When judgment moves: agentic artificial intelligence and the relocation of human discretion in frontline care

1 Introduction

Healthcare policy becomes real when a frontline worker interprets a rule, prioritizes a case, allocates limited time, or responds to a patient whose condition does not fit a standard category. Triage nurses, intake personnel, community health workers, allied health staff, and other frontline actors do not share one profession. However, they occupy a similar position between formal systems and individual persons. Their work involves discretion, which refers to situated judgment where rules cannot fully determine the most appropriate response (Lipsky, 2010; Maynard-Moody and Musheno, 2003).

This discretion is necessary because frontline care is rarely delivered under complete information or unlimited resources. Workers frequently need to respond to uncertainty, competing needs, and differences among patients that cannot be fully captured by standard procedures. Therefore, discretion is not simply a personal choice made by the worker; it is part of how formal policy is adjusted to the real conditions of care. It also allows the worker to explain a decision, respond to vulnerability, and consider circumstances that may not appear in the formal record.

Street-level bureaucracy explains why this judgment matters. Frontline workers do not only apply policy; their repeated decisions shape what people actually receive from public and professional services (Lipsky, 2010). Later work on system-level bureaucracy showed that digital systems can move discretion away from visible workers and place it within technical arrangements (Bovens and Zouridis, 2002; Buffat, 2015). This issue has become more important because healthcare artificial intelligence is moving from prediction and recommendation toward systems that can plan, coordinate, adapt, and initiate actions across connected workflows (Borghoff et al., 2025; Collaco et al., 2026; Hinostroza Fuentes et al., 2025).

Existing concepts do not fully explain this development. Decision support informs the professional, while automation usually performs a predefined task. Delegation involves an authorized transfer of responsibility that remains identifiable and may be withdrawn. Agentic artificial intelligence may go further because the system can interpret a situation, select an action, and determine what happens next. In such cases, the worker may remain present and accountable while no longer being the main author of the decision.

The present article distinguishes agentic artificial intelligence from conventional decision support and prescriptive automation. It then introduces relocated discretion to explain when meaningful judgment moves from frontline professionals to artificial intelligence systems. Based on this distinction, the article discusses the possible human consequences of this movement and the governance conditions required in frontline care. The distinctions among augmentation, automation, delegation, and relocated discretion are summarized in Table 1.

Table 1. Distinguishing relocated discretion from related arrangements.

2 From decision support to agentic action

Clinical decision support systems usually provide alerts, predictions, structured information, or recommended actions. In these systems, the professional remains the principal decision-maker because the system does not normally control the full process from interpretation to execution. Prescriptive artificial intelligence goes further by identifying a preferred action. Rule-based automation may also execute a predefined response. In both arrangements, the pathway is mainly bounded by prior human design.

These arrangements differ according to the amount of judgment retained by the professional. A descriptive system may identify a pattern or estimate a risk. A prescriptive system may recommend a preferred response. Automation may then perform a task that has already been specified. Agentic artificial intelligence extends beyond these functions when it connects several activities and determines how the workflow should continue. The difference is therefore not based only on technical complexity. It is based on whether the system begins to exercise meaningful choice within the care process.

Agentic artificial intelligence is different because it combines goal-directed operation with multistep planning, adaptation, and action initiation. It can take initiative within a wider sociotechnical system that includes human professionals, organizational rules, and digital platforms (Collaco et al., 2026; Salah and Alnoor, 2026). Unlike conventional systems that respond to a single request, agentic artificial intelligence can monitor changing conditions, organize several connected tasks, and determine how a workflow should proceed within defined limits. A conventional triage tool may calculate a risk score for a nurse. An agentic system may collect symptoms, retrieve records, identify missing information, assign urgency, arrange an appointment, and trigger escalation. Therefore, the central issue is whether the professional still interprets the case, develops the possible responses, and can meaningfully change the action before it affects the patient pathway.

Autonomous diagnostic systems provide an earlier example of clinical judgment moving into technical systems. In the pivotal trial reported by Abràmoff et al. (2018), the system generated its diagnostic output at the point of care without specialist interpretation. This system was not agentic in the current meaning because it did not plan and coordinate a wider workflow. However, it illustrates the earlier movement of clinical judgment from professionals to autonomous technologies. Agentic artificial intelligence extends this movement by adding planning, coordination, and execution across several stages of care.

This distinction is also important because human discretion is not always accurate or consistent. Artificial intelligence may reduce information overload, improve standardization, and support better decisions. Ebnali Harari et al. (2025) conducted a randomized controlled trial with 54 participants who had no medical background or previous cardiopulmonary resuscitation (CPR) experience. During a simulated CPR scenario, participants using clinician-supervised ChatGPT achieved higher decision accuracy and reported greater trust than those using autonomous ChatGPT or a paper guide, although they required more time to complete the scenario. The study was conducted in a simulated setting with non-medical participants. Therefore, the findings should not be treated as direct evidence of improved clinical outcomes. However, they suggest that expert supervision may improve the reliability and acceptance of artificial intelligence-supported guidance. This supports the argument that the value of artificial intelligence depends on how professional judgment is retained within the decision process.

This balanced position is important because the argument is not based on the assumption that human judgment is always better. Human workers may be affected by fatigue, inconsistency, limited knowledge, or personal bias. Artificial intelligence may reduce some of these problems. However, the improvement of technical performance does not answer the question of who should exercise judgment in cases involving uncertainty, vulnerability, and moral consequences.

3 Relocated discretion: definition and boundaries

Discretion exists because rules are general while cases are particular. In frontline care, incomplete information, scarce resources, emotional distress, competing risks, and unequal vulnerability require judgment that cannot be fully written in advance. Such judgment also has a moral dimension. Workers decide what is urgent, what is fair, and when an exception may be justified (Maynard-Moody and Musheno, 2003; Zacka, 2017). Resource allocation makes this especially clear because prioritization determines who is seen, delayed, redirected, or excluded (Elgin and Elgin, 2024).

Relocated discretion refers to a shift in where consequential judgment is exercised: interpretive, prioritization, or action-selection authority moves from the frontline professional to an artificial intelligence system that can independently choose, plan, or initiate actions within the care workflow. It occurs when three conditions are met. The system must first select among meaningful alternatives rather than only perform a routine instruction. Its selection must then shape or initiate later workflow activities. Finally, the professional remains answerable for the result while having limited authorship of how the decision was formed.

This concept extends system-level bureaucracy. Bovens and Zouridis (2002) showed that information technology can transfer discretion from frontline workers into system design, while Buffat (2015) explained how e-government changes where discretion is exercised. Agentic artificial intelligence intensifies this development because the system may not only structure the available choices but also interpret changing conditions and execute a sequence of actions. Professional identity and moral agency help explain why this affects more than efficiency. Responsibility gap theory also explains why accountability may remain with the worker after meaningful authorship has moved elsewhere (Santoni de Sio and Mecacci, 2021).

Relocation is not identified by the use of artificial intelligence alone; the central issue is where meaningful choice occurs. It may be acceptable in bounded, reversible, and low-sensitivity activities. Greater concern arises when the system responds to vulnerability, uncertainty, or competing moral claims and the worker cannot adequately inspect or challenge the basis of the action.

4 Human consequences of relocated discretion

Professional identity is closely connected to the judgment exercised during frontline care as it develops through competence, recognition, responsibility, and the performance of meaningful work. Clinical decision support can already affect professional identity when workers perceive reduced control, expertise, or patient connection (Ackerhans et al., 2024). Process design features, including explainability, workflow integration, and individual accountability, were also found to shape trust and professional identity threat in different ways (Ackerhans et al., 2025). The risk may become greater when an agentic system performs the judgment-bearing act rather than only supporting it. A triage nurse may remain the person facing the patient while urgency, scheduling, and escalation have already been determined by the system.

This change may also reduce professional autonomy. Identity concerns how workers understand their role, while autonomy concerns what they are still able and permitted to decide. Evidence on artificial intelligence and physician autonomy remains limited and mixed, with possible gains and losses (Grosser et al., 2025). This caution is important because human discretion may itself be biased, inconsistent, or overloaded. However, formal participation should not be confused with actual authority. A professional who can only approve a completed pathway—without enough time, information, or institutional support to revise it—remains involved in appearance but has limited operational autonomy.

Moral deskilling may not appear immediately. It can develop gradually when professionals no longer practice moral judgment in difficult cases. Moral capacity is maintained through repeated engagement with situations in which rules, needs, and values do not align (Vallor, 2015). Discharge planning provides a useful example. An agentic system may assess readiness, coordinate medication, arrange follow-up, contact community services, and recommend discharge. The clinician may formally approve the plan but may not be able to reconstruct how the system weighted the medical, social, and organizational factors. If this becomes normal practice, professionals may lose important opportunities to exercise situated ethical reasoning.

Relocation may also create a gap between answerability and authorship. Under human discretion, the worker may be questioned because the worker made the judgment. Under relocated discretion, the worker may still be required to explain a decision produced through a process that is only partly visible. Santoni de Sio and Mecacci (2021) distinguish several responsibility gaps, including public and moral accountability gaps. Both are relevant when a patient requests an explanation and the professional cannot provide one, or when the professional cannot understand the result as his or her own decision. Medication management provides a clear example. If an agent reviews interactions, requests tests, changes priorities, and communicates across units, the final professional may be answerable for a chain of actions that no single human fully authored.

These consequences are conceptual risks rather than inevitable outcomes. Their strength is likely to depend on task sensitivity, system transparency, workflow design, professional expertise, and the real possibility of intervention. They therefore identify areas that should be examined before agentic systems become normal across frontline care.

5 Conditions and governance of relocated discretion

Relocated discretion is not always harmful. It may be useful where tasks are repetitive, criteria are stable, consequences are limited and reversible, and delay or inconsistency creates greater risk than standardized action. It may also help professionals manage large amounts of information and coordinate fragmented services. Greater concern arises when evidence is incomplete, patients are vulnerable, decisions involve moral tradeoffs, or the professional remains responsible without meaningful control.

Governance should begin with classification according to clinical, moral, and contextual sensitivity. Low-sensitivity administrative actions may allow greater system autonomy with retrospective monitoring. Medium-sensitivity decisions should require meaningful professional review before execution. High-sensitivity decisions involving serious clinical consequences or individualized vulnerability should remain a human decision before execution. This classification does not create a permanent line between human and machine tasks. It requires healthcare organizations to justify why a judgment-bearing activity has been transferred.

Human review must also be real, as a formal approval button does not preserve discretion where workload, time pressure, system complexity, or organizational expectations make disagreement difficult. Healthcare organizations should provide visible records of what the system considered, selected, and initiated. Professionals should be able to pause or revise an action and should have institutional protection when they override the system. Patterns of override should also be reviewed because they may reveal poor system performance, hidden bias, or unsuitable workflow design.

Professional and moral maintenance is equally important; hence, training should extend beyond technical literacy. Workers need case-based practice in questioning, interpreting, and ethically justifying decisions made within artificial intelligence-mediated environments. Organizations should also monitor whether repeated system use reduces expertise, weakens communication with patients, or changes how workers understand their professional role.

Responsibility should follow actual control. Where an agentic system performs a discretionary act, responsibility should not simply return to the most visible worker. Healthcare managers, developers, vendors, professional bodies, and institutions should share responsibility according to their influence over system design, deployment, and working conditions. This approach reduces the risk that professionals remain answerable for outcomes they did not meaningfully author.

6 Conclusion

Relocated discretion explains a change that may be hidden by the language of automation and efficiency. Judgment does not disappear when an agentic system performs it; its location changes. This movement may improve care where tasks are bounded, auditable, and meaningfully supervised, but it may weaken care where moral or contextual judgment moves into an opaque process while professional responsibility remains unchanged. Future research should examine how relocated discretion can be measured, how professionals and patients experience it, and which forms of human–artificial intelligence cooperation preserve both technical benefit and meaningful human authorship.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. ChatGPT 5.5 and Grammarly AI were used to support language editing, grammar checking, formatting, and readability improvement. The tools were not used to generate data, conduct analysis, create references, or replace the authors' scholarly judgment. The authors reviewed and verified the final manuscript and take full responsibility for its content, accuracy, citations, interpretations, and conclusions.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Niko Männikkö, Oulu University of Applied Sciences, Finland

Reviewed by: Iryna Nyenno, Odessa University, Ukraine