Step 1: Decide what the chatbot answers from
A support chatbot is only as good as the material it answers from. Modern AI chatbots don't need to be "trained" in the machine-learning sense — they retrieve from the sources you connect at the moment a question is asked, and answer from what they find. So the first job is picking those sources.
Start with whatever your team already answers from: a help centre or docs site is the best first source, because it's already organised around real customer questions. Past resolved tickets are the most underrated second source — they show how customers actually phrase things. A sitemap crawl is the fastest way to connect a whole site, but exclude careers pages, old blog posts and marketing pages before the first crawl, or they'll dilute the answers.
- Help centre or docs site — connect first
- Past resolved tickets — connect early; they hold the real phrasing
- Product and pricing pages — useful for pre-sale questions
- Changelog — lets it answer "has this been fixed yet?"
Step 2: Decide what it must never answer alone
Before the chatbot goes live, decide where it should stop. Good AI chatbots give each answer a confidence score and let you set a threshold below which the bot hands off to a person instead of guessing. Set stricter thresholds for anything involving money, refunds or contracts, and looser ones for low-stakes questions like opening hours or password resets.
Some topics shouldn't be automated at all — cancellations tied to a retention conversation, safety complaints, legal or compliance questions. Mark those as never-automate so they route straight to a human however confident the bot is. And make sure any customer who asks for a person gets one immediately.
Step 3: Install the chat widget
Almost every AI chat widget installs the same way: one script tag pasted before the closing </body> tag of your site. It loads asynchronously, so it shouldn't slow your page down. Where that tag goes depends on your platform:
- WordPress: Appearance → Theme File Editor → footer, or a header-and-footer plugin like WPCode if you'd rather not edit theme files.
- Shopify: Online Store → Themes → Edit code → theme.liquid, just before the closing body tag.
- Webflow: Project Settings → Custom Code → Footer Code, then publish (custom code doesn't run in the Designer preview).
- Wix: Settings → Custom Code → Add Custom Code, load on All pages, place in Body – end.
- Squarespace: Settings → Advanced → Code Injection → Footer (needs a Business plan or above).
- Ghost: Settings → Code injection → Site Footer.
- Framer: Project Settings → General → Custom Code → End of <body> tag, then publish.
- React, Next.js, Vue, Svelte, Angular: add the script tag to your root layout or index.html — no SDK needed.
Step 4: Tell it who it's talking to
If visitors are signed in, pass their email, name and user ID to the widget. The chatbot can then look up an order or a plan instead of asking the customer questions your app already knows the answers to — which is the difference between a chatbot that feels like a search box and one that feels like a colleague.
Step 5: Test it with your hardest real questions
Before announcing anything, open your last month of tickets and ask the chatbot twenty of the questions you actually received — including the awkward ones. Check three things for each: did it answer correctly, did it cite the article it used, and did it hand off when it should have? A wrong answer at this stage almost always points to a missing or ambiguous article, not a broken bot; fix the article and ask again.
- Ask a question your docs don't cover — it should say so and offer a person, not improvise.
- Ask about refunds or billing — it should be conservative or hand off.
- Type "talk to a human" — it should hand off immediately.
- Ask in another language — check the reply is in that language.
What "free" AI chatbots usually leave out
Plenty of tools advertise a free AI chatbot. The free tier is usually capped at a small number of AI messages a month, carries the vendor's branding, and sometimes leaves out human handoff entirely — which means once the cap is hit or the bot doesn't know, the customer hits a dead end. When comparing, check what happens at the cap, whether handoff is included, and whether you're billed per AI answer, per conversation or per "resolution" — the last one gets more expensive as the bot gets better.
nDesk has a 7-day free trial with 250 AI answers and no card required, then plans from $29 a month with 500 AI answers included and human handoff on every plan.