Generative AI in Professional Workflows: Adoption Patterns and Reliability Perception

This thesis analyzes the structural transformation of professional workflows driven by the integration of Generative Artificial Intelligence (GenAI), specifically examining the critical divergence between linguistic fluency and factual reliability. Utilizing the theoretical framework of an Extended Technology Acceptance Model (TAM), the research investigates how Perceived AI Risk (PAIR) and Perceived Trust (PTR) mediate the adoption of large language models within the innovation and marketing sectors. The methodology is based on a qualitative A/B stress test comprising 40 experimental instances, comparing a "Naive" control group using zero-shot natural language queries with an "Expert" treatment group utilizing complex prompt architectures characterized by persona adoption and explicit negative constraints. The tests were conducted across three high-stakes professional scenarios: legal compliance (GDPR), financial data extraction (Tesla Q3 2024 Earnings), and academic literature synthesis. Performance was evaluated through a Reliability Evaluation Matrix (REM), specifically quantifying the "Verification Tax"—defined as the human labor time required to audit and validate probabilistic outputs against deterministic ground truth. Findings reveal a significant "Efficiency-Accuracy Trade-off." While expert prompting strategies increased global factual accuracy to 95%, they simultaneously precipitated a 66% increase in the average verification tax. This increase is attributed to "instruction-induced friction," where the human operator is forced to shift cognitive resources from primary factual audit to the verification of secondary stylistic constraints. Qualitative forensic analysis identifies the "Illusion of Fluency" and "Knowledge Overshadowing" as the most critical psychological hazards, demonstrating that polished, authoritative formatting can effectively neutralize human skepticism even when the output is factually hollow. In conclusion, the research defines the "Prompting Paradox," establishing that in deterministic, low-variance tasks, advanced prompt engineering acts as an operational liability that destroys verified velocity. The study proposes a strategic shift from "Input Optimization" to "Audit Specialization," arguing that professional competence in the generative economy resides in the strategic recognition of risk. Ultimately, the human role is redefined from a primary content creator to the final guarantor of institutional trust.