What a general-purpose model with good context actually covers
A well-engineered retrieval layer plus a strong general-purpose model handles the overwhelming majority of enterprise tasks — because the gap is almost always relevant context and tool access, not the model's base capability. Fine-tuning fixes a narrower problem: consistent tone, format, or domain vocabulary, not missing knowledge.
Reaching for fine-tuning to solve a context problem produces a model that's confidently wrong in a new, harder-to-debug way, because the knowledge gap is now baked in rather than retrievable and auditable.
When building or fine-tuning genuinely earns its cost
The honest triggers are: a proven, sustained task-specific need that off-the-shelf models and prompting can't meet after real effort, a data or latency constraint that rules out an external API, or a regulatory requirement for full model control. Absent all three, the ongoing cost of owning training infrastructure, evaluation, and retraining rarely pays back.
Owning a model is a cost centre with a payback period. Most teams reach for it before checking whether the payback exists.MNB AI Practice
Key actions
- Check whether better retrieval and context solve the problem first
- Reserve fine-tuning for tone/format/vocabulary, not missing knowledge
- Require a proven, sustained need before committing to training infrastructure
- Budget for ongoing evaluation and retraining, not just the initial build