RAG vs fine tuning: why AI writes an essay but gets your pricing wrong

The same model that writes a competent essay on monetary policy will quote the wrong price for your service. Not because it is weak. Because it read half the internet and not one line of your price list.

Below: what RAG does about that, when fine tuning is the right answer instead, and the place these projects usually fall over.

Where the difference comes from

A language model learned to predict which word follows which. It knows patterns from an enormous volume of text, which is exactly why the essay works: that task is the assembly of a probable sequence.

Your price list is not probable. It is arbitrary. It came out of a decision somebody made in a meeting in March, and it appears in no text the model has ever seen.

Asked for a price, the model does the only thing it can. It works out which number would sound credible in that sentence. You get something that looks like a price. It is not your price.

What RAG is, in one paragraph

Retrieval augmented generation adds a step before the answer. Before the model writes anything, the system searches your documents, pulls the passages that match the question, and hands them over as source material.

The model stops answering from memory and starts answering from what it was given. That is why a well-built agent can tell you which document an answer came from.

The practical difference: you change a price in a file and the agent quotes the new one the same day. Without that step you would have to retrain.

When fine tuning is the right call instead

This is a fair question and the answer is: sometimes, but almost never for a price list.

Fine tuning changes how a model speaks and reasons. It earns its cost when you need a specific voice, a fixed output format, or reasoning inside a narrow domain.

For facts it fails on three counts. It costs time and money on every change. It cannot cite a source. And you cannot pull a fact back out of it quickly once that fact stops being true.

The working rule: tune the style, retrieve the facts. Pricing, procedures, contract terms and customer data are all facts.

Where these projects actually fall over

Not on the model and not on the retrieval. On the documents.

An agent returns what it finds. If your price list exists in three versions, one of them attached to an email from last year, the agent will find all three and pick one. If the returns procedure is described differently in two places, you get both, in an order nobody chose.

This is the most common cause of failed deployments and the least often quoted line item. Proposals call it connecting the knowledge base. In practice it starts with deciding which file is the real one.

Tidying documents is frequently a bigger share of the project than building the agent. A proposal that says nothing about it will either skip that work or add it later.

What it costs

Three lines, one of them fixed.

Build: $2,500 to $8,000 for one process and one knowledge base, depending on how many sources there are and what state they are in.

Running: $100 to $400 a month for hosting, monitoring and keeping the base current.

Model usage: $15 to $80 a month in a small company. This line scales with how many questions get asked, not with headcount.

Plus one item that appears on no invoice: somebody on your side confirming which document is authoritative. Nobody external can do that, because only you know which version is true.

How to tell whether you are ready

Three questions, half an hour.

For your ten most common customer questions, does one authoritative document exist? Does anybody in the company know where it is? Can you say when it was last updated?

Three yeses means the build takes weeks. A single no means you start by tidying up, and that is good news, because the tidying helps your people too.

What happens when you get in touch

We start with the ten questions your customers ask most and check whether your files contain one unambiguous answer to each. That produces a number, not an opinion.

You talk to the owner, the call runs 45 minutes. If it turns out the problem is not a missing agent but disordered documents, we say so and point you at the cheaper route.

If we work together, somebody updates the knowledge base when your pricing changes, and somebody checks in month three whether the answers still hold. Your code and your data stay with you. The warranty covers repairs at no extra charge.