How to build an AI chatbot trained on your own data that does not make things up

An AI chatbot trained on your own data is only worth launching if it answers from your documents and nothing else. The reliable way to build one is retrieval (RAG), not retraining a model: the bot looks up the relevant passages in your documents for every question, answers only from them, shows the source, and hands the conversation to a person when the documents do not cover it. With an AI-first team a first version takes 3.5 to 7 days and costs 4 times less than a traditional hand-written build, and it is 4 times faster.

The stakes are not theoretical. In February 2024 the Civil Resolution Tribunal of British Columbia ruled in Moffatt v. Air Canada that the airline was liable for a bereavement refund policy its website chatbot had invented. Air Canada argued the bot was responsible for its own words; the tribunal disagreed and ordered it to pay the difference. Whatever your bot tells a customer, your company said it.

What "trained on your own data" actually means

Most people asking for a chatbot trained on their data do not need training at all. There are three ways to give a model your knowledge:

• Prompt with context. Paste a short policy or price list into the instructions. Fine for a page or two of facts, breaks down beyond that.
• Retrieval (RAG). Index your documents, fetch the few relevant passages per question, and answer from them. Works for thousands of pages, updates the moment a document changes, and every answer can cite its source.
• Fine-tuning. Retrain a model on labelled examples. Useful for tone or a narrow output format, poor at facts: it cannot cite a source and needs retraining every time a price changes.

For a business chatbot the answer is almost always RAG, with fine-tuning as a rare extra. That also keeps the budget down, because you are not building a labelled training set.

How to build an AI chatbot trained on your own data in six steps

1. Pick one job and one channel. "Answer delivery and returns questions on the website" is a job. "A chatbot for the company" is not. Narrow scope is what makes accuracy measurable.
2. Clean the knowledge before you index it. Outdated PDFs and three versions of the same price list produce three different answers. Preparing the knowledge base often takes longer than the technical build, so start here.
3. Connect live data where documents go stale. Prices, stock, order status and delivery dates should come from your CRM, shop or database through an API, not from a document someone forgot to update.
4. Write the refusal and handoff rules. What the bot must never answer (discounts, legal, medical, exceptions), when it says "I do not know", and how a person picks up the conversation with full history.
5. Test on real questions before launch. Collect 50 to 200 questions customers actually asked, with the correct answers, and score the bot against them. Rerun that set on every prompt, model or document change.
6. Log, review, own. Keep every question, the passages retrieved and the answer. One named person reviews wrong answers weekly and fixes the source, not the prompt.

We apply the same rules in our own systems. In Sewing Lab the bot identifies the model from the ad a buyer came from, takes the price from the product card, and replies in seconds instead of 1.5 to 2 minutes, but only for the 108 of 122 products where the price is unambiguous; the rest go to a manager. In AI Brain we kept the knowledge base as plain Markdown rather than an opaque store, because when an agent acts on a wrong fact someone has to open the file, find the sentence and delete it.

Ready platform or custom AI chatbot

Ready chatbot platforms are a good start when your answers already live in a help center and the bot only needs to talk. Intercom's Fin, for example, charges per resolved conversation, $0.99 at the time of writing, with no build cost.

A custom chatbot pays off when:

• it needs live data: order status, stock, prices, bookings;
• it has to act: create a lead, update a deal, book a slot in your CRM;
• your handoff rules are specific (by product, language, order value, hour);
• per-resolution fees at your volume add up to more than owning the bot;
• the data cannot leave your infrastructure.

If you are still deciding, our comparison of automation platforms and custom builds walks through the same trade-off for workflows in general.

How much an AI chatbot trained on your own data costs in 2026

The market figures below come from published 2026 guides (Petronella Technology Group, March 2026). Our ranges are the same scope divided by four, in both money and time:

The lower price is not thinner work. AI writes the code and senior engineers direct and review it, so the same working chatbot takes fewer hours. Running costs are separate and third-party: model usage typically runs $100 to $2,000 a month depending on volume. If your bot grows into an assistant that takes actions across several systems, it becomes an agent, and AI agent development cost covers those ranges. Before giving it real traffic, run it through the checklist in why AI agent projects fail.

How we build AI chatbots at DForce

We start with the job, the knowledge audit and the test set of real questions, and only then write prompts. Code is written by an AI pipeline and directed by senior engineers who own retrieval quality, permissions and the handoff rules, which is why the same chatbot ships 4 times cheaper and 4 times faster than traditional hand-written development. You get the bot, its test set, the answer log and a clear rule for what it will never say.

If you want a chatbot that answers from your own documents and data, and knows when to stay quiet, book a discovery call and we will scope the first version with you.

Frequently asked questions

Can I train ChatGPT on my own company data? Not in the sense most people mean. You rarely need to retrain a model. A chatbot trained on your own data almost always uses retrieval (RAG): your documents are indexed, and for each question the bot fetches the few relevant passages and answers only from them, with a link to the source. It is cheaper than fine-tuning, updates the moment a document changes, and lets you see exactly which sentence an answer came from.

How do I stop an AI chatbot from making up answers? Restrict it to retrieved sources, tell it to say it does not know when nothing relevant is found, show the source with every answer, and hand the conversation to a person for anything involving money, refunds or exceptions. Then test it on 50 to 200 real past questions before launch and rerun that test on every change. In 2024 a Canadian tribunal held Air Canada liable for a refund policy its chatbot invented, so this is not optional.

How much does a custom AI chatbot cost in 2026? With an AI-first team, a chatbot trained on your documents for one channel costs $1,250 to $3,750 and takes 3.5 to 7 days, against $5,000 to $15,000 and 2 to 4 weeks with traditional development. A production chatbot connected to your CRM or helpdesk, with live data lookups and handoff to people, is $5,000 to $18,750 over 1.5 to 3 weeks, against $20,000 to $75,000 over 6 to 12 weeks. That is 4 times cheaper and 4 times faster, because AI writes the code and senior engineers direct it. Model usage is extra, typically $100 to $2,000 a month.

Should I use a ready chatbot platform or build a custom one? Use a ready platform when your answers live in a help center and the bot only needs to answer questions. Build a custom chatbot when it has to look up live data such as orders, stock or prices, act in your CRM, follow your own handoff rules, or when per-resolution fees at your volume cost more than owning the bot.