Custom LLM Chatbot vs Website Chat Widget: The Difference

A website chat widget is the chat window on your site. It might connect visitors to a human, follow a fixed script, or run the vendor's own AI. A custom LLM chatbot is the system that does the talking: a language model built around your content, your tone, your rules and your business systems. You can show a custom chatbot inside a widget. So the real choice is between a packaged widget whose AI you rent and configure, and a chatbot built in code for your business.

What a website chat widget is

A chat widget is a front end: the bubble in the corner of a page, the message window it opens and the code snippet that loads it. Most widgets come from a customer service or marketing platform, and what happens after the visitor types depends on how that platform is set up. Most widgets fall into one of three types:

Live chat. Messages go to a person on your team. The widget handles the conversation; your staff supply the answers.
Rule-based bots. The visitor clicks buttons or types keywords, and the bot follows a decision tree someone wrote in advance. It is predictable, but it fails as soon as a question falls outside the tree.
Vendor AI agents. Many platforms now include a language model that reads your help articles and answers in free text. You configure it through the vendor's settings, and it runs on the vendor's infrastructure under the vendor's pricing.

The third type is what most people mean when they compare a widget with a custom chatbot. It is fast to switch on and needs no engineering. The trade-off is that the vendor decides what it can connect to, how it reasons, and how you are billed.

What a custom LLM chatbot is

A custom LLM chatbot is software built for one business. Each part of the system is a design choice rather than a vendor setting:

The model. Which language model answers, and whether different models handle different tasks.
The knowledge. What the bot can look up when it answers: help content, product data, policies, past conversations, records in your own database.
The instructions. How it should sound, what it may and may not discuss, and when it should hand over to a person.
The tools. Actions it can take, such as checking an order, creating a CRM lead or booking a meeting, through integrations you control.
The guardrails and logs. Checks on what goes in and out, and a full record of every conversation stored where you choose.
The interface. Often a chat widget on your site, but it can also sit in an app, a portal, WhatsApp or an internal tool.

Put simply, a widget is the place where the conversation happens. A custom chatbot is the system that decides what gets said and what gets done.

The differences that matter to a buyer

1. What it knows and how it sounds

Packaged AI agents usually learn from a help centre and a set of documents. That works for factual support questions. It works less well when the value lies in how your business talks: how you qualify a lead, which objections you hear and how you answer them. A custom build can draw on sources a vendor product may not accept, such as transcripts of your best sales conversations or live data from your own systems.

2. What it can do, not just say

Answering questions is one job. Taking action is another: looking up an account, changing a booking, writing a qualified lead into the CRM with the right fields filled in. Vendor agents support the integrations the vendor has built. A custom chatbot can connect to whatever your business runs on, including in-house systems and older databases, because the integration code is written for you.

3. Who controls the logic

With a packaged widget, the vendor's product decisions shape your bot's behaviour, and vendor updates can change it. With a custom build, the routing rules, the handover triggers and the prompts are code you can read, test and change. This matters most when the bot's behaviour carries commercial or regulatory weight.

4. How you pay

The two pricing models work in very different ways. Intercom's Fin, a widely used vendor agent, charges 99 cents per resolution, according to Stripe's published case study. A resolution counts when the customer confirms the answer helped, or when they ask for no more help after the last AI answer. Independent pricing guides note that when Fin is used with Intercom, it is billed on top of the platform's seat fees. Outcome pricing is easy to start with, and the bill grows with volume. A custom chatbot works the other way round: you pay for the build up front, then for hosting and model usage. The main cost drivers are:

the number and complexity of integrations;
how much of your content needs cleaning and structuring before the bot can use it;
the number of actions the bot may take, and the safeguards each one needs;
conversation volume and the choice of model, which drive running costs;
who maintains it afterwards.

5. Who owns it

If you leave a widget platform, you leave its AI behind and usually rebuild your setup from scratch elsewhere. A custom chatbot delivered as code you own can be moved, extended or handed to another team. Whichever route you take, check whether conversation logs and any training data can be exported in a usable format.

How a conversation differs in practice

Take a prospect who asks, on a Sunday evening, whether your service works for a business of their size and what it would involve. A rule-based widget offers three buttons, none of which fit. A vendor AI agent finds the nearest help article and summarises it accurately, but in a neutral tone, then offers a contact form. A custom chatbot built on your sales conversations asks the two qualifying questions your best salesperson would ask. It answers in your terms, records the answers in your CRM and books a call if the lead fits.

Each of these can be the right choice in some cases. For a support team answering the same thirty questions, the vendor agent may be the right tool. The custom build pays off when the conversation itself is where the value lies, or when the bot needs to act inside your systems.

Risks that apply to both

Buying a packaged product does not hand the risk to the vendor. Three points apply whichever route you choose.

You are responsible for what it says

In Moffatt v Air Canada (2024), a British Columbia tribunal found the airline liable for negligent misrepresentation. Its website chatbot had wrongly told a customer he could apply for a bereavement fare after travelling. The tribunal rejected the argument that the chatbot was a separate entity. It held that the chatbot was part of the airline's website, and the airline was responsible for its content. The practical lesson: limit the bot to approved sources, make it say when it does not know, and keep a log of what it said.

Prompt injection

The OWASP Top 10 for LLM Applications lists prompt injection as its first risk. Prompt injection means input that alters the model's behaviour in unintended ways. It can come directly from a user, or indirectly from content the bot reads, such as a web page or file. OWASP's mitigations include limiting the model's access to what it strictly needs, requiring human approval for high-risk actions, and adversarial testing. The same list also includes "excessive agency": giving a model too much control over sensitive actions or tools. The more actions your chatbot can take, the more this matters. A custom build lets you set each permission yourself.

Disclosure under the EU AI Act

Article 50(1) of the EU AI Act requires providers to ensure that AI systems which interact directly with people inform them that they are interacting with an AI system. There is no duty to disclose where this is obvious to a reasonably well-informed, observant and circumspect person. These transparency obligations became enforceable on 2 August 2026. Legal commentary notes fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. If you serve EU users, label the bot clearly, whether it is a widget or a custom build.

When a packaged widget is enough

Your questions are mostly factual and already answered in a help centre.
You already use the platform for support and want AI on top of it.
Volume is modest, so per-resolution fees stay predictable.
The bot only needs to answer, not act inside your own systems.

When a custom LLM chatbot makes sense

The bot has to sound like your business, not like a help article.
It needs to read from or write to systems the vendor does not support.
You want to own the logic, the data and the code, and to change them without waiting on a vendor's roadmap.
Volume is high enough that per-outcome fees become a significant line item.
Its answers carry commercial or regulatory weight, so you need to control and test every rule.

Questions to ask before you choose

What exactly should the bot achieve: deflect tickets, qualify leads, book meetings or complete transactions?
Which systems must it read from, and which must it write to?
What should happen when it is unsure, and how does a person take over?
Where are conversations stored, and can you export them?
How is cost calculated at three times today's volume?
Who can change its behaviour, and how are changes tested before they go live?

What we learned building one

In a sales chatbot we built, the business owner's problem was that their best sales conversations did not scale. Every new prospect waited for a person, and a generic chatbot did not sound like the business. Prospects can tell when they are talking to a script. So the chatbot learns from the company's own previous sales conversations rather than from a template. It handles the first conversation and qualification around the clock and talks to new prospects the way the owner would. A template-based widget could not have done this, because the value came from the business's own conversations.

Khive is a software engineering company based in Islamabad, Pakistan, founded in 2017. Our team of 50 engineers builds LLM chatbots and other production software for businesses in the UK, US and EU. We build each system in code and hand it over, so the client owns it.

Next step

If you are weighing a packaged widget against a custom build, tell us what the chatbot needs to do, which systems it has to work with and roughly how many conversations you expect. Send us a brief and we will tell you which approach fits.