AI Augmented Structured Decision-Making : Integrating Large Language Models into Delphi and AHP and Evaluating User Experiences

Sammanfattning

Decision-making in organizations is persistently challenged by cognitivebiases, group dynamics, and the complexity of structured methodologiessuch as the Analytic Hierarchy Process (AHP) and the Delphi method.While these frameworks offer proven scaffolding against poor decisionbehaviour, their adoption in time-constrained business environmentsremains low due to steep learning curves and time constraints. This study was made possible with the help of a case company facing these challenges. It's a large multinational technology company, for whichan LLM-augmented Delphi and AHP decision-making workflow was developed and deployed, raising the question of whether integrating AI improves ordegrades users' perceived ability to make structured decisions. With the case company as the starting point, this thesis examines howLLM augmentation shapes user experience within structureddecision-making workflows, what AI behaviours and features drive thatexperience, and what risks emerge that could harm decision outcomes. A conversational AI augmented tool was developed specifically for the study and integrated within an existing decision platform.Through qualitative methods, including seven semi-structured interviewswith experienced decision-makers conducted after an independent testingof the tool, the study analyses user perspectives. Using an abductiveapproach informed by the Gioia methodology, progressing fromfirst-order concepts to three aggregated dimensions: Decision making, Structured decision processes and AI augmentation. The study revealed that LLM augmentation is broadly perceived asbeneficial, translating intuitive decision behaviours into explicit,systematic processes and enabling users with no prior methodologicalknowledge to complete full AHP and Delphi workflows. Five AI featureswere found to shape user experience: dialectical engagement, automatedgeneration of criteria and alternatives, simulated expert personas,consistency ratio guardrails, and guided process structuring. However,the study also identifies a critical expertise paradox, where users whomost need AI assistance are least equipped to critically evaluate itsoutputs, creating conditions for automation bias and algorithmicopacity. A clear consensus emerged that human accountability cannot bedelegated to algorithms, particularly for decisions involving personnel. The study highlights the importance for organizations of treatingAI-augmented decision tools as complements within a decision process,with final decisions made outside the tool. Practically, organizations should first introduce the decision process with higher perceived ease of use so users can benefit from its greater initial accessibility. They should also integrate AI early in the decision process rather than using it only at a late stage or retrospectively. Finally, personnel-related decisions should be excluded because of confidentiality requirements and the moral importance of human accountability. For researchers, the findings suggest that the identified expertise paradox is not adequately captured by existingtheoretical frameworks. Within the Technology Acceptance Model and itsextensions, perceived usefulness is never treated as a potentialliability -- the possibility that a tool might feel helpful whilesimultaneously degrading the quality of information underlying adecision goes unaddressed. Similarly, Human-in-the-Loop frameworksassume that keeping a human in final authority is sufficient topreserve accountability. The expertise paradox challenges thisassumption: if the human making the final call lacks the knowledge toevaluate what the AI has produced, formal oversight does not translateinto meaningful control.