Lessons from a Physician-CIO on AI Governance with Dr. Stacey Johnston (Chief Information and Digital Execution Officer at Beacon Health System)

Dr. Stacey Johnston joined Beacon Health System in October 2024 as Chief Information and Digital Execution Officer after serving in CMIO and chief applications roles at Beaufort Memorial Hospital and Baptist Health in Jacksonville. She has maintained an active hospitalist practice throughout her technology career and frames her entire AI agenda around reducing EHR burden as a direct cause of physician burnout and lower patient care quality.

When Johnston arrived at Beacon, the organization had no formal IT governance process. Her first action was building a structure with an executive steering committee, eight advisory levels, multiple work groups, and a dedicated AI council. Beacon simultaneously developed two AI policies covering procurement and permitted versus prohibited uses. Prohibited uses include entering protected health information into tools such as ChatGPT or Open Evidence. Vendors must complete a disclosure form covering the LLM used, data modeling, drift and bias monitoring, and data storage. Beacon also built a mandatory AI literacy program for managers and above, and is evaluating monitoring solutions that layer on top of AI systems to detect drift and bias, functioning similarly to network security tools. Johnston requires a defined ROI for any AI initiative to receive advisory council approval; soft or qualitative ROI escalates to the C-suite steering committee, and no AI initiative has been approved without a defined ROI to date.

Beacon's ambient listening solution, an Oracle-built product Johnston describes as comparable to DAX 360 and Abridge, is used by approximately 70 to 80 percent of providers in ambulatory clinics, with roughly 70 to 80 percent of notes generated through it. Documentation time per note dropped from seven and a half minutes to two and a half minutes, and the solution generated an additional 10,000 dollars per physician in revenue over 12 months through better documentation capture. Clinicians trust the tool for documentation review but would not accept a fully autonomous mode in which they cannot read the note before finalization, illustrating Johnston's broader principle that trust in AI must be built incrementally, component by component.

Beacon's revenue cycle team is the organization's most advanced AI adopter, and early success in back-office workflows made it easier to gain acceptance for AI in referral management and then in clinical settings. The most concrete agentic AI case involves colon cancer screening: Beacon ordered 7,000 ColoGuard screenings through a fully autonomous agent, approximately 40 percent were returned, resulting in 250 additional colonoscopy screenings and one early-stage colon cancer caught, treated with resection, and resulting in full survival. Appropriate colon cancer screening also ties directly to four and five star metrics from Medicare Advantage plans, giving the initiative both clinical and financial rationale. A second agentic deployment addressed a scheduling crisis after Beacon acquired and migrated four hospitals onto its tech stack. With 100,000 appointments needing to be back-loaded in two weeks, Johnston estimated manual processing would have required hiring approximately 40 people; instead, Beacon built an agentic scheduling agent in approximately three weeks. Beacon is also moving toward fully autonomous benefits verification and evaluating agentic AI for call return workflows, while autonomous coding is currently operating at an augmented stage.

Johnston identifies inbasket message management as an underappreciated AI use case, noting that primary care physicians have received 85 percent more portal messages since COVID, representing uncompensated burden. She says AI implementation failures are primarily workflow, change management, and training problems rather than technology problems, and that deploying AI without meaningful value to clinicians is counterproductive. If AI pulls from wrong or inconsistently defined datasets, physicians will not trust the data and will not use the system.

Beacon formed a dedicated AI and transformational technologies department reporting directly to Johnston, with the team's sole focus on scanning the environment, meeting business partners, and evaluating new solutions rather than handling day-to-day maintenance. The organization uses a hybrid buy-or-build approach, with externally partnered agents for well-mapped workflows turning around in approximately two weeks and the full cycle from physician request to completed agent taking approximately eight weeks. Beacon is also developing an AI citizen program allowing staff to build low-code or no-code agents through a centralized approval process.

Johnston estimates full clinical AI adoption is one to two years away, contingent on change management, appropriate data, and trust. She predicts that over five to ten years, primary care physicians will increasingly see only sicker patients while lower-acuity cases are managed at home aided by AI and large language models, hospital inpatient acuity will rise, and future hospital rooms will use cameras for contactless continuous monitoring of vital signs. She does not expect AI to eliminate the caregiver workforce in the near term and says having an AI governance council in place before deploying any AI is the single most important first step any AI leader should take.