Radiology AI worklist triage is not one intervention but several, and the distinction determines what a department can measure. At minimum, it separates tools that flag or notify from tools that actively reorder a reading queue, and both differ from systems that assign studies to radiologists. The US FDA defines one category through 21 CFR 892.2080 as software that aids prioritization but neither marks images nor removes cases from a queue [2]. ACR–SIIM guidance places AI integration inside a broader governance and monitoring obligation [1]. Published time-savings claims depend heavily on which of these functions a study actually implemented, making the workflow description as important as the model's standalone accuracy. The systematic review through July 2024 [7], the JACR CTPA retrospective [5], the JMIR prospective chest radiograph study [3], the Radiology primer [4], the AWS Bedrock AgentCore walkthrough [6], and the Intelerad InteleOrchestrator page [8] are cited by their source numbers throughout.

What Triage Software Is Allowed to Do

The regulatory boundary is explicit. Radiological computer aided triage and notification software is a prescription device that notifies designated clinicians of time-sensitive images based on computer analysis, but it does not mark, highlight, or direct attention to a specific image location, and it does not remove cases from the reading queue [2]. The device operates in parallel with standard care, which remains the default option for all cases [2]. That design means the software can change who is alerted and when, but not what the radiologist sees inside the image or whether a study disappears from the list.

This leaves the actual reordering problem to the worklist layer. A Radiology primer notes that standalone AI screening for emergent findings cannot claim the top position on a worklist because it is unaware of study availability, completeness, ordering department, chronicity, and service-level turnaround agreements [4]. A worklist manager with access to the AI flag could reprioritize a suspected pneumothorax upward and reduce confirmation time [4]. The technical constraint is therefore not whether the algorithm detects an abnormality, but whether the surrounding system can receive, interpret, and act on that signal without breaking other queue priorities.

Two Workflow Designs: Notification Versus Active Reordering

The operating difference surfaces in the literature. A systematic review through July 2024 found that patients in the post-triage group had shorter median report turnaround times, with mean differences of 12.3 minutes for pulmonary embolism, 20.5 minutes for stroke, 4.3 minutes for intracranial hemorrhage, and 29.7 minutes for chest diseases, with wide interquartile ranges [7]. Subgroup analysis indicated reductions of 43.7% for algorithms that actively stratified and reordered the worklist, compared with 7.6% for widget-based systems, a statistically significant difference [7]. That gap is a workflow effect, not simply a detector accuracy effect: the same urgency signal produces different operational value depending on where it lands.

A prospective study of 1,054 chest radiographs between November 2023 and January 2024 used a triage system with a four-level notification scale of normal, nonurgent, urgent, and critical, with pneumoperitoneum and pneumothorax deemed critical, and the resulting notification levels informed the sequence of studies on the reporting worklist [3]. The unaided session followed first-in, first-out order, while the AI-assisted session reordered and also generated draft report text [3]. Median report generation time fell from 2 minutes to 0.53 minutes, and mean turnaround time from 876.21 to 82.25 minutes, with reductions across all urgency categories [3]. Because prioritization and report generation were applied together, the study cannot isolate the contribution of triage alone.

A Suggested Local Evaluation: Work Hours Versus Off Hours for a Single Flag

A concrete evaluation can separate detector performance from workflow effect. Suppose a practice deploys an FDA-cleared notification tool for suspected pneumothorax on chest radiographs. The queue remains in first-in, first-out order, but a worklist manager moves flagged studies upward. Measure time from study completion to preliminary report finalization for flagged studies, split into work-hours and off-hours cohorts over a set enrollment period. Compare against a pre-deployment control period using the same inclusion rules and exclude reordered cases if the worklist logic changes mid-study.

A minimum outcome set would include the proportion of flagged studies reported within a pre-specified time threshold, median time to report, and the false-notification rate requiring radiologist review with no actionable finding. The JACR CTPA analysis shows why the day-part split matters, because an aggregate median may hide a non-significant off-hours effect [5]. The systematic review's comparison between active reordering and widget display further suggests that the evaluation should record whether the flag altered queue order or only appeared beside the study, since those two modes have yielded different effect sizes [7]. This is a suggested design, not a trial protocol, and it should be reviewed with institutional governance before use.

The Work Hours Problem in Time-Savings Evidence

Time savings do not distribute evenly across the day. A retrospective analysis of 11,252 adult CT pulmonary angiography examinations for suspected PE at a single tertiary academic center compared PE-positive turnaround time before and after an AI triage device [5]. Clinically observed savings during work hours were 22.2 minutes, significant at P = .004, while off-hour savings were 2.82 minutes and not significant at P = .345 [5]. Model predictions were 29.6 minutes for work hours and 2.10 minutes for off-hours, aligning with observation [5]. The authors attribute the pattern to how workflow parameters, not just detection speed, control turnaround time.

This finding matters for procurement because it means the same deployed tool can show negligible effect at night. Departments that measure global mean turnaround time may dilute a work-hours benefit with off-hours noise. The FDA special controls for triage software require performance testing that estimates clinical effectiveness, with cohorts grouped by confounders, effect modifiers, associated diseases, and acquisition characteristics [2]. A local evaluation should apply the same logic to workflow covariates, including staffing, shift boundaries, and the baseline order in which studies are reported.

Agentic Assignment Versus Triage Classification

A newer architecture moves beyond urgency flags to assignment reasoning. Amazon Web Services published a technical walkthrough on 21 May 2026 describing AI agents on Amazon Bedrock AgentCore that evaluate radiologist specialization, current workload, fatigue patterns, case complexity, clinical urgency, and availability to recommend case assignments [6]. The design routes queries through an orchestrator that delegates to sub-agents for radiologist mapping, image assessment, clinical data collection, metadata retrieval, and shift checking [6]. AWS describes this as moving from deterministic rules to autonomous orchestration [6].

That architecture is distinct from FDA-cleared triage notification software. The AWS example is a vendor-described implementation pattern using generative and agentic components, not an assertion of regulatory clearance or a measured clinical result [6]. Bedrock Guardrails is shown intercepting inbound queries and scanning outbound responses to reduce patient PII exposure [6]. Radiology leaders evaluating such systems should ask whether the agent manages a queue, recommends an assignment, or both, and whether the underlying urgency logic has been validated against the same end points used for triage software.

Interoperability and Local Constraints

Integration planning determines whether any triage signal reaches the worklist. The Radiology primer cautions that custom integrations increase operational and maintenance burden and the likelihood of unanticipated problems [4]. If results appear in a third-party application, that application should integrate with existing software and comply with security measures, and worklist reprioritization may require modifications so the worklist can recognize AI-flagged studies [4]. Standards-based purchasing supports both forward and backward compatibility across RIS, EHR, PACS, worklist, and reporting software [4].

Intelerad's InteleOrchestrator page describes vendor-neutral orchestration with plug-and-play connectivity across imaging systems and over 50 third-party solutions, AI-driven prioritization through InteleGence, and RADPAIR reporting integration [8]. The page presents those functions as product features, not audited clinical outcomes. One quoted customer states the orchestrator makes him 20–30% more efficient and reduces mental fatigue [8]. That is an individual attribution from a named physician and practice, not a controlled measurement, and it does not establish the same result for another organization.