Chatwize
Back to blog
Tutorials

Practical guide

A multilingual AI chatbot live in one day: practical guide

Chatwize teamImplementationPublished on 8 April 20268 min read

Going live with a multilingual AI chatbot on your website isn't a multi-week project. With the right preparation you go from empty dashboard to a bot helping customers in 95+ languages in a single day — no developer needed. Here's how we do it on implementations.

Prep (1 hour, before)

The boring but important part. The quality of your chatbot is directly tied to the quality of the sources you feed it. Before you log in, gather:

  • The 5 most-visited product pages
  • Your current FAQ (or a dump of common email questions)
  • Returns, delivery and payment policies
  • Tone-of-voice guideline — if you have one; otherwise 3 example replies from your best support agent

Most important filter

Only give the bot content you also publish online. Internal docs with margins, agreements or non-public pricing don't belong in the chatbot.

Setup (2 hours, in the dashboard)

Step 1 — Branding in 5 minutes

Paste your primary colour, pick a corner radius that fits your site, upload the logo. Done. This is the part people want to spend longest on — don't. Move on, you'll tweak it later anyway.

Step 2 — Add sources

Add the URLs from your prep. Not your whole sitemap — the deliberate 5–15 pages you need. A focused bot outperforms a bot that's drowning in noise.

Step 3 — Buttons, not open fields

Customers don't love typing. Add 3 to 5 buttons in the welcome screen for the most common paths: “Browse products”, “Delivery time”, “Get in touch”. Lowers friction and steers the conversation to predictable outcomes.

Step 4 — System prompt: short and concrete

Good system prompts fit on an index card. What's your role, for whom, in which language do you reply (answer: the same language as the question), what do you never do, and when do you escalate? Nothing more.

Test (1 hour, before going live)

This is where most implementations save the most time — and where most go wrong. Ask the same 10 questions across 4 languages and don't check whether the bot “says something” — check whether it says it correctly. Examples:

  1. A product question you know the answer to
  2. A question your sources don't cover
  3. A question in a non-default language
  4. An “I want to talk to a human” question
  5. A question with a price or stock component
  6. A returns or warranty policy question
  7. A question outside your domain (test that it politely declines)
  8. A question with a typo
  9. A multi-step question (“I'm looking for X for Y”)
  10. A question from an unhappy customer

What to do when a test fails

Don't hand-write the answer. Add the missing source or sharpen the system prompt. A bot hard-coded with replies is a 2018 chatbot.

Live + the first two weeks

Go live on a quiet moment of the week. Block two review sessions: at 7 and at 14 days, 30 minutes each. Walk the logs, flag the 5 replies you're not happy with, and update the source or prompt in the same session. That's where 80% of your improvement gains live.

1 day

for a full go-live

95+

languages out of the box

30 min

weekly iteration in month one

Ready to make this happen for your team?

Book a short demo and we'll show how Chatwize fits your customer questions, channels and processes.