AI-Powered Workflow Design for Purchase Ledger Data Preparation – Knowledge Transfer and Value Co-Creation through Human-AI Collaboration

Sammanfattning

Introduction: The large-scale collection of public sector spending data by aggregators often yields records rife with severe semantic ambiguity, inconsistent formatting, and missing identifiers. Transforming this heterogeneous data into reliable business assets is a knowledge-intensive process that relies heavily on the tacit judgment of data specialists. Although artificial intelligence (AI) in data curation, human-in-the-loop (HITL) oversight, and knowledge transfer have been studied independently, there remains a gap in understanding how an integrated artefact can jointly support the preparation of highly ambiguous, knowledge-embedded financial data. Research Question: This thesis investigates how an AI-powered workflow can be designed to support knowledge transfer and value co-creation in data preparation. Specifically, it addresses two sub-research questions: (Sub-RQ1) how data specialists’ tacit knowledge can be effectively externalized into cognitive AI nodes, and (Sub-RQ2) what specific dimensions of value are co-created through human-AI collaboration relative to manual preparation baselines. Method: Applying a Design Science Research (DSR) framework merged with Action Research (AR), this study was conducted in collaboration with Tendium AB, a Swedish data aggregator. Using the Framework Method, internal documents and raw datasets were analyzed to diagnose workflow bottlenecks and elicit system requirements. An AI-powered workflow artefact was subsequently designed and instantiated within the n8n orchestration platform. This artefact effectively combined deterministic processing, AI-enhanced contextual reasoning, and a defensive human-in-the-loop (HITL) governance gateway. The artefact was evaluated through a controlled laboratory experiment utilizing 20 stratified historical datasets from public sector entities in Sweden and Finland. Results: The implementation of the artefact significantly improved processing efficiency, reducing the average processing time per dataset from 2 minutes 41 seconds to 2 minutes 18 seconds. The more profound impact was that the human active attention time reduced to zero for 19 out of 20 datasets. Furthermore, the system reached 100\% accuracy in routing structurally corrupt datasets to manual review and successfully resolved complex, multi-row accounting anomalies that typically resist rule-based automation. The study identified three core dimensions of co-created value: cognitive offloading, organizational knowledge retention, and risk mitigation using human-AI agency distribution. Discussion: The findings show that tacit data preparation expertise can be effectively codified into cognitive workflow nodes when safeguarded by a defined HITL boundary. By connecting previously isolated research streams, this study delivers a domain-specific architectural pattern and extends value co-creation theory into company business operations. Limitations of the study include a small evaluation sample size, a single-case context, and a restriction to structured file formats. Future research needs to explore broader generalization, the integration of additional file types, and the longitudinal effects on organizational learning.