"We want an AI trained on our stuff" — that was the brief from a furniture retailer with a 4,000-line product catalogue and a decade of technical-support emails buried in a shared inbox. Customers and staff kept asking the same questions ("does the Callahan sofa come in walnut legs?", "what's the return window on custom orders?") and nobody wanted to dig through PDFs to answer them. He had read that "fine-tuning" was how you make an AI know your business, and wanted a quote for that.
He did not need fine-tuning, and by the end of the call neither he nor his catalogue had to go anywhere near a training run. What he needed was retrieval — hand the model the three or four right paragraphs at the moment someone asks, instead of trying to bake 4,000 products into the model's weights. Here is the distinction that actually decides which one your business needs, and what each one really costs.
RAG — retrieval-augmented generation — means the model never "learns" your content at all. Instead:
The model's knowledge of your business is entirely in what you hand it at query time. Nothing is trained. Update a price or add a product, and the next answer already reflects it — no retraining, just re-embedding the one page that changed.
Fine-tuning keeps training a model on your examples until its weights shift — the knowledge (or the style) becomes part of the model itself, not something you supply per question. It has a real use: teaching a model a consistent voice, an unusual output format, or a narrow skill across thousands of repetitions. It is a poor way to teach a model your product catalogue, because the catalogue changes and the model's weights don't, short of paying to retrain them again.
Here's the detail that catches people out: Anthropic doesn't offer self-serve fine-tuning for current Claude models at all. There is no "upload your data, get a custom model" button on the Claude API — training, uploading files, batching, and every other operation on the platform is documented, and fine-tuning simply isn't one of them. Anthropic's own position is that a well-built system prompt plus the right context gets you most of what people reach for fine-tuning to solve. The only path that exists is a bespoke enterprise arrangement, agreed directly with Anthropic, for large accounts — not something a business with a product catalogue and a support inbox can sign up for on a Tuesday.
If fine-tuning is genuinely what you need, that means a different model provider (OpenAI's API supports it directly) or a serious infrastructure commitment most small businesses have no reason to take on.
RAG, for a 4,000-product catalogue (roughly 200,000 words of content):
Fine-tuning, if you could do it here:
Your content changes — prices, stock, policies, new products. RAG updates the moment the source page does. Fine-tuning is frozen at whatever the training data said.
The model needs to answer from your specific facts. RAG's entire job is grounding answers in real content instead of the model's general training — including citing which page an answer came from, which fine-tuning cannot do.
You want a consistent voice or format, not new facts. This is genuinely fine-tuning's strength when a provider offers it — squeezing out a very particular tone across thousands of repeated outputs. It is rarely the actual problem a small business is trying to solve when they say "train it on our stuff."
Budget and timeline matter. RAG for a business our size can be running in a couple of weeks, for the cost of an afternoon of embeddings and a monthly API bill measured in tens of dollars. Fine-tuning, where it applies at all, starts with a negotiated enterprise contract most SMBs will never reach.
| Factor | RAG | Fine-tuning |
|---|---|---|
| How knowledge gets in | Retrieved at query time from your content | Baked into model weights during training |
| Available on Claude today | Yes — standard API | No self-serve option; enterprise-only |
| Updating content | Re-embed the changed page, done | Retrain, or live with stale answers |
| Typical monthly cost (10k queries) | ~$35–$70 in model calls, plus a few dollars of infra | No published small-business price — enterprise negotiation only |
| What it's actually good at | Answering from your specific, changing facts | Consistent tone or format across huge repetition |
| Setup time | Days to a couple of weeks | Data prep + training run, before any enterprise negotiation even starts |
Neither is a toy version of the other — they solve different problems, and for a small business with a catalogue, a knowledge base, or a support archive, only one of them is actually on the table.
You want customers or staff to get answers from content you already have → RAG. This describes almost every "AI search" or "smart FAQ" request we get, including this one.
You need a very particular writing voice or output format at large scale, and the facts don't change → fine-tuning is the right shape of tool, but check whether it's even offered for the model you want before pricing it out.
Not sure which one your project actually needs → that was true of the furniture retailer too, and it took one conversation to settle. We built him a RAG pipeline over his catalogue and support archive: Claude Haiku 4.5 answering, Voyage embeddings doing the retrieval, running comfortably under $50 a month before his first customer ever asked it a question. Tell us what you want your AI to know and we'll tell you honestly which side of this it falls on.
Tell us what you are building. We reply within 24 hours — no sales team, no follow-up sequence.
A hosted chatbot widget runs $30–$100 a month; one built on your own content costs $1,000–$3,000 once and pennies to run. Here is what each actually answers, what the running bill looks like, and when the honest answer is to skip it.
Gumroad takes 10% and Lemon Squeezy roughly 5%, both with no monthly fee; Sellfy and Podia flip that to a $29–$150 monthly charge with 0% cut. What you actually keep on a $79 sale, once a month and at $5,000 a month, decides which one is worth it.
A gym owner asked for a Next.js site because it looked modern on Twitter. What he actually needed was a $500–$7,500 WordPress site for the marketing pages, plus one thing WordPress could never do well.