60%+ of surveyed firms will feed AI knowledge via retrieval and prompts — resolves 2027-07-14

The debate: a company that wants an AI to know its own material can fetch the relevant pieces into the prompt each time (retrieval-augmented generation, or RAG), load it into a very long context window, or fine-tune the model so the knowledge is trained into its weights.

Why it matters: the two routes differ sharply in cost, in how fast they can be updated and in how easily they can be audited, which decides whether a company needs its own training team.

Who is on each side: SecondSource reads this as a division of labor, with knowledge going in through prompts and retrieval and behavior and format through fine-tuning, while the weights camp argues that writing knowledge straight into the model will erode that split.

SecondSource’s reading: by 14 July 2027, in a named annual survey of how enterprises customize AI (from the venture firm a16z or an equally identifiable study), 60% or more of enterprises will name retrieval or prompt injection, including long context, as their main way of feeding in knowledge. Our confidence is 50%.

The 60% line is an operating threshold that SecondSource set itself, not a figure from any existing survey. We have not checked the current state on the web; it will be read at the deadline.

arXiv (Lewis, Perez, Piktus et al., Facebook AI Research and University College London)
“We explore a general-purpose fine-tuning recipe…”

Read the knowledge-injection-method distribution in the most recent a16z survey, or an equally identifiable annual enterprise AI adoption study, published before the due date.

What would overturn it

A share below 60% flips the reading that knowledge mainly goes in through the context.