Disclosure: Dante AI publishes this guide. Dante AI figures below come from our own pricing page and Help Center, read on 25 October 2026. We do not quote other vendors' prices here; check each vendor's own pricing page.
Training a support chatbot on your own data costs anywhere from $0 to a few hundred dollars a month on a hosted tool, and far more if you fine-tune or build it yourself. Most hosted tools do not bill the training itself: they cap how much content you can load and charge for the replies. The real cost sits in three places: the plan, the replies, and the time it takes to prepare and maintain your content.
In this article
- What does it cost to train a chatbot on your own data?
- What are the ways to train a chatbot on your data?
- How do the approaches compare on cost?
- How do hosted tools price training data?
- What are the hidden costs?
- What would a real business pay? A worked example
- How do you keep the bill down?
- Sources
What does it cost to train a chatbot on your own data?
It depends less on how much data you have than on how you train on it. In 2026 there are three common routes, and they put the cost in different places:
- Fine-tuning a model puts the cost up front, in training runs and engineering, and again every time your content changes.
- Building your own retrieval (RAG) stack puts the cost in infrastructure and engineers: storage, search, model calls and upkeep.
- A hosted chatbot tool puts the cost in a monthly plan with a content limit and a reply allowance, counted in credits or messages.
For a business that wants a support chatbot answering from its website, help articles and documents, the hosted route is usually the cheapest by a wide margin, because someone else has already built and maintains the retrieval part. On Dante AI, for example, adding and re-reading your content uses no credits at all: credits pay for your agent's replies, and each plan sets how much knowledge your agent can hold.
What are the ways to train a chatbot on your data?
Fine-tuning
Fine-tuning retrains a model on examples you supply. It is good at teaching a style or a format. It is a poor fit for facts that change, such as prices, opening hours or stock, because every change means preparing new examples and paying for another training run. You also need people who can prepare the training set and judge the result.
Retrieval (RAG)
Retrieval keeps your content outside the model. When a customer asks something, the system finds the most relevant passages in your content and gives them to the model to answer from. Updating the chatbot means updating the content, not retraining. This is how most support chatbots work today, and it is how Dante AI works: on every reply the agent retrieves the top-matching passages from your knowledge and answers from them, and says so when they do not cover the question. Our guide to training a chatbot on your own data walks through the steps.
Hosted tools
A hosted tool gives you retrieval as a product: you add a website, files or pasted text, and the tool handles crawling, splitting, indexing and answering. You pay a plan price, and the plan usually limits how much content you can load and how many replies you get.
How do the approaches compare on cost?
Where the money goes, by approach. Figures for Dante AI are from our pricing page; the other rows describe how each approach is typically paid for, not any one vendor.
| Approach | What you pay for | Up-front cost | When your content changes | Who maintains it | Best for |
|---|---|---|---|---|---|
| Fine-tune a model yourself | Training runs, model usage per reply, engineering time | High: data preparation and test runs before the first answer | Prepare new examples and run training again | Your engineers | A fixed style or format, rarely changing facts |
| Build your own RAG stack | Hosting, vector database, model usage per reply, engineering time | Medium to high: build the crawler, index, chat widget and handover | Re-index the changed content (you build the scheduler) | Your engineers | Teams with unusual data or strict in-house requirements |
| Hosted tool, priced by credits or messages | Monthly plan; replies counted against an allowance; content limit per plan | Low: usually a free tier or trial | Re-crawl or re-upload; some plans re-crawl on a schedule | The vendor (you keep the content current) | Most small and mid-size support teams |
| Dante AI | Flat monthly plan; credits pay for replies, not for adding content; knowledge capped in characters | $0: free plan, no card | Website re-read weekly on Starter, every 3 days on Advanced, daily on Pro; files, text and Q&A updated by hand | Dante AI (you keep the content current) | Business websites that want answers from their own content plus a human handover |
The table explains why "training data pricing" is a slightly misleading phrase for hosted tools: you rarely pay for the data or the training itself, you pay for the room to store it and for every answer the chatbot gives from it. Our chatbot pricing guide compares the full build paths, platform to custom.
How do hosted tools price training data?
Three numbers decide the bill on a hosted tool. Check all three on any vendor's own pricing page:
- The content limit. How much knowledge each plan holds. Vendors count it in characters, pages, documents or files, so convert to one unit before you compare.
- The reply allowance. How many answers you get a month, in messages or credits, and how many credits one reply costs on the model you pick.
- The refresh rule. Whether the tool re-reads your website on its own, how often, and on which plan.
How Dante AI counts it
| Plan | Price a month | Knowledge limit | Roughly | Credits a month | Website auto-update |
|---|---|---|---|---|---|
| Free | $0 | 150,000 characters | ~75 web pages | 100 | Not available |
| Starter | $40 | 25,000,000 characters | ~12,500 web pages | 2,000 | Weekly or monthly |
| Advanced | $120 | 50,000,000 characters | ~25,000 web pages | 8,000 | Every 3 days, weekly or monthly |
| Pro | $400 | 200,000,000 characters | ~100,000 web pages | 30,000 | Daily, every 3 days, weekly or monthly |
A few rules sit behind those numbers:
- Training is not metered. Credits meter your agent's replies. Adding a website, uploading files, pasting text and re-crawling do not use credits; they use room in your knowledge limit.
- One reply costs 1 to 25 credits, depending on the model. The default model costs 1 credit, standard models about 10, and premium models 25. Every model on the free plan costs 1.
- A page is stored once. If a page is already in your knowledge, adding it again does not count toward your characters.
- Extra credits cost $20 a month per 1,000. Yearly billing gives two months free.
Full detail: knowledge sources, understanding credits and plans and pricing in the Help Center.
See what your own content costs to run. Add your website or a few files on the free plan and watch the knowledge bar and the credit meter. No code, no card.
What are the hidden costs?
The plan price is the easy part. These are the costs that show up later:
- Content preparation. Someone has to find the right pages and files, remove outdated ones and write the answers that exist nowhere yet. This is people time, and it is usually the biggest single cost in the first month. A short FAQ document often does more than a hundred extra pages.
- Content that does not read cleanly. Pages rendered entirely by JavaScript, content behind a login, and scanned images with stylised text can come back thin. Plan time to upload the underlying document or paste the text instead.
- Re-crawls and upkeep. A chatbot that learned your prices in spring will quote spring prices in autumn. Scheduled website re-reads remove most of this work, but on Dante AI auto-update covers website sources only: files, pasted text and Q&A pairs still need updating by hand when they change.
- Testing. On Dante AI, sandbox replies spend the same credits as live ones, after your first 5 onboarding test messages. Budget a few hundred credits for a proper test round.
- The model you choose. The same allowance buys 25 times fewer replies on a premium model than on a 1-credit model. This is the most common reason a bill grows.
- Add-ons. Extra agents, branding removal and extra credits are priced separately on some plans. On Dante AI, Advanced and Pro already include human handover and branding removal.
Our chatbot ROI calculator shows how to weigh these costs against the time the chatbot saves.
What would a real business pay? A worked example
An invented business: a home-services company with a 60-page website, 10 PDF guides and price sheets, and about 600 chat conversations a month at 3 visitor messages each, so about 1,800 replies.
Step 1: size the knowledge. Using the Help Center's rough guide (about 2,000 characters per web page and 5,000 per short PDF): 60 pages x 2,000 = 120,000 characters, plus 10 PDFs x 5,000 = 50,000. Total: about 170,000 characters. That is over the free plan's 150,000 and a tiny fraction of Starter's 25,000,000.
Step 2: size the replies. 1,800 replies a month. On a 1-credit model that is 1,800 credits. On a standard model at about 10 credits a reply it is about 18,000 credits.
Step 3: pick the plan.
| Option | Fits the knowledge? | Fits 1,800 replies? | Website kept fresh | Monthly cost |
|---|---|---|---|---|
| Free | No (170K over 150K) | No (100 credits) | Manual | $0, too small |
| Starter, 1-credit model | Yes | Yes (2,000 credits) | Weekly or monthly | $40 |
| Advanced, 1-credit model | Yes | Yes (8,000 credits) | Every 3 days, weekly or monthly | $120 |
| Advanced, standard model | Yes | Yes, with 10 x 1,000 extra credits | Every 3 days, weekly or monthly | $120 + $200 = $320 |
| Pro, standard model | Yes | Yes (30,000 credits) | Daily available | $400 |
Step 4: add the people time. Count the hours to choose and clean the 60 pages and 10 PDFs, write 20 Q&A pairs for the questions the site does not answer, and check the PDFs each time a price changes. If prices change often, the extra $80 a month from Starter to Advanced buys website re-reads every 3 days instead of weekly, plus human handover, which usually costs less than the hours it saves.
The result: this business would pay $40 to $120 a month on a 1-credit model, and the model choice, not the amount of content, is what would move it toward $320 or $400. Our guide to training an AI chatbot on your documents shows the setup itself.
How do you keep the bill down?
- Start on a 1-credit model and move up only for agents that need more reasoning. On Dante AI the model is set per agent, and switching never changes what the agent has learned.
- Load less, better content. Remove outdated pages and duplicate files. A clean knowledge base answers better and stays inside a smaller limit.
- Turn your top 20 questions into Q&A pairs. They take priority over crawled pages for matching questions.
- Let the schedule do the re-reads if your website changes often, instead of re-crawling by hand.
- Use extra credits for a busy month, upgrade for a busy year. On Dante AI, 1,000 extra credits cost $20 a month; for sustained traffic, the next plan is usually better value.
- Pay yearly once you know your plan: it gives two months free.
Ready to see your own numbers? Start free at dante-ai.com with Get started, add your site, and the knowledge bar shows how much of your limit it uses.
Sources
- Dante AI: pricing page, Plans and pricing, Understanding credits, Knowledge sources and AI models (read 25 October 2026).
- Fine-tuning, RAG and hosted-tool rows describe how each approach is generally paid for. They make no claim about any named vendor's price; check each vendor's own pricing page.