Adding an AI Chatbot to Your Website: Options & Costs

You can add an AI chatbot to your website in two ways: install a hosted widget for roughly $0-$150 a month and go live in a day, or build a custom bot on your own data. A custom build with a Turkey-based team typically costs $3,000-$15,000, plus $30-$300 a month to run.

Which route fits depends on three questions, not on the technology. What data will the bot answer from? How many conversations will it handle each month? Who takes over when it cannot help? If you cannot answer those three yet, the choice of tool will not save the project.

The stakes are real. Gartner surveyed 3,566 B2B and B2C customers in February and March 2026 and found that only 27% would try a chatbot again after a negative experience (IT-Online). A weak launch does not just fail quietly. It teaches your visitors to avoid the channel.

This guide takes the decision-maker's view: bot types, what hosted widgets really cost, RAG in plain language, token economics, the handoff to a human, the risks, and the compliance basics. We took every price from the official pricing pages in September 2026.

Table of Contents

• Rule-Based vs LLM: Which Type of AI Chatbot Do You Need?
• Hosted Chatbot Widgets: What You Get and What You Pay
• A Custom Bot on Your Own Data: How RAG Works
• Token Economics: What Does One Conversation Cost?
• Buy or Build? Cost and Decision Table
• Human Handoff: The Feature That Decides Success
• Risks: Wrong Answers, Prompt Injection and Liability
• Privacy and Transparency: Say It Is a Bot
• How to Add an AI Chatbot to Your Website in 6 Steps
• Frequently Asked Questions

Rule-Based vs LLM: Which Type of AI Chatbot Do You Need?

Products sold under the word "chatbot" rest on three different technologies. Knowing the difference tells you what you are paying for when you read a quote.

A rule-based bot is a decision tree. The visitor taps buttons and the bot shows the answer you wrote in advance. It does not understand free text; type "where's my parcel??" and it usually replies "Sorry, I didn't get that." In return, you know exactly what it will say, each conversation costs close to nothing, and it carries the least legal risk.

An LLM-based AI chatbot runs on a large language model. It understands free text, typos and messy phrasing, and it writes a fresh answer every time. That is both its strength and its risk. Left alone, a model will invent a confident answer about a policy it has never seen. Serious deployments therefore tie the model to the company's own documents; the RAG section below explains how.

A hybrid bot combines the two. Tasks with fixed steps, such as starting a return, booking a slot or collecting a quote request, run as rule-based flows. Open questions go to the LLM. This is the setup we recommend for most sites.

Question variety should drive the choice, not traffic volume. If visitors keep asking the same five things, a rule-based flow will do. If the pool of questions is wide and the answers already live in your documents, the LLM side earns its place.

Hosted Chatbot Widgets: What You Get and What You Pay

A hosted widget is a SaaS product that runs from one script tag on your site. You point it at your URL or upload your PDFs, the platform crawls the content, and within hours you have an AI chatbot that answers questions. Most vendors also ship plugins for WordPress, Shopify and similar platforms.

The detail that matters most is the pricing model. Vendors bill in four different units: per agent seat, per conversation, per message credit, and per resolved request. The same traffic produces very different invoices depending on the unit.

The strength of a hosted widget is speed. You launch the same day, the vendor handles maintenance, and the live chat inbox and reports come in the box. The weaknesses show up in three places:

• Access to live data. Questions about order status, stock or account balance need the bot to reach your systems. Hosted products offer this on higher tiers or in a limited form.
• Invoice scale. Per-resolution and per-conversation fees grow in a straight line with traffic. At $0.99 each, 1,000 resolved requests a month means $990.
• Data control. Transcripts sit on the vendor's servers. Read the contract for the retention period, the deletion process and the hosting country.

There is also a performance cost. Third-party chat scripts add JavaScript to every page. Load the script when the visitor taps the bubble or when the page goes idle, not during the initial render. You can measure the effect with our site analysis tool, and our guide to speeding up a website for Core Web Vitals covers the technique.

A Custom Bot on Your Own Data: How RAG Works

A language model does not know your company. It has never read your shipping policy, your price list or your warranty terms, so it guesses. The approach that fixes this is RAG, short for retrieval-augmented generation. The term comes from a 2020 paper by Lewis and colleagues, and it is now the standard architecture for bots that speak from private data.

The idea is simple. Before the model writes anything, the system searches your documents and lets the model answer only from what it found. Think of an open-book exam. The student does not answer from memory; the student answers from the page on the desk.

The process runs in four steps:

1. Preparation. The system splits your FAQ pages, product descriptions, contracts and manuals into small chunks. It converts each chunk into a numeric vector that represents its meaning and stores it in a vector database.
2. Retrieval. When a visitor asks a question, the system finds the few chunks closest in meaning. It matches meaning, not keywords, so "when do I get my money back" finds the chunk titled "Refund timeline."
3. Generation. The system sends those chunks to the model together with the question. The instruction is strict: answer only from this text, and say you do not know if the text does not cover it.
4. Citation. The bot links the source page under its answer. The visitor can verify it, and you can trace a wrong answer back to the document that caused it.

A common question: would it work better to train the model on our data? For most websites, no. Fine-tuning costs more, needs a new training run every time your price list changes, and does not stop the model from inventing things. With RAG, the answer changes the moment you update the document.

In a custom AI chatbot project, most of the effort goes into the data, not the model. Two return policies that contradict each other, a price PDF from 2023, a shipping table that exists only as an image: each one produces wrong answers. Picture a 40-person furniture manufacturer selling through dealers. Most incoming questions concern delivery times, fabric options and dealership terms. The first two weeks of that project go into merging scattered documents on those three topics into one current source, not into code.

The second critical piece is the evaluation set: a list of 50-100 questions with known correct answers. You run the bot against that list after every change. Without it, you judge quality by feel. The real advantage of a custom bot, though, is live data: order lookup, stock, booking calendars or CRM records. Our guide to API integration for business explains how those connections work.

Token Economics: What Does One Conversation Cost?

LLM providers bill by the token, the unit a model uses to read and write text. According to Anthropic's documentation, one token equals roughly 4 characters or 0.75 words in English. Other languages, especially agglutinative ones such as Turkish or Finnish, use more tokens for the same sentence. Providers price input and output separately, and output costs several times more everywhere.

Sources: official pricing pages of Anthropic, OpenAI and the Gemini API, September 2026.

Let us make that concrete. Anthropic's pricing page includes a customer support example: about 3,700 tokens per conversation, and 10,000 conversations on Claude Haiku 4.5 for roughly $37 in total. As an illustrative calculation, a site with 2,000 conversations a month would pay about $7.40 under the same assumptions. Even if RAG context and a token-heavy language triple the bill, it stays near $22.

Now assume a platform that charges $0.99 per resolution closes half of those 2,000 conversations: the invoice reads $990. The comparison needs a fair caveat. The platform fee covers the inbox, reporting, guardrails and maintenance, none of which a raw API gives you. Still, the gap explains why a custom AI chatbot enters the conversation as volume grows.

Three things quietly inflate a token bill:

• Conversation history. Every new message sends the earlier messages to the model again. The last turn of a ten-turn chat uses several times the input of the first.
• Context bloat. RAG does not need to attach ten pages to every question. Three relevant chunks usually do the job.
• Model choice. Answering FAQs does not call for the most expensive model. A small model handles most support questions, and routing only hard questions to a larger one keeps the bill balanced.

The brakes are just as clear. On Anthropic's pricing page, tokens read from the prompt cache cost one tenth of the standard input price. System instructions and fixed documents repeat on every request, so support bots benefit directly.

One last warning: prices move. Google's pricing page states that the current $0.75 / $3.75 rate for Gemini 3.8 Flash holds through December 31, 2026 and rises to $1.50 / $7.50 on January 1, 2027. Gartner predicts that by 2030 the cost per resolution for generative AI in customer service will exceed $3, higher than many offshore human agents (CX Today). Build on a layer that lets you swap models, not on a single model name. For a deeper look at token maths, see our guide on how to integrate AI into a mobile app.

Buy or Build? Cost and Decision Table

The build ranges below are typical bands from Master Web projects delivered by our Turkey-based team. They are not a market average, and a firm quote follows a scope analysis.

An honest starting point: for a site with a few hundred conversations a month, a hosted widget almost always makes more sense. Against a $40 subscription, a custom build would take years to pay for itself through savings alone. What justifies a custom bot is not the saving. It is the work a hosted product cannot do.

A custom AI chatbot makes sense when:

• The bot must read live data such as orders, stock, bookings or account balances.
• Conversation volume has pushed per-resolution or per-credit fees into hundreds of dollars a month.
• Transcripts must stay on your own servers, which healthcare, finance and legal work often require.
• The bot needs to create records in your custom CRM, draft quotes or talk to a dealer portal.

The practical advice is to move in stages. Start with a hosted widget and collect two or three months of data: what visitors really ask, which questions the bot fails, how many chats go to a human. That data lets you scope the custom build from measurement, not from guesswork.

Location changes the build price more than the technology does. Our guide to outsourcing software development to Turkey puts mid-to-senior engineers in Turkey at $25-$50 an hour, against $95-$180 for North American onshore teams. For the rest of your site budget, the website cost calculator and our software project budget planning guide help you pin down each line.

Human Handoff: The Feature That Decides Success

What defines an AI chatbot is not the questions it answers but what it does when it cannot. In the same Gartner survey, 87% of customers said companies that use generative AI in customer service must keep a route to a human agent (CX Dive).

The gap between intent and behaviour tells the same story. In that survey, 49% of customers said they would have used a chatbot if the company had offered one, yet only 7% used one in their most recent service interaction. One dead end pushes a visitor away from the channel for good.

A healthy handoff flow has four parts:

1. Triggers. The handoff starts when the visitor types "agent" or "human", when the bot fails the same question twice, or when the message mentions a complaint, a cancellation or legal action. Do not bury the human option at the bottom of a menu.
2. Context transfer. The agent sees a summary of the chat, the page the visitor was on, and the answers the bot gave. Making the visitor repeat everything wipes out the value of the handoff.
3. Out-of-hours behaviour. If nobody is online at 11 p.m., the bot says so, takes an email address or phone number, and logs the request. "Connecting you now" followed by silence is the worst option.
4. Feedback loop. Every chat that went to a human points to a gap in the knowledge base. Read those chats weekly and update the documents; that is how the handoff rate falls.

Another Gartner prediction points the same way: half of the companies that cut customer service staff because of AI will rehire for similar roles by 2027. Put the bot in front of your people, not in place of them.

Risks: Wrong Answers, Prompt Injection and Liability

What your AI chatbot says, you said. The clearest example comes from Canada. In February 2024, British Columbia's Civil Resolution Tribunal held Air Canada liable in Moffatt v. Air Canada (2024 BCCRT 149) for wrong refund information its website chatbot gave a customer. The airline argued the chatbot was a separate entity; the tribunal rejected that. The core of the ruling fits in one sentence: a company answers for all the information on its website, whether it comes from a static page or a chatbot (case summary).

The second risk is technical. Prompt injection ranks first on OWASP's 2025 list of risks for LLM applications: user prompts that alter the model's behaviour or output in unintended ways. In the direct form, a visitor types "ignore your previous instructions." In the indirect form, the malicious instruction hides inside a web page or file the bot reads.

The measures that work in practice:

• Limit authority. The bot cannot grant discounts, approve refunds or commit to prices. Those actions go to a rule-based flow or to a person.
• Tie answers to sources. The bot answers only from the knowledge base and links the source. With no source, it says "I don't have that information, let me pass you to our team."
• Narrow tool permissions. If the bot looks up orders, it sees only the orders of the signed-in user. OWASP calls this least privilege.
• Filter input and output. Set separate filters for personal data, abuse, competitor comparisons and legal advice.
• Keep logs. You should see later which question drew which answer from which document. In a dispute, that record is your only evidence.

Our website security guide covers the wider picture; a bot is one more attack surface on that list.

Privacy and Transparency: Say It Is a Bot

This section is not legal advice; check it with your own counsel. Visitors type names, phone numbers, order numbers and sometimes health details into a chat box. Every AI chatbot therefore processes personal data, and GDPR or your local equivalent applies.

In practice, that means four habits. The first message of the chat window carries a short privacy notice with a link to the full text. A masking layer on your server strips ID numbers, card details and phone numbers before the message reaches the model provider. You set a retention period for transcripts and delete them when it expires. You also check where the model provider processes data, because a transfer outside your region triggers its own rules. Our website cookie compliance guide covers the cookie and form side of the same picture.

If you serve users in the EU, one more rule applies. Article 50 of the EU AI Act applies from 2 August 2026. Providers must make sure people know they are interacting with an AI system, unless that is already obvious. Fines reach up to €15 million or 3% of worldwide annual turnover (European Commission).

Turkey, where our team works, moved early on this topic. The Turkish data protection authority published an information note on chatbots on 8 November 2024. It asks developers to run a risk assessment before processing personal data, state the legal basis, inform users at the point of collection, and apply privacy by design and by default.

Legal duty aside, naming your bot "Emma" and letting it pose as a person damages trust the moment a visitor notices. Opening with "I'm an AI assistant" is the cheapest trust investment you can make.

How to Add an AI Chatbot to Your Website in 6 Steps

Step 1: Write down the goal and the metric. "We want a bot on the site" is not a goal. "Close half of delivery and returns questions without a human" is one you can measure. Pick the metric up front: resolution rate, handoff rate or satisfaction score.

Step 2: Build a question inventory. Go through the last three months of emails, chat logs and call notes. List the 30 most frequent questions. That list becomes both the skeleton of the knowledge base and the first draft of your evaluation set.

Step 3: Clean the knowledge base. Merge contradictory, outdated and image-only information into a single source. Give every answer an owner and a review date. Skip this step and the bot will mirror the mess on your site straight back to visitors.

Step 4: Choose the route and run a pilot. Use the decision table to pick a hosted widget or a custom bot. Launch on one group of pages first, for example only the support section. Set up the handoff flow and the out-of-hours behaviour during the pilot, not afterwards.

Step 5: Test against the evaluation set. Before launch, run your 50-100 questions. Review every wrong answer: does the fault sit in the document, the retrieval or the instruction? Add prompt injection attempts and out-of-scope questions to the set.

Step 6: Measure and maintain weekly. In the first month, read the handed-off chats every week, update the knowledge base and watch token spend. Set a daily spending cap and an alert. An AI chatbot does not behave like a plugin you install once. It behaves like a content channel you keep current.

These six steps show that the work is more about content and process than about code. The code takes weeks. The right questions, clean documents and a working handoff decide the quality of the result.

Frequently Asked Questions

How much does it cost to add an AI chatbot to a website?

Hosted widgets start in the $0-$150 a month range, and the invoice grows with conversations or resolutions. A custom RAG bot on your own data costs $3,000-$6,000 with a Turkey-based team, or $6,000-$15,000 with order and CRM integrations. Running costs for a custom build stay between $30 and $300 a month.

How long does it take to add an AI chatbot to a website?

A hosted widget goes live in 1-3 days, and most of that time goes into preparing the knowledge base. A custom RAG bot takes 3-5 weeks. A version with live data connections and a handoff console takes 6-10 weeks.

What is RAG and why does an AI chatbot need it?

RAG means the model searches your documents before it writes and answers only from the text it found. A language model does not know your prices, policies or products, so without RAG it guesses. When you update a document, the answer updates too, with no retraining.

Can I add an AI chatbot to my website for free?

Yes, within tight limits. As of September 2026, Chatbase's free plan includes 50 message credits a month and Tidio's free plan includes 50 one-off AI conversations. Those quotas suit a trial; real traffic will push you to a paid plan.

Can I add an AI chatbot to a WordPress or Shopify site?

Yes. Most hosted widgets offer a plugin or a single script snippet that works on WordPress, Shopify and custom sites alike. A custom bot connects the same way: a small front-end script talks to your own back-end service.

Will an AI chatbot replace live chat agents?

No. It works in front of them. In Gartner's 2026 survey, 87% of customers said companies using generative AI in customer service must keep access to a human agent. A well-designed AI chatbot closes repetitive questions and passes complex or emotional cases to a person with full context.

Who is liable if the chatbot gives wrong information?

The website owner. In Moffatt v. Air Canada (2024), a Canadian tribunal ruled that a company answers for all information on its website, including what its chatbot says. Remove the bot's authority to promise prices, discounts or refunds, and tie every answer to a source document.

Before you add an AI chatbot, the question to settle is not "which tool" but "which questions and which data." If traffic is modest and the answers already exist in your documents, start with a hosted widget and collect data. If the bot must read live data, if the invoice has grown with volume, or if transcripts must stay on your servers, the time for a custom build has come.

At Master Web we run the whole process, from knowledge base preparation and RAG architecture to order and CRM integrations and the handoff console. You can see what we do on our web software development page. To review your questions and your data together and find the route that fits, get in touch with us.