Patient-Held AI Moves the Evaluation Question Away From Documentation Burden
Ambient AI scribes have entered clinical care with a narrow reputation: they are being built to reduce note-taking, cut clerical work, and spare clinicians from the after-hours grind of documentation. This preprint pushes the conversation in a different direction. The focus shifts from the clinician-facing scribe to a patient-held AI tool, framed as a Personal Health Assistant that could help patients communicate with clinicians, retain complex information, and stay oriented after the visit ends.
The distinction is immediate at the bedside. A tool that lives in the patient’s hands serves a different purpose than one embedded in the clinician workflow.
Rather than asking only whether the technology saves time, the investigators set out to define what success should look like from the patient and caregiver perspective. They partnered with a large technology company and used the Nominal Group Technique to co-design an evaluation framework for a patient-held AI tool in real-world settings. The exercise was deliberately structured to surface outcomes that matter to the people who have to live with the consequences of a visit once the encounter is over.
Patients and Caregivers Prioritized Coordination, Not Billing Convenience
The study brought together 33 patients and caregivers and 10 multidisciplinary clinicians in structured virtual sessions. Those participants identified outcomes for future studies of patient-held AI, and the priorities diverged sharply from the commercial emphasis now surrounding ambient documentation products.
Industry attention has clustered around the scribe function, billing support, and operational efficiency. The stakeholder group in this project looked elsewhere. They placed greatest weight on longitudinal care management and post-visit coordination, signaling that the real value of patient-held AI may lie in what happens after the note is signed, not merely during the encounter.
Communication and respect emerged as the most critical outcome. That choice is clinically revealing. Patients did not frame AI primarily as a shortcut for paperwork; they saw it as a mechanism to improve the human-to-human dynamic, preserve shared understanding, and reduce the friction that often develops when complex instructions are compressed into a rushed conversation. In practical terms, a tool that records, organizes, and returns information to patients could help them recognize medication changes, follow-up tasks, warning signs, and next steps with fewer gaps in recall.
Consent, Trust, and Access to the Record Became Central Clinical Problems
The sharpest hurdle was not technical performance. It was clinician trust and the need for formal consent processes.
That finding matters because patient-held AI sits inside a sensitive boundary: it touches the encounter, the record, and the relationship between the patient and the health system. If a tool captures the clinical conversation, questions about who controls the transcript, how it is shared, and whether clinicians are comfortable with its use become central rather than peripheral. The study places those governance concerns on equal footing with usability.
The article also draws attention to a basic but consequential gap in current practice. Patients cannot assume they will automatically have access to the transcript of their own visit when that record is created and retained by the health system. A patient-held tool could fill that void by preserving discussion points and follow-up steps in clearer detail than a typical clinical note or visit summary. For patients navigating a new diagnosis, a medication change, or a complicated care plan, that access could be the difference between informed action and a missed instruction.
Evaluation Should Measure Advocacy and Engagement, Not Only Efficiency
The framework proposed by this work asks health systems to measure something broader than workflow compression. It suggests that future studies of patient-held AI should track patient-side advocacy, engagement, and coordination across the care journey.
This is a different evidentiary standard. If the only metric is whether clinicians chart faster, the technology will be judged through an administrative lens. If the endpoint includes whether patients feel heard, whether they understand the plan, and whether they can participate more effectively in subsequent care, then the same tool is being tested against a clinical communication standard. In chronic disease management, complex specialty care, and post-visit recovery, those outcomes may matter more than minutes saved in the exam room.
The investigators argue for a partnership model in which patients, caregivers, clinicians, and developers shape how the technology is studied. That approach is especially relevant for tools that blur the line between documentation support and care navigation. A patient-held AI that improves comprehension but erodes trust would fail a bedside test even if it performs well on a software dashboard.
What This Framework Means for Real-World AI Deployment
No sample size in this preprint is large enough to settle questions of safety, equity, or generalizability. The work is conceptual and qualitative, designed to define outcomes rather than test clinical effect. Even so, the priorities surfaced here are likely to influence how subsequent studies are built. If future trials ignore patient-defined metrics, they risk measuring the wrong thing with impressive precision.
The most consequential implication is that patient-held AI should not be treated as a clone of the ambient scribe with a different user interface. Its job is not only to transcribe what clinicians say. It may also need to reinforce understanding, support adherence to follow-up instructions, and preserve a record the patient can actually use.
Data privacy restrictions limited access to the qualitative dataset, and the preprint remains unreviewed. Those are real constraints, especially for a technology area moving quickly and often ahead of regulation. Even so, the paper makes one point with unusual clarity: if patient-held AI is going to enter clinical care, its evaluation has to begin with the patient’s version of utility, not the vendor’s.