The real question is not model size
Businesses often begin by asking which model they should use. That matters, but it is rarely the first decision. The better question is what business behavior the AI system must improve and how much company context it needs to do that reliably.
General-purpose AI is strong when the task is broad, exploratory, or low risk. A custom LLM system becomes valuable when the product must understand internal knowledge, follow company-specific rules, connect to tools, and produce outputs that can be evaluated.
When general-purpose AI is enough
- Drafting early ideas, summaries, and broad research.
- Supporting internal productivity where mistakes are easy to notice.
- Testing a workflow before investing in product-grade infrastructure.
When to build custom
- Your answers depend on proprietary documents, menus, policies, data, or workflows.
- The AI must cite sources, call tools, or take structured action.
- You need repeatable quality, cost controls, logging, and human review.
A practical path
Start with a narrow use case. Build retrieval, prompts, tool calls, and evaluations before assuming fine-tuning is required. Then measure what fails. The winning architecture is usually the one that solves the business problem with the least unnecessary complexity.