How much data do you need to build an AI model?
Short answer
Key takeaways
- The "you need big data" belief comes from training giant models from scratch — a task almost no business actually does.
- A classical ML model for a focused task often trains well on a few hundred to a few thousand good, labelled examples.
- Fine-tuning a pre-trained model or using retrieval over your documents delivers value with far less data than training from scratch.
- Data quality, labelling, and representativeness matter more than raw volume — a small clean dataset beats a large messy one.
The "we don't have enough data" objection almost always comes from a misconception: that every AI project trains a massive model from the ground up. That's the one scenario where enormous datasets are genuinely required — and it's a scenario virtually no business is in. The frontier models were pre-trained by their makers on internet-scale data already; your project builds on top of that, or uses smaller purpose-built models, neither of which needs a warehouse of your own history to start working.
For a focused, well-defined task, classical machine learning is often the right tool, and it's remarkably data-efficient. A model that classifies documents, scores a lead, flags an anomaly, or forecasts next month's demand can frequently be trained on a few hundred to a few thousand good examples — the kind of history most established businesses already have sitting in a spreadsheet or database. Choosing the right technique for the problem, rather than defaulting to the largest possible model, is exactly what keeps the data requirement modest.
Several techniques stretch small data further. Fine-tuning adapts an already-capable pre-trained model to your specific task with a comparatively tiny set of your examples. Retrieval-augmented generation (RAG) sidesteps training on your data almost entirely — it looks up the relevant passages from your documents at answer time, so a knowledge assistant over your policies or manuals can work with no model training at all. And for some pattern and vision tasks, pre-trained models handle your case out of the box, with your data used only to check and tune, not to teach from zero.
The real lever is quality, not quantity. A few hundred clean, correctly labelled, representative examples will outperform a huge but messy or biased pile every time — and messy data is where projects actually stall, because the pipeline work to clean it is the hidden cost. The right first move is rarely "collect more data"; it's to scope honestly what you already have, decide the leanest technique that fits, and start there. If the data genuinely isn't ready, a good partner will say so plainly rather than train a model destined to underperform.
Still have a question? Ask a human.
Tell us what you're trying to figure out and we'll give you a straight answer.