Legislators hear promises and pitfalls of AI in health care, from documentation to diagnostics

Tanya Malik, a lawyer and telehealth entrepreneur, and Michelle Hager, managing partner of BlueSears, told members of the Vermont Senate Health and Welfare committee on March 19 that artificial intelligence is already embedded in many health-care tasks — from ambient documentation and image analysis to revenue-cycle automation — and that the state should weigh benefits against privacy, bias and access risks.

The presentation covered clinical uses now in practice, workforce and patient benefits, and legal and policy questions for state lawmakers. "I am embracing and think the industry should embrace artificial intelligence in health care while balancing privacy and safety concerns and maintaining some control," Malik said, summarizing the view she presented to the committee.

Why it matters: Committee members repeatedly returned to two constraints that could limit AI's value for Vermont residents — data privacy and rural access. Witnesses said AI can reduce documentation time, help detect disease in images and speed claims processing, but many tools require broadband, large datasets and explainable algorithms; without those, new systems risk replicating bias or creating coverage denials driven by cost models rather than clinical judgment.

Key takeaways from the presentations and committee discussion

- Current clinical uses and workforce effects: Malik and Hager described "shadow" and ambient AI that listens to clinician-patient visits and drafts notes, tools that flag possible findings on images and algorithms that predict readmission risk. Malik said tools can shift routine tasks away from clinicians, allowing them to “operate at the top of their license,” and gave a community mental-health example in which a documentation tool cost about $75,000 but yielded measurable staff retention benefits for the purchaser.

- Digital twins and precision care: Malik described "digital twin" models — virtual replicas of a patient or population updated with real-time data — that can be used for precision treatment planning and employer-level population health. She said employers receive deidentified aggregated results to preserve privacy.

- Revenue-cycle automation and claims: Hager described AI that "cleans" claims, performs straight-through processing and flags likely denials or fraud. She said those tools can reduce accounts receivable days and speed appeals, but also highlighted pending litigation alleging insurers used predictive tools to deny post-acute care; Hager cited ongoing cases against UnitedHealthcare and Humana brought by plaintiffs who say algorithms curtailed medically necessary services.

- Explainability and oversight: Hager invoked guidance from the American Medical Association and the Office of the National Coordinator (ONC), saying AI tools should be "transparent, explainable, and evidence based" and that clinicians must participate in testing and validation before live use.

- Bias, privacy and deepfakes: Speakers warned that training datasets skewed toward lighter skin tones or better-resourced patients can produce biased results in dermatology and other fields. Malik and Hager flagged privacy risks tied to ambient recording and to data aggregation, and Malik demonstrated a behavioral-health transcription product that integrates notes into an electronic medical record if the practice chooses.

- Equity and rural access: Committee members asked how AI can help Vermont's rural communities. Malik said the technology can extend care where telehealth is established, but stressed that broadband and local telehealth capacity remain prerequisites to realizing those benefits.

What legislators asked and what remains unresolved

Committee members pressed on where data are stored, who bears liability for AI-driven decisions, and whether patients should be notified if a chatbot or automated process is used. Woody, a committee questioner, asked whether models and datasets are cloud-hosted and what backup exists; Hager confirmed many production AI systems run in cloud environments but that storage and disaster-recovery arrangements vary by organization. Senator Lyons and others asked whether AI might steer clinicians or systems toward costlier care; Hager and Malik said that while algorithms could theoretically promote cost-driven options, clinical and ethical guardrails and validation processes must prevent that outcome.

No formal votes or policy actions were taken during the session; presenters and committee members framed the meeting as the start of an ongoing conversation and identified several options for follow-up, including studying existing state laws, consulting ONC and AMA guidance, and inviting vendors for demonstrations.

Ending note

Both witnesses urged the committee to involve clinicians, patient advocates and technology developers when shaping state responses. Malik and Hager recommended focusing on explainability, clinician validation of tools and improving the telehealth and broadband foundations that will determine whether Vermont residents benefit from — or are left behind by — rapid AI adoption.