Workflow refinement, not model performance, drives successful AI adoption

At ViVE 2026, Christus Health deputy CIO Stuart James described how his team begins evaluating AI tools by going to the “gemba,” a Japanese term that refers to the real place where work happens. The term comes from the lean management system, which emphasizes improving processes by observing work directly rather than relying on reports or assumptions. Before any code touches company workflows, James’s team maps what clinicians and staff actually do.

Research consistently supports this approach. A 2024 review in the Journal of Medical Artificial Intelligence, which synthesized findings from more than 40 peer-reviewed studies, concluded that workflow integration is “a challenge often overshadowed by a focus on AI model performance.” The review documented case after case in which strong models faltered because they were grafted onto workflows no one had examined.

How workflow gaps undermine AI adoption in healthcare

Other leaders at ViVE pointed to the same challenges in their own deployments. “It’s that legacy of meaningful use of IT,” said Michael Han, CMIO at MultiCare Health System. “We’re still addressing meaningful use here with AI now.” Health systems that digitized outdated workflows during the EHR rollout preserved their inefficiencies, and clinicians who struggled with paper processes often encountered the same issues in digital form. Instead of fixing the problem, the technology reproduced it, and the same dynamic is now repeating with AI.

Haider Warraich, a heart failure cardiologist and program manager at ARPA-H, noted that “AI is not like a new drug that you can just replace in the workflow. It’s not a new artificial hip that you just swap in and out. You’re changing how medicine works.” He pushed the point further, arguing that healthcare’s underlying AI problem is execution rather than knowledge. The therapies, the data, and the protocols already exist for most chronic conditions. What is missing is the operational capacity to match what clinicians know to be true with what they actually do at the bedside.

The clinical reasoning AI is meant to support “doesn’t start with a microphone,” argued Anil Jain, chief innovation officer at Innovaccer and a former Cleveland Clinic IT executive. Tools focused on capturing the patient encounter address documentation, but the harder problem sits upstream, in the handoffs and decision points that move data across an enterprise. That orchestration layer is where most organizations lack both visibility and governance, Jain said, and without it, AI tools end up operating without a clear understanding of the workflows around them.

Why workflow design matters in AI deployments

The orthopedics deployment at MedStar Health illustrates the difference between jamming AI into an existing process and redesigning the process itself to take advantage of new technology. The intake protocol that fed MedStar’s LLM-powered subspecialty matching began with 16 questions clinicians wanted patients to answer before booking. Jennifer McCraw, MedStar’s vice president of enterprise access and digital transformation, described the gap between what providers wanted to ask and what patients would tolerate as the central design problem. Her providers wanted to ask roughly a hundred questions across protocols, but patients didn’t want to answer any.

McCraw said the team cut the protocol to four questions while preserving clinical accuracy on subspecialty matching. The result was higher conversion to appointments and physician satisfaction with the appropriateness of patient routing. The workflow, not the AI model, is what changed to achieve this success.

This success shows up at scale, too. A 2025 McKinsey Global Survey of 1,491 organizations using AI tested 25 organizational attributes against bottom-line impact from generative AI. Workflow redesign had the largest effect of any variable the researchers tested, outranking executive oversight, dedicated adoption teams, and KPI tracking. Only 21 percent of organizations had fundamentally redesigned at least some workflows alongside their gen AI deployments. The rest were applying new tools to processes they had not examined.

What strong AI proposals have in common

Most of the projects Christus evaluates now move forward, James said, because those with weak workflow assumptions are filtered out before they reach the evaluation stage. Renown Health uses a similar filter. Steven Ramirez, the system’s chief information security and technology officer, said proposals stall at the phase-gate review when teams cannot articulate the problem they are trying to solve or define what success looks like. James described a parallel problem at Christus, where 80 to 90 percent of working sessions begin with no documented workflow at all. His team must often map the work from scratch before any automation conversation can begin.

Projects that get the green light look different. They begin with a clearly defined need, examine data infrastructure ahead of any model training, and map the workflows they intend to automate in detail. AI tools enter at the end of that process, not the beginning, after the work of understanding the workflow is already done.