We examine the Intermediate Knowledge Problem (IKP): the challenge of maintaining coherent, goal-aligned intermediate representations during multi-step reasoning. The issue is particularly acute in automated knowledge elicitation and engineering, where analysis relies on provisional and revisable artifacts. Using Grounded Theory as a running example, we illustrate why existing symbolic systems and large language model based pipelines provide limited support for such intermediate knowledge, identifying a central obstacle to automating knowledge elicitation and engineering methods.