When a business asks for a custom AI chatbot trained on your data, it usually pictures a model being fed thousands of company documents until it has learned them. That is almost never how these chatbots are built, and knowing the difference will save you money and a few bad decisions.
AI development is one of our core services, and this is the explanation we give in the first call: what "trained on your data" really means, what the build involves, what it costs, and what makes the difference between a chatbot people trust and one they stop using in a week.
What "trained on your data" really means
In most projects the AI model is not retrained at all. The chatbot uses a method called retrieval-augmented generation, usually shortened to RAG. It works in three steps every time someone asks a question:
- Search. The system searches your documents for the passages most related to the question.
- Hand over. It gives those passages to the AI model along with the question.
- Answer. The model writes an answer using only what it was handed, and can show which document each part came from.
Think of it as an open-book exam. The model is not expected to remember your refund policy. It is given the page and asked to answer from it.
Why retrieval beats retraining for most businesses
- Updates are instant. Change a document and the next answer uses the new version. A retrained model would need retraining again.
- Answers can show their source. A link to the exact policy or manual page is what makes staff and customers trust the answer.
- Access can be controlled. The search step can be limited to the documents a particular user is allowed to see.
- It costs far less. Retraining a model is a specialist job with its own data preparation; retrieval is engineering.
Retraining, called fine-tuning, still has a place. It helps when you need a fixed tone or a very specific output format. It is not how you teach a model your facts.
What goes into building a custom AI chatbot trained on your data
The chat window is the smallest part. Most of the work is in the steps behind it.
- Collect the sources. PDFs, help articles, product sheets, past support answers, database records. Decide which are current and which are not.
- Clean and split. Documents are divided into passages small enough to search well. Scanned PDFs, tables and duplicated old versions need the most care here.
- Index. Each passage is stored in a search index that finds text by meaning, not only by matching words.
- Write the rules. The model is instructed to answer only from the passages, to say when it does not know, and to hand over to a person when needed.
- Connect the channels. Website widget, WhatsApp, your app or an internal tool. Each has its own login and formatting rules.
- Test with real questions. A set of real questions with known correct answers is run before launch and after every change.
If the chatbot should also do things, such as book an appointment or check an order, that is a step further. We explain where that line sits in what an AI agent really is.
What it costs to build
One November 2026 cost guide puts a custom chatbot that answers from your own documents, with one or two integrations, at $12,000 to $25,000 and six to nine weeks when built by a small or offshore team. It puts the same scope from a US-based team at $60,000 to $120,000. Very large document sets and many channels sit above those ranges.
The build is only part of the bill. There are two running costs:
- Model usage. The AI model is billed by the amount of text it reads and writes, so cost rises with the number of conversations and with how much text is sent per question.
- Hosting and upkeep. The search index, the server, monitoring, and someone to review failed answers each month.
For how these figures compare across other kinds of AI work, see our breakdown of AI development cost in India.
How to stop it making things up
An AI model that is not given the answer will often produce a confident one anyway. In a business chatbot that is the main risk, and it is handled by design, not by hope.
- Answer only from sources. If the search finds nothing relevant, the chatbot says it does not know and offers a person.
- Show the source. Every answer links to the document it used, so a wrong answer is easy to spot and fix.
- Keep a test set. Fifty to a hundred real questions with agreed answers, re-run after every change.
- Review the misses. Unanswered and down-voted questions are the best list of documents you still need to write.
- Fix the documents first. If two documents disagree, the chatbot will too. Most accuracy problems are content problems.
Privacy and where your data goes
Ask three questions before any build: where the documents are stored, whether the AI provider may use your content to train its models, and who can see conversation logs. For sensitive data, the options range from a commercial model with a no-training agreement to an open model on your own server. We compared those choices in open-source AI models for business.
Also check the rules on telling users they are talking to an AI. They are tightening in several markets, and we summarised them in AI disclosure rules for 2026.
When a ready-made tool is enough
You do not always need a custom build. If you have a few dozen help articles, one website and no need to connect other systems, a subscription chatbot product will do the job for a monthly fee. Build custom when you need control over the sources, per-user access, your own channels such as WhatsApp or an app, or integration with your systems.
If you are unsure which side you are on, send us a short description of your documents and channels on the contact page. Our AI development page shows how we scope this kind of work.
