The network had done the hard commercial work: it had licensed regulator-cleared AI triage models built to flag urgent findings in imaging studies. What it had not bought, because the vendors did not sell it, was any connection between those models and the way radiology actually runs across 14 hospitals. The picture archive, the radiologist worklists and the on-call escalation paths all sat on one side of a gap, and the models sat on the other.

Overnight was where the gap hurt most. Studies queued strictly first-in-first-out, so a scan with a potentially urgent finding waited behind routine work with nothing to move it up the list. If a model had flagged it, no one would have known: there was no route from a model flag to the right reader, and no record of whether anyone acted on a flag at all.

For radiologists the licensed models were invisible; for department leads they were an unaccountable line item. The network could not say which studies the models would have prioritised, could not audit what happened after a flag, and could not defend continuing to pay for capability that never touched a patient pathway.

We built the integration layer the vendors left out, treating the triage models as services to be wired into the network's existing workflow rather than tools to be worked around. A DICOM router identifies eligible studies as they arrive and streams them to the third-party triage services, so the models see the right scans without anyone lifting a finger.

An orchestration service takes the flags coming back and writes prioritisation directly into the PACS worklists radiologists already use, so a flagged study rises up the queue inside the familiar screen rather than in a separate portal. When a flag needs a human now, mobile on-call notifications reach the right reader, and full audit logging records every flag, every notification and every acknowledgement.

Rollout ran hospital by hospital behind feature flags, so each site could be switched on, observed and tuned without risking the rest of the network. Dashboards tracking pipeline latency, uptime and flag volumes gave the programme team live evidence at every stage, which is how all 14 hospitals went live without a single episode of unplanned PACS downtime.