You've probably met the classic chatbot: a menu of buttons, "press 1 for delivery, 2 for returns." A customer types a normal question, the bot can't find it in its script, and replies "I didn't understand, please pick an option." That's where its usefulness ends. An AI bot — an assistant built on a language model (LLM) — starts from the opposite end: it understands what the person wrote in their own words, even if no one planned for that question in advance.
This isn't a cosmetic upgrade. It's the difference between a form that talks and an assistant that actually answers. Let's look at what it adds over a plain chatbot, and where it has to be kept on a short leash.
The real difference from a scripted chatbot
A scripted chatbot only works along the paths you drew for it. Every new question needs a new branch, coded by hand. Ten ways to ask "how much is shipping" mean ten rules, or you lose the customer.
An AI bot doesn't run on branches. It understands the intent behind the phrase. "My parcel came dented, what do I do?" and "the product arrived damaged" lead to the same answer, even though you never wrote them out separately. More than that, it can answer from your documents: feed it your return policy, your product list, your user guide, and it answers from those, not out of thin air.
What an AI assistant actually does
It isn't magic, and it doesn't replace your team. It takes over the repetitive, clear-cut pieces, much like a well-trained new hire.
- Answers from your knowledge base. Questions about hours, shipping, warranty, specs — pulled from your documents, not a generic internet.
- Drafts replies for support and sales. It reads the customer's message and prepares an answer your agent checks and sends in seconds.
- Qualifies leads. It asks the right questions, captures the contact and the need, and flags who's worth calling first.
- Triages and hands off to a human. When a question is sensitive or the bot isn't sure, it doesn't force an answer — it passes it to a colleague with the full conversation context.
For us, the qualification part often ties to real data: a new lead can be enriched automatically from Datero.md, and a confirmed order can flow straight into your process — the way Pivoteka.md routes orders to the pickup point and courier on its own after MAIB payment.
Where the bot can live
You don't have to pick one place. The same assistant, with the same knowledge base, can sit in several channels:
- Telegram — where many customers in Moldova and the region already write.
- A website widget — the chat corner on your pages, wired to your CRM.
- WhatsApp — for businesses where the customer prefers it there.
The idea is the same throughout: the customer writes where it's convenient, the bot answers just as well everywhere, and you run one brain, not three separate bots.
The limits, said honestly
An AI bot can be wrong. A language model sometimes "hallucinates" — it produces an answer that sounds confident but is false. Left unchecked, it can promise a customer a discount that doesn't exist or a wrong return policy. That's why we build with three firm rules:
- Grounded in your data. It answers from your documents, not from what the model "thinks" in general.
- With clear guardrails. It doesn't promise prices, confirm orders, or invent policy. Sensitive topics go to a human automatically.
- Under human review. At the start you read its answers and correct them. The human stays on top, always.
And one practical tip: don't start with "a bot that knows everything." Start with one repetitive, high-volume task — shipping questions, lead qualification, the first draft of a reply. Get that working well, then expand. A narrow, accurate bot beats a broad, unreliable one every time.
It's worth it when you have volume: dozens of messages a day with the same questions, a small team stuck on repetitive replies, customers who write in the evening when no one's at the desk. It's not worth it for five messages a week — there a person answers better and cheaper.
At KERNEX we build and integrate these assistants, wired into the systems you already run. Email us at hello@kernex.md and we'll start from a real task, not from promises.