In Simple Terms

A foundation model out of the box (GPT-4o, Claude Sonnet, Gemini, Llama) is a generalist. It has seen the public internet and knows about almost everything. What it has not seen is your specific business, your products, your industry's jargon, your customers' typical situations, or the tone your brand uses. For many tasks, the generalist is good enough. For some, it is not.

A custom model closes that gap. Take the generalist model, continue training it on your business's own data (transcripts of past sales calls, examples of past customer support answers, your own product descriptions), and the result is a specialised version that talks like your business does. Same underlying capability. Better aim at your specific use case.

For most SMBs, the question of whether to build a custom model usually answers itself: when the off-the-shelf model keeps making the same kinds of mistakes for your business, despite good prompting, fine-tuning is the next lever. When the off-the-shelf model is fine, do not bother. Building a custom model adds complexity (training data preparation, retraining schedule, evaluation) that the team has to maintain.

Three Ways an SMB Customises a Model

When Custom Models Are Worth Building (and When They Are Not)

Use a custom model when:

The off-the-shelf model gets the task consistently wrong, despite reasonable prompting. Examples: a customer support bot whose answers keep contradicting your actual policies. A classifier that miscategorises your documents because the categories are specific to your industry. A scoring system that keeps weighting the wrong signals.

You have enough labelled training data that the model can actually learn the pattern. Rule of thumb: 200 to 500 examples for a fine-tune to do anything useful, 1,000 to 5,000 for noticeable lift on most tasks. Below 100, prompt engineering will usually beat fine-tuning.

The cost of the work the model does justifies the maintenance burden. A custom model needs a refresh schedule, a way to measure whether it is still performing, and a fallback for when it stops working. If the task is low-stakes, prompt engineering plus an off-the-shelf model is usually the right answer.

Do not build a custom model when:

The off-the-shelf model works well enough. The marginal lift from fine-tuning rarely justifies the work if the baseline is already good.

Your data is too sparse or too dirty. A fine-tune on bad data produces a worse model than no fine-tune at all. Cleaning your data first is the higher-leverage activity.

The use case is something a foundation model with RAG can solve. Retrieval-augmented generation lets the off-the-shelf model use your data at query time without permanent training. Cheaper, easier to update, and usually the right first answer for "let the AI use our company knowledge."