AI Chatbot for Your Business Website: What It Can and Cannot Do

It is 23:00. A potential customer is on your website with a simple question: do you deliver to their town, and roughly how much would it cost? Nobody is at the desk. The contact form promises a reply "within one business day", so they open a competitor's site instead.

That gap is what AI chatbots are good at closing. They have also been oversold. This guide separates what a chatbot can realistically do for a small business from what it cannot, and shows how to launch one you can trust.

What an AI chatbot actually is today

The old chatbots were decision trees: press 1 for opening hours, press 2 for prices. They broke the moment a visitor typed something unexpected.

A modern AI chatbot uses a large language model to understand a question in natural language and answer it in natural language. The important part for a business is the second ingredient: it is connected to your own content (your FAQ, price rules, service descriptions, policies), and it is instructed to answer from that material. This approach is often called retrieval-augmented generation. The model provides the language ability, and your content provides the facts.

That is the difference between a bot that sounds fluent and a bot that is actually useful. Fluency without your facts is how you get confident nonsense.

What it does well

• Answers repeat questions instantly. Opening hours, delivery areas, how a process works, what is included in a service, what to bring, how to cancel. If your inbox is full of the same ten questions, this is the biggest win.
• Covers the hours nobody is working. Evenings, weekends and holidays.
• Replies in the visitor's language. A well-configured bot can answer in Latvian, English and other languages from the same source content, which helps a lot for businesses serving mixed audiences.
• Captures leads. It can ask a couple of qualifying questions and collect a name and phone number or email, so a visitor who would have left becomes an enquiry.
• Routes people to the right place. Book a call, message on WhatsApp, open the right page.
• Reduces interruptions. Staff spend less time on questions with a written answer and more on work that needs a person.

What it cannot (and should not) do

• It cannot know what is not in its content. If your prices depend on a case-by-case assessment and you have not written down the rules, it has nothing sound to say.
• It can be wrong. Language models sometimes produce plausible but incorrect answers. Good setups reduce this a lot by restricting the bot to your sources and allowing it to say "I am not sure, let me connect you with a person". They do not eliminate it.
• It should not make binding promises. Firm quotes, contract terms, refunds, medical or legal advice, and anything with real consequences should be handed to a human.
• It is not a replacement for good service. It buys you speed and coverage. It does not fix a slow team, confusing prices or an unclear offer.

When a chatbot pays off, and when it does not

It usually pays off when:

• You get the same questions again and again by email, phone or DM.
• Enquiries arrive outside working hours and slow replies cost you customers.
• You serve people in more than one language.
• Each new lead is worth real money, so a missed one hurts.
• Your services are explainable in writing.

It usually does not, yet, when:

• Your site gets very few visitors, so there will be few conversations.
• Your sales are deeply consultative and every case is different.
• Nobody has time to maintain the source content. A chatbot with outdated information is worse than no chatbot.

A rough way to check the numbers: estimate the extra leads you would capture each month, multiply by your close rate and average job value, add the staff hours you would save, and compare that with the setup cost plus the monthly platform and upkeep. If the result is clearly positive on conservative assumptions, it is worth trying. If you need optimistic numbers to justify it, wait.

The trust and legal side (EU and Latvia)

This is not legal advice, but there are three points every business should address before launch.

1. Tell people it is an AI. EU rules on AI transparency expect users to be told when they are talking to an AI system rather than a human. Even without that, it is simply good practice. Put it in the greeting and the widget header ("Velinx AI assistant"), and give an easy way to reach a person.
2. Treat chats as personal data. Names, emails, phone numbers and free-text messages fall under GDPR. You need a clear privacy notice, a lawful basis, sensible retention limits, and a data processing agreement with the vendor that runs the AI. Find out where the data is processed and whether conversations are used to train models.
3. Do not invite sensitive data. Instruct the bot not to request health details, ID numbers or payment card data in chat, and make it point people to the proper channel instead.

How to launch a chatbot that people trust

1. Define the scope. List the 20 questions you get most often. Start there, not with "everything".
2. Write the source content properly. Clear answers, current prices or price rules, delivery areas, policies. Whatever is vague in your content will be vague in the bot.
3. Set the tone. Friendly and direct is usually better than corporate. Decide how it introduces itself.
4. Add guardrails. Answer only from your sources. Say "I do not know" instead of guessing. Escalate automatically for complaints, negotiations, urgent problems and anything legal or medical.
5. Build the handover. WhatsApp, phone, email or a contact form, with the conversation summary passed along so the customer does not have to repeat themselves.
6. Test with real questions. Use your actual customers' phrasing, including typos, slang, Latvian declensions and questions designed to trip it up.
7. Launch small, then review weekly. Read the conversations. Every question it could not answer is a gap in your content, and every wrong answer is a rule to add.

What to measure

• Conversations per week and how many are outside working hours.
• Resolution rate: how many were answered without a human.
• Handover rate: how many needed a person, and why.
• Leads captured and how many became customers.
• Unanswered questions: the most valuable list you will get. It tells you what your website and FAQ are missing.

The SEO angle people forget

A chatbot does not make your site rank, and text that appears only inside a chat widget is not something search engines read as page content. Two consequences:

• Keep important information on normal pages. Prices, services, delivery areas and policies belong on crawlable pages, not hidden behind the bot.
• Turn the bot's unanswered questions into content. Real visitor questions are the best FAQ and blog topics you will ever find, and they match how customers actually search. Our guide to local SEO in Latvia shows how to use that content.

Also make sure the widget loads lightly. A heavy script can hurt your page speed, so load it after the page is usable. We explain why in our Core Web Vitals guide.

What about phone calls?

The same principles apply to an AI receptionist that answers calls: give it a narrow job (answer common questions, take messages, book appointments), be clear that it is automated, comply with rules on recording calls, and always provide a route to a human. For appointment-based businesses such as clinics, salons and workshops, missed calls are missed revenue, and this is where an AI receptionist can earn its place.

Want to see one?

You can see an example in the chatbot section of our homepage. We scope every chatbot and receptionist individually, trained on your own content and languages, so tell us what your customers ask most and we will send a free estimate within 24 hours.

Frequently asked questions

Will an AI chatbot replace my staff?

No. It takes the repetitive, information-only conversations off your team so they can focus on work that needs judgement and a personal touch. Think of it as extended opening hours for basic questions.

Can an AI chatbot speak Latvian?

Modern models handle Latvian reasonably well, and a bot can answer in Latvian and English from the same source content. Quality depends on the domain, so test it with your own terminology, product names and the way your customers actually write, including inflected forms.

How long does it take to set one up?

For a small business with clear content, days to a couple of weeks. Most of the time goes into preparing accurate source content and testing, not into the technology.

Is an AI chatbot GDPR compliant?

That depends on how it is set up, not on the technology itself. You need a privacy notice, a lawful basis, retention limits, a processing agreement with the provider and clarity about where data is stored. If you are unsure, ask a data protection professional to review your setup.

What if the chatbot gives a wrong answer?

Reduce the risk by restricting it to approved sources, letting it say it does not know, escalating sensitive topics and reviewing conversations regularly. You remain responsible for what your website says, including its chatbot, so monitoring is not optional.