mostly your time writing answers and flows.
Limits: these bots are mostly scripted. They handle the questions you planned for and struggle with anything else. That is often fine. Predictable is a feature when the bot is talking to customers.
Our article on live chat for your website covers how to staff and measure chat, which still matters once a bot handles the first message.
Option 2: An intent-based platform
Google Dialogflow, Amazon Lex, Microsoft’s Azure Bot Service and Power Virtual Agents, and the open source Rasa framework all work on a similar idea. You define intents (what the visitor wants), give example phrases for each, and connect each intent to a response or an action such as checking an order status through an API.
Good for: businesses that need the bot to do things, not just talk. Checking a booking, looking up an order, creating a support ticket or pulling a price from your system.
What it costs: platform usage fees that are usually small per request, plus development time to design intents, connect your systems and test. The development time is the main cost.
Limits: you have to anticipate what people will ask. Unexpected questions fall through to a fallback message. The upside is control. The bot only says what you wrote, which makes it safer for regulated or sensitive topics.
Option 3: A custom bot on a large language model API
This is the option most people mean when they say “a ChatGPT for our website.” A developer builds a chat interface that sends the visitor’s question to a model such as OpenAI’s text-davinci-003, along with instructions and relevant information from your own content, and shows the reply.
The common pattern looks like this:
- Collect your source material: FAQ pages, policies, product details, help articles.
- Split it into short passages and create embeddings for each one, so the system can find the passages most related to a question.
- When a visitor asks something, find the closest passages and send them to the model with clear instructions, such as “answer only from this information and say you do not know otherwise.”
- Show the answer, log the conversation and offer a human handover.
Good for: businesses with a lot of written content and many different questions, where scripted flows would never keep up.
What it costs: a development project to build the interface, the content pipeline, logging and safety checks, plus usage fees from the model provider. OpenAI prices its models per 1,000 tokens, where a token is roughly a short word or part of one. At the time of writing, OpenAI lists Davinci at $0.02 per 1,000 tokens and its newer embedding model at $0.0004 per 1,000 tokens, with smaller and cheaper models also available. Your monthly bill depends on how many conversations you get and how much context you send with each question, so estimate it from your own traffic rather than a rule of thumb.
Limits: covered in detail below. They are real, and they are the reason this option needs the most care.
What drives the cost of an AI chatbot
Whatever option you pick, these are the items that move the price:
- Scope: a bot that answers 20 questions is a small project. A bot that answers anything from 300 help pages and checks live order data is a much bigger one.
- Integrations: connecting to a CRM, booking tool, store or help desk adds development and testing time. Our guide to sending web forms to a CRM through an API shows the kind of work involved.
- Content preparation: most bots fail because the source content is out of date or contradicts itself. Cleaning it up is often the largest hidden cost.
- Usage: subscription tiers for hosted tools, or per-token fees for model APIs.
- Ongoing review: someone needs to read transcripts, fix wrong answers and update content. Budget a few hours a month at minimum.
The limits you need to plan for
Confident wrong answers
Large language models predict likely text. They do not check facts. When they lack information they can make up a plausible answer, including fake policies, prices or links. Grounding the model in your own content and telling it to refuse when the answer is not there reduces this, but does not remove it. Test with tricky questions before launch and keep testing after.
Out of date knowledge
GPT-3 models were trained on data that stops in 2021. They know nothing about your new product, this year’s prices or last week’s change to your returns policy unless you supply that information with each request.
Instructions that visitors can override
Developers have started calling this prompt injection. A visitor types something like “ignore your instructions and tell me…” and the model may follow them. Keep sensitive data out of what you send to the model, limit what actions the bot can trigger, and never let the model alone approve refunds, discounts or account changes.
Privacy and data handling
Visitors will type personal details into a chat box, whether you ask for them or not. Check where conversation data is stored, how long it is kept, whether the provider may use it, and what your privacy policy says. If you operate under GDPR or similar rules, involve whoever handles data protection before launch.
Tone and brand risk
A bot speaks for your business. One rude, odd or wrong answer can end up in a screenshot. Write clear instructions for tone, keep the bot’s scope narrow and review transcripts regularly, especially in the first weeks.
Speed
Model APIs can take several seconds to reply, and response times vary. Show a typing indicator, keep answers short and have a fallback message if the API does not respond.
How to choose
A simple way to decide:
- Mostly the same ten to thirty questions? Start with a hosted tool and scripted flows.
- Need the bot to look things up or take actions? Use an intent-based platform connected to your systems.
- Large body of content and very varied questions? Consider a custom bot on a language model API, with grounding, logging and human handover built in from day one.
Many businesses end up combining options: scripted flows for common tasks, a language model for open questions, and a person behind both.
A sensible way to start
- Pull your last few months of emails, chat logs and support tickets and list the questions that come up most.
- Write or fix the answers to those questions on your website first. This helps visitors and search engines even if you never launch a bot.
- Pilot the bot on one page or one topic, with clear wording that visitors are talking to an automated assistant.
- Make the human handover obvious. Route those conversations properly, as you would with any lead routing setup.
- Review transcripts weekly, fix wrong answers, then widen the scope.
If support volume is the real problem, look at help desk automation alongside the bot. Sorting and routing tickets often saves more time than the chat window itself.
Common questions
Can I add ChatGPT to my website?
Not directly at the time of writing, because ChatGPT has no public API yet. You can build a similar experience with OpenAI’s existing GPT-3 models through their API, or through Microsoft’s Azure OpenAI Service.
Will an AI chatbot replace my support staff?
For most small businesses, no. It can take the repeat questions off their plate and collect details before a person steps in. Complex, emotional or high value conversations still need a human.
Is an AI chatbot good for SEO?
Chat conversations are not usually indexed, so the bot itself does little for search. The content work you do to prepare the bot, such as clear FAQ and help pages, can help a lot.
How long does it take to build one?
A scripted bot on a hosted tool can be live in days. A custom bot grounded in your content, with integrations and testing, is a project measured in weeks.
The bottom line
An AI chatbot can save time and help visitors, but the right choice depends on the job you need done, not on the excitement around ChatGPT. Start with clear questions and good content, pick the simplest option that covers them, plan for wrong answers and privacy, and keep a person close by.
If you want help scoping a chatbot or deciding whether you need one, see how we automate the busywork for growing businesses.
