Advancing value creation in B2B market research – a human–AI collaboration framework for large language model

Purpose This study aims to examine how prompt design influences the quality of large language model (LLM)-generated business-to-business (B2B) market research outputs. It also proposes a switch-point governance mechanism for converting LLM-generated drafts into decision-ready market intelligence artefacts. Design/methodology/approach The study uses an exploratory pilot evaluation of three market-facing LLM-based tools: ChatGPT, Gemini and Perplexity AI. The tools are compared across three benchmark B2B market research tasks and three prompt types: basic, context-augmented and role-specific. In total, 81 outputs were generated. A role-specific benchmark subset of 27 outputs was then assessed using automated metrics and blinded human ratings. The automated metrics included relevance through cosine similarity, reference overlap through ROUGE-F1, cross-run semantic stability, latency, token length and readability through Flesch Reading Ease. Human evaluators rated the outputs on relevance, completeness, factuality, instruction-following and insightfulness. One-way ANOVA and non-parametric checks are reported as exploratory task-level tests, not as confirmatory evidence of model superiority. Findings Role-specific prompts produced the strongest descriptive scores in this pilot setting. Within the benchmark subset, Gemini showed the strongest observed profile, with higher average relevance, reference overlap and semantic stability, along with lower latency. ChatGPT produced relevant outputs but tended to generate longer and less readable drafts. Perplexity generated concise and readable summaries, but its outputs were less aligned with the reference brief and decision-facing requirements. Human ratings followed a similar descriptive pattern. Cross-run semantic stability showed directional alignment with higher decision-readiness ratings, offering preliminary support for P1. The proposed switch-point matrix illustrates P2 by mapping automated confidence signals and human judgement to four escalation actions: accept, judge, review/redo and rewrite. Research limitations/implications The study is an exploratory pilot case. The task set is small, and the task-level ANOVA is based on only three observations per tool. The findings may also vary across domains, model versions, prompt libraries and organisational contexts. Future research should expand the task bank, include more B2B sectors, use larger rater pools, compare additional model versions and test the switch-point thresholds in live market intelligence workflows. Practical implications The study provides a replicable workflow for B2B market intelligence teams. It shows how teams can standardise prompt design, screen LLM-generated outputs, combine automated checks with human evaluation and allocate decision rights through transparent escalation rules. Social implications The framework supports accountable use of LLM-based market research by reducing the risk that persuasive but weakly grounded claims enter managerial decision-making. It also highlights the need for human oversight; evidence checks and traceable governance in AI-supported knowledge work. Originality/value The study extends service-dominant logic to LLM-mediated B2B market research by treating LLM outputs as candidate value propositions rather than finished knowledge products. Its main contribution is the operationalisation of decision-rights governance through cross-run semantic stability and switch-point logic.