Recoverability-Centered Trajectory Governance for Long-Horizon Clinical AI Agents

Clinical AI is moving from systems that generate discrete predictions and recommendations toward agents capable of pursuing clinical objectives through repeated tool use, communication, planning, and state-changing actions. Emerging systems support multistep electronic health record workflows, and increasingly envision longitudinal care across repeated encounters. This transition expands safety assessment beyond discrete outputs and actions. Through its actions, an agent can help create the clinical state on which its subsequent decisions depend. Consequently, individually defensible actions may compose into an unsafe trajectory, even when no single action is clearly erroneous. We propose recoverability-centered trajectory governance as a clinical synthesis of established safety approaches, linking patient-specific corrective capacity to the continuing justification for delegated authority. Its central objective is clinical recoverability: preserving the opportunity for the human-AI care system to redirect care before clinically acceptable corrective pathways are exhausted. We define recovery margin as the remaining opportunity for such correction, and reauthorization boundaries as conditions requiring delegated authority to end or narrow pending renewed human judgment. This framing distinguishes reversibility, primarily a property of an action, from recoverability, a property of an evolving clinical trajectory. The operational framework combines persistent obligation tracking, independent monitoring, explicit reauthorization, authority contraction, graceful degradation, independent escalation, and verified human takeover. Oversight should concentrate at consequential trajectory junctions. Evaluation should combine trajectory stress testing, component-specific recovery-margin indicators, interruption-and-takeover testing, and population-level surveillance. Trajectory governance complements product lifecycle oversight by examining whether continued patient-specific delegation remains justified as care evolves. Because corrective capacity depends partly on access to timely monitoring and human rescue, recoverability is also an equity issue. The defining safety question is therefore not only whether an action is acceptable but whether continued delegation remains justified. The proposed indicators and governance mechanisms require domain-specific operationalization and empirical validation, including comparison with established safeguards.

Authors

Publication Details

Journal
Journal of Medical Internet Research
Published
2026-11-19
DOI
https://doi.org/10.2196/110563
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Recoverability-Centered Trajectory Governance for Long-Horizon Clinical AI Agents

Siru Liu, Jialin Liu
Journal of Medical Internet Research
Artificial Intelligence in Healthcare and Education
article

Recoverability-Centered Trajectory Governance for Long-Horizon Clinical AI Agents

Siru Liu, Jialin Liu
article en

Abstract

Clinical AI is moving from systems that generate discrete predictions and recommendations toward agents capable of pursuing clinical objectives through repeated tool use, communication, planning, and state-changing actions. Emerging systems support multistep electronic health record workflows, and increasingly envision longitudinal care across repeated encounters. This transition expands safety assessment beyond discrete outputs and actions. Through its actions, an agent can help create the clinical state on which its subsequent decisions depend. Consequently, individually defensible actions may compose into an unsafe trajectory, even when no single action is clearly erroneous. We propose recoverability-centered trajectory governance as a clinical synthesis of established safety approaches, linking patient-specific corrective capacity to the continuing justification for delegated authority. Its central objective is clinical recoverability: preserving the opportunity for the human-AI care system to redirect care before clinically acceptable corrective pathways are exhausted. We define recovery margin as the remaining opportunity for such correction, and reauthorization boundaries as conditions requiring delegated authority to end or narrow pending renewed human judgment. This framing distinguishes reversibility, primarily a property of an action, from recoverability, a property of an evolving clinical trajectory. The operational framework combines persistent obligation tracking, independent monitoring, explicit reauthorization, authority contraction, graceful degradation, independent escalation, and verified human takeover. Oversight should concentrate at consequential trajectory junctions. Evaluation should combine trajectory stress testing, component-specific recovery-margin indicators, interruption-and-takeover testing, and population-level surveillance. Trajectory governance complements product lifecycle oversight by examining whether continued patient-specific delegation remains justified as care evolves. Because corrective capacity depends partly on access to timely monitoring and human rescue, recoverability is also an equity issue. The defining safety question is therefore not only whether an action is acceptable but whether continued delegation remains justified. The proposed indicators and governance mechanisms require domain-specific operationalization and empirical validation, including comparison with established safeguards.

Journal of Medical Internet ResearchVol. 28
Artificial Intelligence in Healthcare and Education
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon