Companies mostly fine-tune AI for tone and format; under 30% use it to add knowledge — resolves 2027-07-14

The debate: a company can teach a model its business by looking up relevant documents at question time and handing them over (retrieval), or by retraining the model on its data (fine-tuning).

Why it matters: the wrong choice wastes money, because fine-tuning knowledge in is the costliest way to do retrieval's job and must be redone whenever the facts go stale.

Who is on each side: one camp expects fine-tuning to take over knowledge injection; SecondSource reads the division of labor as intact, with knowledge going through retrieval and behavior through fine-tuning.

SecondSource’s reading: through 2027-07-14, in a named annual enterprise AI survey (a16z or equivalent), knowledge injection and fact updating will make up less than 30% of enterprise fine-tuning uses, which remain mostly about behavior, format, and style.

The 30% line is our own working threshold. It was not checked online when the call was written and will be verified at the due date. Our confidence is 0.45.

This is one half of our call that knowledge goes into the prompt and behavior into the model. The other half asks whether, in the same survey, at least 60% of companies use retrieval or prompting as their main way to feed in knowledge.

arXiv (Lewis/Perez/Piktus/.../Riedel/Kiela; FAIR/UCL)
“We explore a general-purpose fine-tuning recipe…”

Read the breakdown of enterprise fine-tuning use cases by sub-category in the same a16z, or equally identifiable, report.

What would overturn it

A knowledge-injection share of 30% or more blurs the line in the behavior-versus-knowledge division of labor.