Quick Buzz Feed
This article discusses why most AI tools fail to address the real complexities of professional workflows, particularly in healthcare. It highlights a decade-old paper mapping 191 distinct tasks performed by primary care physicians during patient visits, emphasizing the 'bird's-eye vs. mouse-eye' problem in workflow analysis and the critical need for granular understanding to build effective AI solutions.
This section begins by highlighting a key finding from a decade-old paper shared by Dr. Paulius Mui, which meticulously mapped the complete workflow of primary care physicians during patient visits. The researchers astonishingly identified 191 distinct tasks in a routine clinical encounter. This extensive study involved observing 30 primary care physicians across 17 diverse clinics in Wisconsin and Iowa, encompassing urban and rural settings, academic and community practices, and both electronic health record (EHR) and paper-based systems. Every task performed by physicians was recorded and systematically coded, resulting in a comprehensive list categorized into 12 major areas, 189 subtasks, and 191 total distinct tasks. These tasks ranged from fundamental actions like gathering chief complaints and performing physical exams to more granular, often overlooked details such as logging into the EHR, reviewing scratch paper notes, and managing interruptions like answering pages mid-visit. The author labels this disparity between perceived simplicity and actual complexity as the "bird's-eye vs. mouse-eye problem" in workflow analysis. The bird's-eye view, a high-level perspective (100 feet above the workflow), outlines major steps like patient arrival, vitals, physician entry, history, exam, plan, and departure. Conversely, the mouse-eye view provides an on-the-ground, detailed perspective, capturing every nuanced action: patient check-in, insurance card exchange, front desk verification, waiting times, medical assistant actions (e.g., placing blood pressure cuff, recording vitals on paper, EHR login, chart updates). The analogy of a hawk hunting a mouse illustrates how high-level observers miss the granular, often challenging, details a practitioner navigates daily. This fundamental misunderstanding leads to most tools being built from a bird's-eye perspective, missing the true intricacies of physicians' daily tasks. The original paper, released during the rollout of EHRs under the ACA's meaningful use requirements, aimed to equip clinics with a workflow mapping tool before significant changes, promoting intelligent planning over reactive chaos. Today, with almost all physicians using EHRs, countless AI companies are attempting to integrate solutions into these complex workflows, often doing so "blind." The article emphasizes that direct observation (shadowing) is the gold standard for understanding these workflows, as interviews often miss crucial nuances. For instance, the common setup of a computer facing a wall behind the patient forces physicians to constantly turn, impacting efficiency and bedside manner—a detail only visible in the room. The author recounts his own experience pre-medical school, observing front-desk operations to create granular process maps, which allowed for precise identification of bottlenecks and pain points, leading to solutions that address real problems rather than assumed ones. This detailed approach enables effective iteration, timing of tasks, and pinpointing the actual issues a solution must solve.
This section underscores the critical relevance of granular workflow understanding amidst the current rapid AI implementation wave across healthcare. Health systems are being inundated with pitches for various AI tools, including ambient documentation tools, clinical decision support systems, automated coding, AI scribes, and diagnostic assistants. A significant issue highlighted is that most of these tools are developed by individuals who have never directly observed a physician's work, thus designing from a "bird's-eye view." These tools often operate under faulty assumptions: that workflows are perfectly linear, that physicians have ample time to review AI-generated summaries, that EHR interfaces are always intuitive, and that physicians consistently work from a fixed desk rather than in dynamic environments like hallways, managing multiple tasks simultaneously (e.g., holding a phone and a faxed chart). For practices and health systems evaluating new AI solutions, the article provides essential questions to demand from vendors: First, inquire if they have mapped your specific, actual workflow; if they have shadowed your physicians, timed tasks, and accurately identified existing bottlenecks. Second, ask for a detailed explanation of how their tool integrates at the task level, specifying which of the 191 identified tasks it replaces, which new tasks it introduces, and where it might create new handoffs or require additional clicks. Third, vendors must explain contingency plans for when the tool malfunctions or disrupts the workflow, including fallback procedures, the amount of required training, and the time it takes a physician to correct or override an incorrect AI output. The article warns that if a vendor cannot provide satisfactory answers to these questions, they are operating from a superficial, "hawk's perspective," and implementing their tool will likely burden clinicians with more work. In conclusion, the article firmly states that clinical workflows are inherently complex and were never simple. Any entity developing tools for physicians without committing to thorough, ground-level research to understand these intricate workflows is inevitably setting both the users (physicians) and themselves up for failure, exacerbating existing inefficiencies rather than solving them.