The short version: the right AI chat tool for your website is the one that answers your customers’ actual repeat questions correctly, hands off to a human the moment it’s stuck, and costs you less than the support hours it saves. Everything else, the flashy demo, the “powered by GPT-4” badge, the case study logos, is noise. Pick based on your last three months of support transcripts, not a sales call.
Start with your own transcripts, not the tool’s demo
Every AI chat vendor will show you a beautiful demo where a customer asks “what are your opening hours” and the bot answers instantly. That tells you nothing. Your customers don’t ask polite, easy questions. They ask “why hasn’t my order arrived, it’s been nine days and I paid for express” and “can I use this on a Mac if I already bought the Windows version.”
Before you look at a single tool, pull your last 200 support tickets or live chat transcripts. I did this with a client, a small accountancy practice in the north west with about 40 staff, before we installed anything. We found that 61 percent of their inbound chats were one of six questions: pricing, deadline extensions, document upload problems, “are you VAT registered,” appointment rescheduling, and “can I speak to someone now.” That list became our test script for every tool we trialled.
If you skip this step, you’ll buy on vibes and end up with a bot that can chat fluently about nothing your customers ask.
The uncomfortable bit nobody selling you a chatbot wants to say
Here’s the part most comparison articles gloss over: a badly trained AI chat tool is worse than no chat tool at all. Not slightly worse. Meaningfully worse, because it answers with confidence even when it’s wrong, and customers trust confident-sounding answers.
The clearest public example is Air Canada. In February 2024, a tribunal in British Columbia ordered the airline to pay a customer after its website chatbot gave him incorrect information about bereavement fare refunds, information that directly contradicted the airline’s actual policy published elsewhere on the same site. Air Canada argued the chatbot was “a separate legal entity responsible for its own actions.” The tribunal did not agree. The company had to honour what its bot promised.
That case matters for you even if you’re nowhere near airline-sized. It proves the core risk: an AI chat tool trained loosely on your general website content, rather than tightly on your actual policies, will eventually make something up and say it with total confidence. If a customer acts on that and it costs them money, you’re the one who answers for it, not the software company. This is why “does it sound natural” is the wrong first question. “Where does it pull its answers from, and can I lock that down” is the right one.
The six things that separate good tools from expensive toys
1. What it’s trained on
Ask directly: does it only answer from documents I upload and pages I approve, or does it also draw on general internet knowledge to fill gaps? You want the former. Tools like Intercom’s Fin, Zendesk’s AI agents, and Ada let you restrict answers strictly to your help centre, policy pages, and product docs. Tidio’s Lyro and Crisp’s AI let you do the same on a smaller scale. If a tool can’t show you, in plain terms, which sources fed a specific answer, don’t shortlist it.
2. How it hands off to a human
This is the single biggest driver of whether customers end up angry or satisfied. A good tool escalates the moment it detects low confidence, a complaint, a refund request, or the words “speak to a person.” A bad one keeps looping the customer through the same three canned responses until they give up or start swearing at it in the chat window. Before you buy, ask the vendor to show you the exact escalation trigger settings, not describe them. Watch them set it up on screen.
3. Where the conversation lives after it starts
If your team already handles website chat, email, and WhatsApp, check whether the AI tool sits inside that same inbox or forces your staff to check a separate dashboard. Fragmented inboxes are how customer messages get missed for two days. If WhatsApp is part of your support mix, it’s worth reading up on converting a personal WhatsApp account to a business account so that channel feeds into the same place as your website chat, rather than living on someone’s personal phone.
4. Pricing model, and what happens when volume spikes
This is where people get caught out. AI chat pricing usually falls into one of three shapes: per seat (a flat monthly fee per human agent, AI included), per resolution (you pay only when the AI solves something without a human), or per conversation (you pay whether it solved anything or not). Intercom’s Fin charges per resolution, roughly $0.99 each when I last checked, which sounds cheap until your resolution rate is low and you’re paying for a lot of “sort of solved.” Tidio’s Lyro bundles a set number of AI conversations into its paid plans starting around $29 a month, then charges per conversation after that. Zendesk and Freshdesk tend to fold AI into higher-tier seat pricing, which suits businesses that already pay for seats anyway. None of these is objectively best. The trap is assuming a low headline price stays low once your traffic doubles in December.
5. What it does with customer data
If customers might type in an order number, email, or medical detail (even something as simple as “I need this because of my allergy”), you need to know where that text is stored, whether it’s used to train the vendor’s wider model, and whether you can delete it on request. UK businesses handling any personal data through the chat need this documented for UK GDPR purposes, not as an afterthought.
6. How easy it is to correct when it’s wrong
Every AI chat tool will get something wrong in its first month. The question is how fast you can fix it. Some tools let you correct an answer in two clicks and it’s live everywhere within minutes. Others require you to re-upload whole documents and wait for reprocessing. Ask to see the correction workflow before you sign anything, not after your first bad review mentions it.
A short, honest story about getting it wrong first
Back when I set up my first AI chat widget on a client site, we picked a tool mainly because the sales rep was likeable and the free trial was generous. Within the first week, the bot told a customer asking about a service package that a feature was included when it had been discontinued eight months earlier, because it had pulled the answer from an old cached blog post rather than the current pricing page. The customer complained, reasonably, and we had to honour a discount to smooth it over. It cost us less than a bad Google review would have, but it taught me the lesson: restrict the source material first, personality second. The tool wasn’t broken. We’d set it up badly. That’s true of most AI chat disasters you’ll read about online. The tool rarely fails on its own; the setup does.
A step by step way to choose one
- Pull 100 to 200 recent support conversations and tag the top recurring topics. This is your test set, and it takes about an hour with a spreadsheet.
- Shortlist three tools that fit your existing budget and platform (check integration with your website builder or ecommerce platform first, if you’re mid rebuild it’s worth reading about choosing a DIY website builder so chat and site aren’t fighting each other).
- Feed each shortlisted tool your real help centre content and policy pages only, nothing generic.
- Run your top ten transcript questions through each tool live and score the answers for accuracy, not fluency.
- Set escalation rules deliberately, don’t accept the defaults, and test what happens when you type “I want a refund” or “this is a complaint.”
- Pilot on a low-traffic page or a single product line for two to four weeks before putting it on your homepage.
- Read every transcript from the pilot yourself, weekly, for the first month. Don’t delegate this to a dashboard summary.
If you’re weighing this up alongside a broader live chat decision rather than an AI-specific one, it’s worth reading what to look for in live chat software before you spend anything, because some of the same traps (poor handoff, hidden per-agent fees, weak reporting) apply whether or not the chat is AI-powered.
When you shouldn’t bother with AI chat at all
If your website gets under 200 visitors a day and your current inbox is manageable within a couple of hours, an AI chat tool is solving a problem you don’t have yet. A well written FAQ page and a simple “message us” form will outperform an AI bot that answers ten questions a week, because you’ll spend more time correcting its mistakes than it saves you. AI chat earns its cost when repeat, simple questions are eating hours of a real person’s week. Below that threshold, spend the money on something else, maybe better product photos, or finishing the customer engagement extras that differentiate a small site rather than a bot nobody notices.
And if the whole area feels like more than you want to work out alone, it’s a fair reason to bring in outside help rather than guess. I work with small businesses on exactly this kind of decision through AI implementation coaching, mostly because the setup mistakes (wrong source content, weak escalation, unclear pricing tier) are cheaper to avoid up front than to fix after three months of bad transcripts.
What good looks like, six months in
You’ll know you chose well when your team stops dreading Monday morning’s backlog of overnight questions, when the AI’s resolution rate for your top five recurring questions sits comfortably above 70 percent, and when the handoffs to humans arrive with full context attached rather than a customer having to repeat themselves. You’ll know you chose badly when customers start typing “let me talk to a real person” within the first two messages, every time.
Frequently asked questions
How much does an AI chat tool for a small business website cost?
Expect somewhere between $29 a month for a basic entry plan on tools like Tidio, and $300 to $1,000 a month for a mid-size business using Intercom or Zendesk with AI resolution add-ons, depending on conversation volume. Per-resolution pricing, common with Intercom’s Fin, tends to sit around $0.99 per solved query, which can be cheaper or pricier than a flat plan depending on how many questions your bot resolves without help.
Can an AI chatbot replace my human customer service team completely?
No, not for most small businesses, and treating it that way is how you end up with angry customers and no escalation path. AI chat tools handle repetitive, well-documented questions well; refunds, complaints, and anything emotionally loaded still need a human, and your escalation rules should reflect that from day one.
What’s the biggest mistake businesses make when choosing an AI chat tool?
Testing the tool on easy, generic demo questions instead of their own real support transcripts. A bot that handles “what are your hours” perfectly can still fail badly on the specific, messy questions your actual customers ask, which is why pulling your last few months of real chats before shopping matters more than any feature comparison chart.
Is it risky, legally, to let an AI chatbot answer customer questions on my website?
Yes, if it’s not restricted to accurate, current company information. A British Columbia tribunal ordered Air Canada to honour a refund policy its own website chatbot had misstated in 2024, confirming that businesses are responsible for what their AI tools tell customers, even when the answer was wrong. Lock the bot’s source material down to your current, approved content to avoid the same problem.
