Fine tuning

LLMs, RAG & Orchestrationalso: fine-tuningalso: finetuning

In one sentence

Fine tuning continues training a model on your own examples so it behaves the way you want, changing the behavior of the model rather than its knowledge of your current facts.

Last reviewed 3 August 2026

Not to be confused with Retrieval-augmented generation.

Definition

Fine tuning trains a model on your own examples so it behaves the way you want. It changes how the model behaves, not what it currently knows.

It can make an agent sound like your business and keep answers short. It cannot keep your prices up to date. That is a job for retrieval.

The distinction between behavior and knowledge is the single most commercially important thing about fine tuning, and it is the one most often misunderstood. Fine tuning shapes behavior. It does not reliably teach current facts, and it cannot cite where an answer came from.

What fine tuning is genuinely good for

  • Style and tone. Making responses consistently match a brand voice.
  • Format adherence. Reliably producing output in a required structure.
  • Task specialization. Improving performance on a narrow, repeated task such as classification or query rewriting.
  • Brevity. Training a model to give consistently short answers, which is unusually valuable in voice.
  • Domain language. Improving handling of specialized terminology.

What fine tuning is bad for

  • Teaching current facts. Prices, policies and stock change, retraining every time is impractical, and the model will still confidently state stale information.
  • Anything requiring citation or auditability, since fine-tuned knowledge cannot be traced to a source.
  • Rapidly changing information of any kind.

The comparison with retrieval, the decision most teams face

  • Retrieval supplies knowledge at question time, updates instantly, and can be cited.
  • Fine tuning shapes behavior, requires retraining to change, and cannot be cited.
  • They are complementary rather than alternatives. A fine-tuned model that answers briefly and in brand voice, grounded by retrieval over current documents, is the combination that works.

The practical costs

  • Data preparation is the bulk of the effort. Examples must be numerous, consistent and representative.
  • Poor training data produces confidently wrong behavior that is harder to diagnose than a bad prompt.
  • Every base model update potentially requires redoing the work.
  • Small models are far cheaper to fine tune, which is where the technique is most economically sensible.

Lighter alternatives to try first

  • Better prompting, which is free and instantly reversible.
  • Few-shot examples in the prompt, which cost tokens and require no training.
  • Retrieval, for anything knowledge-related.
  • Fine tuning is the last resort after these, not the first move.

Common misconception

That fine tuning is how you teach a model about your business. For facts, retrieval is the correct mechanism. Fine tuning teaches the model how to behave, not what is currently true.

Why it matters commercially

Teams frequently pursue fine tuning to solve knowledge problems, spend substantially, and arrive at a system that is stale by design. Understanding the boundary between behavior and knowledge saves real money.

In voice specifically

Brevity is where fine tuning earns its place in voice specifically. A long spoken answer is punishing to sit through, whereas the same verbosity is skimmable in text, so training a model to answer consistently short pays off more in a voice agent than it would in a chat window.

Where AsqVox fits

True RAG over uploaded documents is the knowledge mechanism, precisely because business facts change and must be updatable by re-uploading rather than by retraining.

Visual

Behavior, not facts

Behavior, not factsBehaviorFine tuning is the answerBrand toneOutput formatBrevityNarrow task specializationDomain terminologyRequires retraining to changeCannot be citedFactsRetrieval is the answerPricesPoliciesStockOpening hoursProduct specificationsUpdates by re-uploading a documentCan be citedComplementary, not alternativesFine-tuned for brevity and brand voiceGrounded by retrieval over current documentsTry these first, fine tuning is the last resort1. Better prompting, free and instantly reversible2. Few-shot examples in the prompt3. Retrieval, for anything knowledge-related4. Fine tuning, last resort not first move

It changes how the model behaves. It does not know your prices changed.

Start with the router question: what are you trying to change, behavior or facts? Behavior goes left to fine tuning, facts go right to retrieval, and the two combine rather than compete. Before reaching for fine tuning at all, work up the ladder from prompting through few-shot examples to retrieval. And remember the standing warning: every base model update potentially means redoing the work.

Statistics

Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.

There is no reliable published benchmark quantifying fine tuning benefit for business voice applications. Results depend entirely on task, data quality and volume, so any specific improvement figure is dataset-specific.

-no reliable figure

A vendor quoting a single uplift percentage for fine tuning is quoting their own task on their own data. The honest position is that the benefit cannot be stated as one number.

Small models are substantially cheaper to fine tune than frontier models, which is why task-specific adaptation is most economically sensible at the small end.

Cheapest at the small endindustry range

Engineering convention across model providers, 2026 - A cost relationship rather than a measured figure. It is the reason narrow, high-volume tasks are the ones where fine tuning pays for itself.

Fine-tuned knowledge cannot be attributed to a source, which makes it unsuitable where grounding or auditability is required. This is a structural property, not an implementation limitation.

-industry range

Structural property of fine tuning, 2026 - No better implementation fixes this. Baked-in knowledge has no passage to point at, so it fails exactly where citation is the requirement.

Under GDPR, purpose limitation restricts repurposing data collected for one purpose. Using accumulated transcripts to fine tune a model is a distinct purpose from the one they were collected for, and requires a lawful basis covering it.

A distinct processing purposeindustry range

GDPR, purpose limitation principle, 2018 - Whether customer data is used for training is a standard procurement question. Transcripts gathered to run a call are not automatically available to train a model.

India DPDP Act 2023, consent-primary with Rules notified November 2025 and phased compliance by 13 May 2027, makes fine tuning on transcripts a consent scope question.

Compliance by 13 May 2027industry range

Digital Personal Data Protection Act 2023, Rules notified November 2025, 2023 - A consent-primary regime, so the question is whether the consent obtained covers training use, not only the original call.

All-in production voice cost lands around USD 0.11 to USD 0.33 per minute. Fine tuning is a one-off cost that may reduce per-inference cost if it permits a smaller model.

USD 0.11 to 0.33 per minuteindustry range

Industry-reported production voice cost range, 2026 - The recurring number is inference. Fine tuning is a capital cost that only pays back if it lets a cheaper model do the job.

Examples

In practice

A company fine tunes a model on its product catalog to improve accuracy. Three months later prices have changed and the model confidently states old ones, with no way to trace or correct the claim short of retraining. The catalog is moved to retrieval, where a re-upload updates the answers immediately, and the fine tuning is retained only for response brevity and tone.

The everyday version

Fine tuning is training the AI on your own examples so it behaves the way you want, sounding like your business and keeping answers short. What it is not is a way to teach it your prices. Prices change, and a trained-in price is baked in until you retrain. Your documents belong in the retrieval system, where updating them updates the answers.

Usage

Who says it

  • AI engineers and ML practitioners.
  • Buyers, frequently as a proposed solution to knowledge problems, which is the misuse worth correcting.

Where it turns up

  • In an RFP it sits next to customization, brand voice, data used for training, purpose limitation and data handling by the model provider.
  • Whether customer data is used for training is a standard and important procurement question.

Common misuse

  • Using it to teach current facts.
  • Attempting it before exhausting prompting and retrieval.
  • Overlooking that transcripts used for fine tuning constitute a separate processing purpose under data protection law.

Questions people ask

What is the difference between fine tuning and RAG?

Fine tuning changes how the model behaves by continuing its training on your examples, and it requires retraining to update. RAG leaves the model alone and supplies knowledge at question time from documents you can re-upload and cite. For behavior such as tone, format and brevity, fine tuning is the tool. For facts such as prices and policies, retrieval is the tool. They combine well and do not replace each other.

Can I fine tune a model to know my business facts?

You can, but you should not rely on it. Fine-tuned facts go stale the moment prices, policies or stock change, the model will still state the old ones confidently, and there is no way to trace or correct a claim short of retraining. Put facts in retrieval, where a re-upload updates the answer immediately and the source can be cited.

When is fine tuning actually worth it?

After prompting, few-shot examples and retrieval have been exhausted, and for behavior rather than knowledge. It earns its cost on narrow, high-volume tasks and on consistent style, format or brevity, and it is most economically sensible on small models where training is cheap. It is a last resort, not a first move.

Are there legal issues with fine tuning on call transcripts?

Yes. Using transcripts collected to run a call to instead train a model is a distinct processing purpose. Under GDPR purpose limitation it needs a lawful basis that covers training, and under India DPDP Act 2023 it becomes a consent scope question about whether the consent obtained extends to training use. Whether customer data is used for training is a standard procurement question for a reason.

Share this definition

Last reviewed 3 August 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.