How a Custom AI Chatbot Differs From a Generic Chat Widget

Almost every website now has a small chat bubble in the corner. From the outside they all look alike: a friendly greeting, a text box, perhaps a row of suggested questions. But two bubbles that look identical can behave in completely different ways once a visitor starts typing. One of them recites a handful of canned answers and gets lost the moment a question steps outside its script. The other knows the business it speaks for, can look things up, remembers what was said two messages ago and knows when to bring in a person.
This article explains where that difference comes from. It describes what a generic chat widget really is, what a custom-built AI chatbot adds, which sources a custom chatbot is built from, and how you can judge the difference yourself instead of taking a vendor's word for it. It contains no promised results and no customer figures, because what a chatbot achieves always depends on your website, your content and your visitors.
What people usually mean by a generic chat widget
A generic chat widget is a product that is built once and installed many times. The company behind it designs the bubble, the window and the logic, and every customer receives the same thing with a different logo. Setup usually means pasting a snippet of code into your website and filling in a form.
Under the surface, most generic widgets work in one of a few ways:
- Canned answers. You enter questions and answers, and the widget matches a visitor's message against them. If the match is close, the stored answer appears. If not, the widget falls back to something like "Sorry, I did not understand that."
- Decision trees. The visitor clicks through a fixed set of buttons ("Orders", "Returns", "Something else"). The path is predefined, and free text is either ignored or only used to pick a branch.
- A general-purpose language model with a thin layer on top. Newer widgets connect a general AI model and give it a short description of your company. The replies sound fluent, but the model has no access to your actual records and no reliable knowledge of your specifics, so it tends to answer with plausible generalities.
None of these is useless. For a small site with ten recurring questions and no customer data behind them, a simple widget can be a perfectly reasonable choice. The trouble starts when a business expects more from it than it was built to do. Our comparison of generic chatbots and custom email automation looks at that decision from the angle of email; here we look at the same question for chat on your website.
Three things a generic widget cannot do
When a generic widget disappoints, it is almost always for one of three reasons.
It has no access to real data
Visitors rarely ask abstract questions. They ask "Where is my order?", "Is this jacket still available in blue?" or "Do you have an appointment free on Thursday?" Those are questions about facts that live in a shop system, a stock database or a calendar. A widget that has never been connected to any of them can only answer in general terms ("Orders usually arrive within a few days") or send the visitor away to another page. That is the moment where a visitor decides the chat is not worth their time.
It has no real memory of the conversation
Many widgets treat each message as an isolated question. Ask "How much is the Team plan?" and then "And does it include the timeline view?" and the second question may arrive without any idea what "it" refers to. People talk in follow-ups. A chat that cannot hold the thread forces visitors to repeat themselves and makes the conversation feel like filling in a form.
It does not know where its knowledge ends
A canned-answer widget fails visibly: it says it did not understand. A generic AI widget can fail invisibly, because a fluent model will happily produce an answer that sounds right but is not backed by anything. Neither behaviour is what a business wants. The honest behaviour, "I do not have that information, but I can pass your question to a colleague", needs to be designed in. It rarely comes with a template.
What a custom AI chatbot adds
A custom AI chatbot is built around one specific business. It is still based on modern language models, so it can understand free-form questions and answer in natural language, in the visitor's own language. What is different is everything around the model: what it is allowed to know, what it can look up, how it behaves when it is unsure, and where it hands over.
At InboxFirma, this is the idea behind the custom AI chatbot for your website: we build and configure it together with you rather than handing you a template. Which questions it covers, which sources it uses and which actions it may take are agreed during setup and can be adjusted afterwards.
How a custom chatbot is built: four sources
The easiest way to see the difference is to look at what a custom chatbot is built from. There are four typical sources, and a project may use some or all of them.
Your website
Your website already contains a great deal of what visitors ask about: services, products, opening hours, delivery information, contact details. A custom chatbot can learn from the pages and product information you choose to include. The words "you choose" matter. Not every page belongs in a chatbot's knowledge, and an outdated page should not become an outdated answer.
Your documents
Much of a business's real knowledge never appears on its website: manuals, price lists, policy documents, onboarding guides, internal checklists. A custom chatbot can use the documents you approve, so that an answer about a warranty condition or a technical detail comes from your document and not from a guess.
Your FAQs
Your existing FAQ is a record of the questions your visitors really ask, in the wording your team has settled on. A custom chatbot can answer common questions in line with that wording, which keeps its replies consistent with what you already tell people elsewhere. The difference from a canned-answer widget is that the visitor does not need to phrase the question in exactly the way the FAQ does.
Custom integration with your internal systems
This is the source that separates a chatbot that talks about your business from one that works with it. Where a custom integration exists, the chatbot can look things up in your own systems: an order in your shop, stock in your inventory, free slots in your booking calendar. It is the part that has to be scoped individually, because every business has different systems and different interfaces. Which systems can be connected, and what the chatbot may read or do there, is assessed per project. Our integrations page describes how we approach this, including that it is a custom piece of work and not a ready-made connector for every product.
The four sources are also the four boxes you will find under "How We Train Your AI Chatbot" on our demo page, which we come back to below.
What a custom chatbot can do that a generic one cannot
With these sources in place, the behaviour changes in ways visitors notice quickly.
It grounds answers in real data
When a visitor asks about an order, a custom chatbot with the right integration does not guess. It retrieves the record and builds its answer from what it finds: the status, the carrier, the expected date. When the visitor asks whether a product is available, the answer comes from the stock data, not from the product description. The rule that matters is simple: facts such as statuses, dates, quantities and availability should only ever come from the data, never from the language model's imagination.
This also changes what happens when nothing matches. If the order number does not exist in the system, a well-built chatbot says so, asks the visitor to check the number and offers a person. It does not fabricate a status to be helpful.
It holds a natural multi-turn conversation
A custom chatbot keeps the conversation in view. If a visitor mentions an order number in the third message and asks "And when will it arrive?" in the sixth, the chatbot knows which order is meant. It can greet, answer a small question about opening hours, move on to a real lookup, take a follow-up and close with a thank-you, all as one flow rather than a string of unrelated question-and-answer pairs. It also writes the way a person at your business would: briefly, politely, in the visitor's language.
It can take limited, defined actions
Where a custom integration allows it, a chatbot can do more than answer. It can propose free appointment slots from a real calendar and, once the visitor confirms, write the booking. It can collect the details of a problem, summarise them and create a support ticket for a colleague. These are deliberately narrow actions with clear rules, not open-ended autonomy. Which actions exist at all is decided in setup, and many businesses start with the read-only ones.
It hands over to a person at the right moment
A custom chatbot is designed to know its limits. Complaints, legal questions, sensitive matters or simply a visitor who asks for a human should lead to a person, not to a longer chatbot answer. A good hand-over also saves the visitor from starting again: the chatbot gathers what a colleague needs, summarises it and says clearly that someone will follow up, without promising a response time it cannot guarantee.
A short illustration
To make this concrete, here is a fictional example of the same question put to two different chat bubbles on a fictional online shop. The shop, the order number and the replies are invented for illustration.
A visitor writes: "Where is my order LU-10482?"
A generic widget might reply: "Thanks for your message! Orders usually ship within 2-4 working days. You can find tracking information in your confirmation email." That is not wrong, but it does not answer the question.
A custom chatbot with a shop integration can look up the order and reply along the lines of: "Your order LU-10482 with the espresso machine has been shipped with DHL and should arrive on Monday. Here is your tracking number." If the visitor then asks "Can I also book a product advice call for tomorrow morning?", it can check the calendar, notice that tomorrow is a Sunday on which no calls are offered, and suggest the next free slots instead.
The visitor never had to choose a category, repeat the order number or leave the chat. That is the difference in one example.
See the difference yourself
Reading about this only goes so far. For that reason we built a live AI chatbot demo where you can experience it directly. You choose one of six fictional businesses (an online shop, a software company, an agency, a real estate agency, a hotel and a local hair salon) and chat with its assistant. Next to the chat, an "AI Actions" panel shows what the assistant does behind the scenes: which lookup it ran, what the system returned and, where relevant, the booking or support ticket it created.
The businesses, orders, stock levels and calendars in the demo are all fictional, and the demo is not connected to any real company's systems. It is meant to show how a conversation with a custom chatbot feels and how clearly the real data and the conversation are kept apart. We suggest you try a few things that a generic widget would struggle with: say hello and ask something unrelated, then ask about an order, refer back to it in the next message, book an appointment and finally ask to speak to a person.
A related product, not the same one
A word on how this fits with our other product. The AI Email Assistant and the AI chatbot are related but different. The AI Email Assistant reads and answers customer emails inside the InboxFirma application, with approval rules that decide what may be sent automatically and what a person must check. The chatbot talks to visitors on your website, in real time. They can share ideas, such as grounding answers in your own data and handing sensitive matters to a person, but they solve different problems in different places and are set up separately. If your visitors write to you as often as they chat, you might eventually want both, but neither requires the other.
Questions to ask before you choose
Whether you are looking at a template or a custom build, these questions help separate marketing from substance:
- Where do the answers come from? Can the provider show which sources the chatbot uses, and can you control them?
- Can it look anything up? If it should answer questions about orders, stock or appointments, which systems does it connect to and how?
- What happens when it does not know? Ask to see how it behaves with a question it cannot answer. The honest "I am not sure" is a good sign.
- How does the hand-over work? Does a person receive context, or does the visitor start from zero?
- Does it remember the conversation? Try a follow-up question that only makes sense in context.
- What does it do with sensitive topics? There should be a clear rule that such questions go to a person.
- Who maintains it? A chatbot is not a one-off task. Content changes, products change and rules need review.
Starting small
A custom chatbot does not have to begin with every integration at once. A sensible first step is a chatbot that answers questions from your website, documents and FAQs, with a clear hand-over to your team. Once you see which questions come up most and where visitors need real data, you can decide which integration is worth building first. That order keeps the risk low and lets the chatbot earn its place before it is given more responsibility.
If you would like to talk through what this could look like for your own website, you are welcome to get in touch. We will start with your visitors' real questions and work out which of the four sources they point to.


