The PACS worklist is the surface a radiologist stares at when they sit down to start their shift. How it is ordered determines which study gets read first, which patient waits longest, and which urgent finding might be discovered an hour later than it should have been. Most departments take that ordering for granted because it has always worked the same way. Understanding what actually changes when an AI triage layer enters that picture, specifically on day one of deployment, is useful grounding before anyone commits to a configuration.
How Manual Prioritization Currently Works
In most radiology departments, the PACS worklist defaults to ordering studies chronologically by time received, sometimes with a manual priority field that can be set by the ordering clinician, the radiology scheduler, or the technologist. Stat orders appear first because the ordering clinician or ward staff designated them as stat. Everything else appears in order of arrival.
The problem with this arrangement is that it treats stat designation as the only prioritization signal. A study ordered as routine that contains an acute finding does not get escalated because there is no mechanism to detect the finding before the radiologist reads the study. The study sits where it landed in the queue, regardless of what is inside it.
Manual supplemental prioritization exists in various forms. Technologists who complete a study and notice something on the scout or localizer images can call the reading room or flag the study. On-call radiologists take calls from ward teams about suspected urgent findings. Some departments have radiologist assistants or advanced practice radiographers who triage physically before passing to the radiologist. All of these mechanisms require a human to recognize that escalation is warranted before the escalation can happen. Under normal volume conditions, they work adequately. Under high volume conditions, the mechanisms break down precisely when they are most needed.
What AI Triage Actually Does in the Worklist
An AI triage layer operates between study completion and worklist delivery. When a study finishes acquisition, the DICOM images are pushed to the PACS per standard workflow. Simultaneously, the images are sent to the AI pre-read system, which runs analysis in parallel while the PACS processes the study. The analysis output, a priority classification or finding flag, comes back within a defined latency window, typically two to five minutes for most modalities at reasonable infrastructure capacity.
The triage output then updates the worklist position of the study. This update happens via an HL7 order update message or a DICOM Worklist modification, depending on how the PACS integration is configured. The radiologist looking at the worklist sees studies in a new order: flagged studies near the top regardless of when they arrived, unflagged studies in their original chronological order below them.
On day one, this is the concrete change. The worklist looks different. Studies that the AI has flagged appear at or near the top. The radiologist still reads every study. They still sign every report. The AI has not changed who reads what. It has changed when they encounter it.
The Day-One Adjustment Period
The first day or two of an AI triage deployment surface a predictable set of observations from radiologists. Some flagged studies will be read and found to be normal or near-normal. The AI flagged based on a finding that the radiologist, on full review, does not consider significant. This is expected. A triage system calibrated for high sensitivity at the cost of some specificity will surface borderline studies along with genuine urgent findings. Radiologists who understand this calibration will read the flagged study and move on. Radiologists who expected the AI to be perfectly selective will be frustrated by what they perceive as noise.
Setting expectations before deployment matters here. AI triage is not meant to be a perfect filter. It is meant to ensure that studies with findings of genuine concern are not buried in a chronological queue. A small number of false positives, studies flagged that turn out not to require expedited attention, is an acceptable cost of ensuring that genuine urgent findings are not missed because they arrived at 10:02 AM in a department with a 9:45 AM backlog.
The second observation on day one: the lower part of the worklist, the studies that did not get flagged, still exists and still needs to be read. AI triage does not reduce total volume. It reorganizes the order. A radiologist who has internalized the idea that AI will "handle" routine studies will be disappointed, because the routine studies are still there. They are just ordered after the flagged ones.
Configuration Decisions That Affect the Day-One Experience
How the integration is configured determines how disruptive or comfortable day one feels. The key decisions are: what finding categories trigger a flag, what the worklist display looks like with the new priority tier, and whether the AI flag is visible to the radiologist or whether it is purely a queue-ordering signal.
Showing the AI flag to the radiologist, with a brief indication of the finding type that triggered it, is generally better than hiding it. A radiologist who knows why a study was elevated to the top of the worklist is mentally prepared when they open it. A radiologist who just sees a study at the top without knowing why is mildly disoriented and may wonder if the worklist ordering is malfunctioning.
Displaying the flag also preserves the independent review structure important for the CDS framework discussed in other contexts. The radiologist knows the AI flagged the study. They read the study. Their report reflects their independent assessment. The flag is a prompt, not a conclusion.
What to avoid: worklist configurations where AI flags trigger automatic escalations to clinical teams before the radiologist has read the study. The AI pre-read is an internal workflow tool. Clinical notification of findings should come from the radiologist's report, not from the AI pre-read output. A workflow that notifies the ordering physician based on the AI flag, before radiologist review, creates a clinical communication loop that is disconnected from the professional read and creates liability exposure.
What Does Not Change on Day One
The radiologist's reading workflow does not change. They open the study in PACS the same way. They dictate the report the same way. The report is signed the same way. The downstream clinical communication happens through the signed report, same as always.
The technologist workflow does not change materially. Studies are acquired and pushed to PACS per standard protocol. The AI system receives images from the PACS in a configuration that does not require the technologist to do anything differently.
Quality review processes do not change. Peer review, case conference selection, and QA auditing all operate on the radiologist's signed reports, not on AI triage outputs. The AI triage layer is upstream of the report and does not appear in the peer review record.
We are not saying that nothing requires adjustment. Radiologists need a brief orientation to the flagging logic so they know what finding categories the AI is trained to detect and at what sensitivity. Imaging IT needs to understand the HL7 or DICOM integration touchpoints in case of PACS upgrade or integration failures. The operations lead needs a dashboard to track triage-to-read latency, flag rates by modality, and flag override rates. These are real setup requirements. But for the radiologist sitting down on day one to read a shift, the experience is recognizable. The worklist just has a different order at the top.