Abstract
<p>Large language models (LLMs) are rapidly entering qualitative analysis workflows, yet the field's discussion of their integration has unfolded as though "qualitative research" were a single tradition with a single epistemological stake. Drawing on a review of fifty-one studies on LLM-assisted qualitative analysis published between 2024 and 2026, we observe that approximately 80% do not articulate an epistemological stance, while those that do rarely work within their stated paradigm as a constraint on methodological decisions. The result is a debate that conflates important issues: interpretivist scholars mount forceful objections to LLM use that may not extend to applied qualitative work, while pragmatic integrators describe sophisticated workflows without consistently operationalizing the standards by which their tradition assesses analytic adequacy. In this article, we argue that the integration question is methodology-specific and must be answered on the terms each tradition sets for itself. We then offer a worked answer for one pathway through the qualitative research landscape: a pragmatic narrative inquiry combining Framework Analysis with Codebook Thematic Analysis, drawn from the CHAT project, a PCORI-funded study of adolescent substance use involving 86 semi-structured interviews and 5,466 coded excerpts analyzed through Muse QDA, a purpose-built LLM-assisted qualitative analysis platform. Our framework rests on three principles: epistemic capacity, which specifies what each agent (human, LLM) is positioned to claim at each phase of analysis; analytic significance, which distinguishes calibration interventions (instrument-tuning) from meaning-making interventions (analysis-shaping) and locates the latter at tradition-specific decision points; and trajectory documentation, which preserves the path of analytic judgment without claiming reproducibility of outputs. The framework's distinctive theoretical contribution is a strict form of complementarity: human and LLM epistemic capacities are non-substitutable within this workflow, such that neither agent can do the other's work without epistemic loss. We illustrate the framework through a three-track codebook development process, a documented case of conceptual sharpening between two parent codes, and a "What did I miss?" exercise in which LLM coverage prompted human meaning-making. We close by considering the framework's limits, including resource demands, tool dependence, and the question of who gets to do LLM-assisted qualitative research, and by arguing that QDA platform design in the LLM era is itself a methodological matter. The integration question is not whether LLMs can do qualitative research, but for which tradition, under which paradigmatic commitments, do which kinds of integration satisfy that tradition's own standards.</p>