Generic AI tools are impressive right up until a customer asks about your refund window, your onboarding steps, or page 47 of your product manual. At that point, an answer that sounds confident but is not grounded in your actual documents is worse than no answer at all. The fix is to train an AI agent on the files you already have, so every reply comes from your PDFs, help articles, and policies instead of the open internet.
This guide walks through why document-grounded answers beat generic ones, which files work best, and the exact steps to go from a folder of PDFs to a working AI agent on your website. You can do the whole thing on a free plan with no credit card required.
Why Training on Your Own Documents Beats Generic AI
A general-purpose AI model is trained on the public internet. It can write a poem or summarize a news story, but it has never read your warranty terms, your integration guide, or the spec sheet for the product you launched last month. Ask it a question about your business and it will either admit it does not know or, worse, produce a plausible-sounding guess.
An AI agent trained on your documents works differently. When a visitor asks a question, the agent first retrieves the most relevant passages from the content you uploaded, then writes its answer from those passages. This approach is known as retrieval augmented generation, and you can read a plain-English breakdown in our guide to RAG-powered AI agents.
The practical benefits show up fast:
- Accuracy. Answers come from your files, so the refund window it quotes is your refund window.
- Coverage of long-tail questions. The obscure question buried on page 47 gets answered as reliably as your top FAQ.
- On-brand responses. The agent reflects your terminology, product names, and policies rather than generic industry language.
- A single source of truth. When your documents change, the agent's knowledge changes with them. No retraining project, no copy drift.
What Documents Work Best for Training
Almost any written material you already maintain can become training data. The sources that consistently perform well:
- PDFs. Product manuals, spec sheets, onboarding guides, brochures, and whitepapers. If your knowledge lives in one big manual, our chat with PDF guide shows how much you can get from a single well-structured file.
- Word documents and text files. Standard operating procedures, sales one-pagers, internal playbooks, and process docs.
- Help center articles. Usually your best source, because they are already written as answers to real customer questions.
- Policies. Returns, shipping, warranty, privacy, and terms of service. These generate a huge share of support questions, and they are exactly where a wrong guess is most costly.
- Your website itself. You do not have to upload anything to start. Paste your site URL and the agent trains on your published pages, then layer documents on top for depth.
A useful rule of thumb: if a new support hire would need to read it in week one, upload it.
How to Train an AI Chatbot on Your Documents, Step by Step
Here is the full process with Dante, from empty account to a live AI agent on your site.
Step 1: Create a free account
Sign up at www.dante-ai.com. The free plan costs $0, needs no credit card, and includes one AI agent with up to 700 onboarding credits, then 100 credits per month. One credit equals one message, so you can run a meaningful pilot before paying anything.
Step 2: Upload your documents or paste your website URL
Add your PDFs, docs, and policy files, or simply paste your website address and let the agent ingest your published content. Most teams do both: the website provides broad coverage, and the uploaded documents provide depth on products, policies, and procedures.
Step 3: Let the agent train
There is no configuration screen to fight with. Dante processes your sources automatically, and you have a working AI agent in about 60 seconds. Larger document sets take a little longer to fully index, but you can start testing almost immediately.
Step 4: Test in the sandbox with real customer questions
Before anything goes on your website, ask the agent the questions your customers actually send. Pull ten or twenty real tickets from your support inbox and run them through the sandbox. Check that answers are correct, check the tone, and note any gaps. Gaps almost always mean a missing document, not a broken agent, so upload the missing source and test again.
Step 5: Embed with one script tag
When you are happy with the answers, add the agent to your site with a single script tag. It works on plain HTML, Webflow, WordPress, and Squarespace. From there you can connect Shopify, HubSpot, Zendesk, and more than 9,000 other apps through Zapier, and built-in human handover passes the full conversation context to your team whenever a person needs to step in.
How Automatic Re-Crawl Keeps Knowledge Fresh
Documents go stale. Prices change, policies get revised, and features ship. An AI agent that answers from last quarter's PDF is quietly wrong in a way nobody notices until a customer does.
Dante handles this with automatic re-crawling of your website content. On the Advanced plan your knowledge refreshes weekly, and on the Pro plan it refreshes daily, which matters when your site changes often. On any plan, you can upload a replacement document the moment something changes, and the agent answers from the new version right away. Plan details are on the pricing page, and every tier is a flat monthly fee with no per-resolution charges.
Quality Tips That Make Answers Noticeably Better
Training an AI agent is easy. Training a great one takes a little editorial discipline:
- Curate rather than dump. Ten authoritative documents beat a hundred random ones. Every file you upload is a source the agent may quote, so treat the upload list like a reading list for a new employee.
- Remove stale files. When a policy is updated, delete the old version instead of leaving both. Two conflicting sources force the agent to pick one, and it may pick the wrong one.
- Test with real questions, not demo questions. Your phrasing is not your customers' phrasing. Real tickets surface real gaps.
- Pick the model tier that fits. GPT-5.6 tiers Sol, Terra, and Luna are available from the Starter plan up, so you can match response quality to how demanding your use case is.
- Keep an escalation path. Configure human handover for the conversations that deserve a person, like billing disputes or cancellations. The agent handles volume, and your team handles judgment.
Common Mistakes to Avoid
- Uploading every file you own. Old proposals, outdated price lists, and draft policies all become potential answer sources. Noise in means noise out.
- Contradictory versions of the same document. The most common cause of inconsistent answers is two files disagreeing with each other.
- Skipping the sandbox. Embedding an untested agent on your homepage is the fastest way to lose trust in the whole approach. Twenty minutes of testing prevents weeks of doubt.
- Treating training as one-and-done. The launch is the start. Review conversations, spot unanswered questions, and feed the missing documents back in.
- No human handover. Even a well-trained agent should have a graceful exit to your team for the conversations that need one.
Frequently Asked Questions
Can I train an AI chatbot on PDF files?
Yes. PDFs are one of the most common training sources. Upload product manuals, guides, spec sheets, or policy PDFs and the AI agent answers questions from their contents, alongside anything it learned from your website.
How do I train an AI chatbot on my own documents for free?
Create a free Dante account, which costs $0 and needs no credit card. Upload your documents or paste your website URL, and you get one AI agent with up to 700 onboarding credits, then 100 credits per month, where one credit equals one message.
How long does it take to train an AI chatbot on documents?
With Dante you have a working AI agent in about 60 seconds. Large document libraries take a little longer to fully index, but you can start testing in the sandbox almost immediately after uploading.
Does an AI chatbot update automatically when my documents change?
Website content re-crawls automatically, weekly on the Advanced plan and daily on the Pro plan. Uploaded files can be replaced at any time on any plan, and the AI agent answers from the new version straight away.
What file types can I use to train an AI chatbot?
PDFs, Word documents, text files, help articles, and your website content all work. Paste a site URL for broad coverage, then upload documents for depth on products, policies, and procedures.
Your documents already contain the answers. A trained AI agent just makes them available at the moment a customer asks, at any hour, without a queue. Start free and have your files answering questions in about 60 seconds.