AI chatbot for your website: should you, and which kind?

An AI chatbot is worth it when customers keep asking the same handful of questions (price, opening hours, policies, order status) and you already have accurate written answers for the bot to work from. For most small and mid-sized businesses the sensible choice is an off-the-shelf AI chat widget that answers from your own content, with a "talk to a person" button that is always visible. Build a custom assistant only when the bot needs to look things up in your own systems (orders, availability, stock) or when customer data needs tight control.

And if your customers mostly message you on WhatsApp or pick up the phone, a website chatbot may not be the first place to spend.

Three kinds of chatbot, in plain English

Vendors call everything an "AI chatbot", but there are really three very different things:

A scripted chatbot (rule-based): the bot follows buttons and answers you wrote in advance. The customer taps "See prices" and the price list appears; ask anything off-script and it is lost. Cheap and safe, but rigid.
An off-the-shelf AI widget: chat software you rent by the month (Tidio, Intercom and the like) with an AI inside that understands free-form questions. You feed it your FAQ, policies and price list, and it answers in natural language.
A custom AI assistant: built for your business, wired into your own systems (orders, bookings, inventory), following your rules, with your data stored where you decide.

The phrase that matters most is "answers from your own data". It means that before replying, the bot looks up your documents to find the relevant passages, and is only allowed to write its answer from those passages. Engineers call this RAG (Retrieval-Augmented Generation, roughly "look it up first, then write").

An analogy: an AI without lookup is a brilliant new hire who talks well but hasn't read the company handbook. They answer everything fluently, including when they're guessing. An AI with lookup is that same hire made to open the handbook before answering, under instructions: "if it isn't in the handbook, say you don't know and pass it to the manager."

What each kind costs (market prices)

Publicly listed prices of popular tools at the time of writing:

Kind	Market price example	Where the real cost sits
Scripted	Tidio has a free plan (50 conversations a month); its Flows automation starts at about $24/month (Tidio)	Time to write the scripts
Off-the-shelf AI widget	Tidio's Lyro AI starts at $32.50/month for 50 AI conversations; Intercom's Fin charges $0.99 per resolved outcome, plus seats from $29 per person per month (Intercom)	Fees grow with chat volume; effort to write the content the bot reads
Custom assistant	AI usage is cheap: for example Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens (Anthropic)	Development, system integration, maintenance

A token is the unit AI is billed in: each small chunk of text the model reads or writes.

To make that concrete, here's my rough estimate from the price list above: if each chat uses about 20,000 input tokens (including the documents the bot looks up) and 2,000 output tokens, then 1,000 chats on Claude Haiku 5.5 cost about $3 in AI usage; on the more capable Claude Sonnet 5.5 ($2 / $10 per million tokens) it's about $60. That's an illustration, not a quote.

The lesson: with a custom assistant, the AI bill is almost the smallest cost. What you mainly pay for is building it, connecting it to your data, and keeping its answers right. A rented widget is the reverse: fast to set up, but the bill climbs with your traffic.

Why it matters to your money: the upside and the real risks

The upside is clear: customers get an answer at 11 p.m., your staff stop typing "how much is it?" for the hundredth time, and questions answered well move customers on to booking, buying or leaving their details.

But there are three risks, and they cost real money:

Customers hate being blocked from a human. A Gartner survey of 5,728 customers found that 64% would prefer companies didn't use AI for customer service, and 53% would consider switching to a competitor if they found out a company was going to. Their top worry was that it would get harder to reach a person (Gartner). Gartner also predicts that by 2027, half of the organizations that planned to cut their customer service workforce significantly because of AI will abandon those plans (Gartner).
The bot gets it wrong, and you're liable. In 2024 Air Canada's website chatbot misstated the airline's refund policy to a customer. The airline argued the chatbot was "a separate entity"; the tribunal rejected that, treated the chatbot as part of the website, and ordered the airline to pay (American Bar Association). What your bot says counts as what your business says.
Customer data. People type phone numbers, addresses and sometimes health details into chat boxes. Most AI services process that text on servers in another country, which can bring data-protection rules into play (GDPR if you serve customers in the EU, for example). Ask where the data goes, and check with your own legal adviser.

A good chatbot isn't one that can answer everything. It's one that knows when to say "let me connect you with a member of the team", and does it fast.

Where do your customers actually chat?

Before choosing a bot, ask: do my customers chat on my website at all, or do they message me somewhere else?

Many customers would rather use the messaging app already on their phone (WhatsApp, Messenger, or whatever is standard in their market) than a box on your site. What that means for you:

Treat the website chat as one channel, not the only one. Plenty of visitors read the bot's answer and then switch to WhatsApp, email or the phone. Your bot should offer those hand-offs.
The most valuable part, the set of accurate documents the bot looks up, can be reused for an assistant on WhatsApp later. Build the knowledge base once, use it on several channels.

How to choose: a checklist and the questions to ask

Checklist before you buy

Can you list the 10–20 questions customers ask most? Without that list, you're not ready for a bot.
Do you have accurate, up-to-date documents (prices, returns policy, hours) for the bot to look up?
Is there an always-visible "talk to a person" button, and is a person actually there to pick it up?
Is the bot instructed to say "I don't know" when the answer isn't in your documents?
Who reviews the chat history every week to fix wrong answers?
Which channel do customers use most: website, WhatsApp, email or phone?

Questions to ask your web team

"Which of my documents does the bot answer from? When I change a price, how long before the bot knows?"
"What does the bot do when it isn't sure? Show me a question that isn't covered."
"How does the hand-off to a person work? Does my staff member see the earlier conversation so the customer doesn't repeat themselves?"
"Where is what customers type stored, in which country, and for how long? Is it used to train AI models?"
"Is pricing per month, per conversation or per resolution? If chat volume grows fivefold, what does the bill become?"
"Is there a report showing what customers ask most and where the bot gets it wrong?"

A serious team will answer each one clearly. If the answer is "the AI figures it all out, don't worry", that's your cue to stop.

Next step

In short: an AI chatbot is worth having when it answers from your real data, admits when it doesn't know, and always lets customers reach a person. Start small: one set of common questions, an off-the-shelf widget or a lean assistant, measure what customers ask, then expand.

For the bigger picture of what AI can do on a website beyond chat, read adding AI to your website. Once you know what you need, my AI integration service helps you pick the right kind, build the bot on your own data and keep its answers accurate over time. See pricing or get in touch.

Sources:

An AI chatbot for your website: should you use one, and which kind?