EMS providers draw a sharp line between helpful automation and unsafe clinical intrusion
Emergency medical services move at the pace of seconds, not clinic appointments. In that environment, the appeal of AI decision support is obvious: retrieve the right protocol faster, trim documentation time, and surface medication details before a mistake reaches the patient. A participatory design study with EMS providers explored exactly where artificial intelligence fits into that workflow and, just as important, where it does not belong.
The work centered on designing and critiquing candidate AI applications for time-critical prehospital care. EMS clinicians were asked to identify practical uses, then challenge those ideas against the realities of noisy scenes, limited bandwidth, split-second decisions, and the need to preserve professional judgment. Their responses separated low-risk support functions from applications that could distort clinical reasoning or disrupt patient care.
One message came through with unusual clarity: AI can assist the crew without replacing the crew.
Protocol retrieval and documentation support emerged as the most acceptable use cases
Providers identified several functions that could reduce friction without changing the fundamental structure of emergency care. Fast access to protocols and medication references ranked among the clearest opportunities. In a system where a paramedic may need to confirm dosing, contraindications, or a treatment algorithm while under pressure, an AI tool that retrieves the right reference quickly could shorten delays and reduce cognitive load.
Speech-based documentation support also drew interest. EMS documentation often happens under less-than-ideal conditions, sometimes while the team is still managing airway, circulation, or transport logistics. A voice-enabled system that generates a draft record during care could reduce charting burden and free attention for the patient. The appeal here is not abstract efficiency; it is the possibility of preserving situational awareness while the record is assembled in real time.
Providers also saw value in AI-generated patient “snapshots” that synthesize relevant history from previous encounters. For prehospital teams, the difference between a vague complaint and a concise history can determine whether they anticipate deterioration, medication interactions, or an evolving chronic condition. A structured summary of prior encounters could help crews orient themselves before arrival at the hospital.
Medication recognition rounded out the favorable use cases. Identifying home medications and surfacing safety-relevant information could help EMS teams avoid avoidable adverse events, especially when patients cannot communicate clearly or are unable to provide an accurate list. In prehospital medicine, medication uncertainty is not a minor inconvenience. It is a common setup for error.
Diagnostic AI and ambient surveillance raised immediate concern
The cautions were just as instructive as the endorsements. Providers resisted clinical reasoning support when it could be interpreted as a diagnostic assertion. In EMS, the first assessment is often provisional by necessity, based on incomplete information, dynamic symptoms, and a scene that can change in minutes. An AI-generated diagnosis may project false certainty onto a process that is deliberately probabilistic and iterative.
They were also wary of ambient “always-on” monitoring designed to flag errors or workflow deviations. Such systems can sound appealing in a quality-improvement deck, but they carry a different meaning on the street. If every action becomes a potential alert, the technology risks becoming a second supervisor rather than a clinical aid. That kind of surveillance can erode trust, distract from urgent tasks, and make teams less willing to rely on the tool at all.
Voice-based exchanges with AI in front of patients drew similar hesitation. EMS encounters are intimate, time-sensitive, and often emotionally charged. A device that speaks back in the room may introduce confusion, reduce rapport, or make patients feel they are being managed by a machine rather than cared for by a clinician. The setting matters: what looks seamless in a prototype can feel intrusive at the bedside or in the back of an ambulance.
Workflow fit and trust calibration mattered more than technical novelty
Across concepts, providers repeatedly returned to the same design constraints. Seamless workflow fit was essential. If a tool interrupts a handoff, slows scene management, or requires too many steps to retrieve a useful answer, it fails before it earns trust. Robustness in noisy environments mattered for obvious reasons: sirens, traffic, bystanders, and radio chatter are not edge cases in EMS. They are the operating environment.
Transparency was another nonnegotiable feature. Users wanted to understand where an AI recommendation came from, what data it used, and how confident it might be. Without that visibility, trust calibration becomes guesswork. Overtrust can be dangerous when a tool is wrong. Undertrust can make a helpful system functionally invisible. In emergency care, either failure mode has consequences.
Professional autonomy remained central to adoption. EMS clinicians did not describe a desire for an algorithm that quietly takes over judgment. They wanted support that strengthens their decisions without flattening them into machine output. That preference reflects a core reality of prehospital medicine: decisions are made in context, often with incomplete histories and rapidly changing physiology. A tool that cannot tolerate ambiguity is poorly matched to the field.
The study points to a narrow but practical role for EMS AI
The most clinically credible uses of AI in EMS were the least theatrical. Retrieval, summarization, documentation assistance, and medication recognition all reduce friction around the edges of care while leaving the clinician in charge of the decision itself. Those functions align with how prehospital teams actually work, especially when minutes matter and documentation cannot be postponed indefinitely.
The less acceptable ideas shared a common flaw: they tried to push AI into spaces where human judgment, patient interaction, or scene awareness is already doing essential work. Diagnostic assertions can harden uncertainty into false confidence. Always-on monitoring can turn support into surveillance. Voice interaction in front of patients can disrupt a therapeutic encounter before it has a chance to settle.
The study’s value lies in its restraint. Rather than treating EMS as a showcase for every available AI feature, it uses the field to define boundaries. Those boundaries will matter as vendors and health systems move from abstract enthusiasm to real deployment. In a setting defined by noise, urgency, and limited margin for error, the safest AI may be the kind that quietly reduces administrative burden and improves access to information while leaving clinical authority where it belongs — with the EMS crew on scene.