Matt Willsmore
Why enterprises are taking a fresh look at customizing large language models
I’ve been asked a lot lately whether fine-tuning large language models (LLMs) still matters. For a while, most people decided it was too expensive, too complex, or unnecessary given what modern foundation models can already do out-of-the-box.
But the landscape is changing fast. New, more efficient techniques—and the simple reality that 95 percent of GenAI projects fail to deliver measurable business value—mean it’s worth revisiting the idea.
My short video below, recently presented to our Special Interest Group, examines the issue.
Why Fine-Tuning Deserves a Second Look
A recent MIT study found that 95% of GenAI projects fail to deliver sustained ROI. The 5% that succeed do three things well:
- They focus on clear, measurable business problems.
- They collaborate with experienced partners who understand both AI and domain context.
- They adapt models to enterprise workflows — ensuring technology fits how people actually work.
Fine-tuning helps bridge that gap. By adjusting a model with your own data, acronyms, and tone, you can reduce hallucinations, improve accuracy, and lower the cost of complex prompts.
The New Era of Fine-Tuning
Modern techniques like parameter-efficient fine-tuning (PEFT) — including LoRA, adapters, and prompt-prefix tuning — let teams customize models using only a few hundred examples. Cloud platforms such as OpenAI’s fine-tuning API, Google Vertex AI, and new players like Thinking Machines’ Tinker now make this process accessible without deep ML expertise.
Fine-tuning won’t replace Retrieval-Augmented Generation (RAG) or prompt engineering, but it complements them. For regulated, jargon-heavy, or data-rich organizations, it’s becoming an essential way to get AI that truly understands your business.
If you would like to explore this further, please CONTACT US for a free consultation. You may also find some of the resources below useful.