The central question
General-purpose AI models are useful, but they are not enough for every business. As AI moves deeper into workflows, companies will need systems that understand their data, terminology, constraints, and operating logic.
API-based AI solved the first adoption problem
Third-party AI APIs made experimentation fast. Companies could add language-model capabilities without building infrastructure. But convenience comes with limits once AI becomes central to operations.
Limits of generic APIs
- Limited control over model behavior and interpretation.
- Privacy and compliance concerns when sensitive data leaves the organization.
- Unpredictable costs as usage scales.
- Generic outputs that miss industry-specific terminology, regulation, or workflow context.
Customization has three main paths
Not every company needs to train a model from scratch. Most will choose between prompting, fine-tuning, LoRA, retrieval, or a combination. For deeper model adaptation, the three core training paths are fine-tuning, LoRA, and pre-training.
Fine-tuning adapts an existing model
Fine-tuning uses domain-specific examples to adjust a base model toward a task or industry. It is useful when the model needs deeper behavioral or domain alignment than prompting can provide.
Fine-tuning trade-offs
- Useful for legal, medical, financial, support, and other domain-specific assistants.
- Requires high-quality examples and careful evaluation.
- Costs less than pre-training but still needs model access and technical skill.
LoRA makes adaptation lighter
LoRA adds small trainable adapters to a base model instead of retraining the full model. It is attractive when teams need lower cost, faster iteration, or several specialized variants.
LoRA trade-offs
- Efficient for style, behavior, and narrower domain adaptation.
- Can support multiple adapters on one base model.
- May be less suitable when deep factual specialization is required.
Pre-training gives control, but at high cost
Training from scratch gives the most control over architecture, data, and ownership. It is also the most expensive and operationally demanding path, suitable only for organizations with major resources and a strong reason to own the full stack.
Starting points for custom AI
- Define the business use case and failure criteria.
- Decide whether retrieval, fine-tuning, LoRA, or pre-training fits the need.
- Collect, structure, and govern domain data.
- Plan compute, evaluation, monitoring, and ongoing improvement.
The practical point
Custom AI is not automatically better than a strong general model. It becomes valuable when control, privacy, domain accuracy, or workflow fit matter enough to justify the extra complexity.