Adding an AI customer assistant to a website is technically easy these days. A few lines of code open a chat window in the bottom-right corner that answers visitors' questions fluently. The hard part is controlling what that assistant says on your company's behalf. In this post we explain what a website AI assistant should and shouldn't do, and the points that often get skipped during setup.
What an assistant is good for, and what it isn't
A well-designed website assistant is especially useful for: giving a first answer to questions that arrive after hours, pointing visitors to the right page, answering frequently asked questions again and again, and collecting quote requests in an orderly way. Instead of filling out a form, visitors describe their need in their own words, and the assistant gathers the necessary details and passes them to your team.
There are also cases where an assistant is useless or even harmful. Letting AI speak on its own about binding matters such as exact prices, delivery dates, stock availability, warranty coverage or legal interpretation is risky. Language models can fill a gap in their knowledge with a sentence that sounds perfectly reasonable. An assistant on your site that says "this item ships tomorrow" has made a promise on your behalf, even if the item is out of stock.
Do: decide the source of knowledge yourself
The assistant should generate its answers from content you provide, not from general internet knowledge. That means your service pages, FAQs, business hours, contact details and approved statements are defined as the assistant's knowledge source. The assistant is instructed: "rely only on these sources; if it isn't in the source, say you don't know."
This approach is a small-scale example of AI that answers from company documents. If you want to work with a larger document archive, take a look at our post on AI that answers from company documents (RAG).
When setting up the knowledge source, keep one thing in mind: if your site has outdated or contradictory information, the assistant will repeat it. Setting up an assistant is often a good opportunity to review your site content as well.
Don't: let the assistant talk about everything
Draw the assistant's topic boundaries clearly. For questions unrelated to your company (homework, small talk, comments about competitors), a short, polite answer is enough. Otherwise you get unnecessary costs and odd answers that someone can screenshot and share.
Behaviors to avoid:
- Quoting prices, discounts or promotions that aren't in the source
- Making firm promises about delivery times or stock
- Giving personalized advice on health, legal or financial matters
- Presenting itself as a human; visitors should know they are talking to AI
- Asking visitors for more personal information than necessary
A plan for human handoff
The best assistant is the one that knows when to stop talking. When a visitor asks for a quote, makes a complaint, or the assistant fails to find an answer twice in a row, the conversation should be handed to a person. During business hours this can go to a live representative; after hours it becomes a record with a message like "our team will get back to you through this channel."
For the handoff to work, the conversation log must reach your team cleanly: who it was, which page they came from, what they asked and what the assistant answered. This log can be sent to a mailbox or a CRM. These connections are part of integration work, and they are often what makes the assistant genuinely useful.
KVKK and security
Everything typed into the chat window is data processing. Visitors may type their name, phone number or even health details. Under KVKK (Türkiye's Personal Data Protection Law), that means:
- The chat window should include a short privacy note and a link to the privacy notice.
- If the AI service's servers are abroad, this may count as a cross-border transfer of personal data, and the KVKK transfer rules need to be assessed.
- You should decide how long conversation logs are kept and who can access them.
- Check the provider's contract to see whether your data is used to train its models.
The cleanest approach is to handle these topics together with a KVKK compliance project.
Pre-launch checklist
| Check | Why it matters |
|---|---|
| Knowledge source is current and single | So the assistant doesn't repeat outdated information |
| Price and promise statements are disabled | Risk of binding wrong answers |
| The "I don't know" answer has been tested | It should redirect instead of making things up |
| Human handoff works | So no request gets lost |
| Privacy note is visible | KVKK obligation |
| Conversation logs are monitored | So wrong answers are caught early |
In the first weeks after launch, reading conversation logs regularly is the fastest way to improve the assistant.
How we do it at Globya
We don't set up the website assistant as a standalone plugin, but as part of your website, your mailbox and your CRM if you have one. We prepare the knowledge source with you and put topic boundaries and handoff rules in writing. KVKK requirements such as the privacy notice and log retention are part of the setup at no extra charge. In corporate website projects, we can also launch the assistant together with the first version of the site.
Frequently asked questions
Will an AI assistant replace our live support team?
Not entirely. It reduces the team's workload by handling repetitive questions; quotes, complaints and special cases should still be handed to a person.
What happens if the assistant gives wrong information?
To reduce this risk, the assistant is set up to answer only from approved content and to stay away from binding topics. Even so, logs should be read regularly and the content corrected based on any wrong answers.
Do we need to tell visitors they are talking to AI?
Yes, we recommend stating this clearly. It is the honest approach and it sets the visitor's expectations correctly.
Which languages can the assistant work in?
Current language models are fluent in many languages. However, the assistant is more likely to answer correctly in the languages your knowledge source is prepared in.
The Globya assistant is online 24/7; it answers right away and passes your question to the team if needed.