Healthcare
Average first-hold time, down from 52 seconds
Routing accuracy achieved on audit, up from a baseline with an 18% mis-route rate
About
The network sought to turn chaotic phone queues into fast, accurate care routing. Average hold time was 52 seconds, 18% of emergencies were mis-routed to finance or the wrong department, and regulators had begun flagging the pattern. We built an LLM-based triage layer that classifies urgency in real time and pushes patient context directly into the EMR.
Industry
Healthcare
Company size
1,000 – 5,000 employees
Founded
2000
The Company
A hospital group whose phone queue had become a risk
The network operates a multi-facility hospital and clinic system fielding a high volume of daily patient calls covering everything from appointment scheduling to genuine medical emergencies, all arriving through the same intake queue.
Call routing depended on front-line staff correctly triaging urgency and department in real time under high call volume — a process prone to error precisely in the moments, like a genuine emergency, where getting it wrong mattered most.
The challenge
52-second holds and emergencies routed to finance
Average hold time before a call was even answered ran 52 seconds, and 18% of emergency-classified calls were mis-routed — in some cases landing in the finance department rather than clinical triage. Regulators had begun flagging the pattern as a compliance concern.
The routing errors weren't a training problem so much as a volume-and-pressure problem: staff triaging calls in real time under load will make more classification errors than a system built specifically for that task, and the consequences in a healthcare setting are more serious than in most other call-routing contexts.
The Solution
An LLM triage layer with HIPAA-grade logging into the EMR
We built a Twilio telephony integration passing calls to a large language model that classifies urgency in real time, logs HIPAA-grade transcripts, and pushes patient context back into the electronic medical record through FHIR APIs — so by the time a call reaches a human, the relevant department already has the context rather than starting from zero.
The classification model was tuned specifically on the urgency and department-routing distinctions that mattered most for the network's own call patterns, rather than a generic intent classifier, to hold routing accuracy to the standard a healthcare setting requires.
The Results
Hold time down to 14 seconds, routing accuracy audited at 97%
First-hold time fell from 52 seconds to 14 seconds, and routing accuracy audited at 97%, a marked improvement over the baseline mis-route rate that had drawn regulatory attention. No further routing-related regulatory complaints have been recorded since rollout.
Clinical staff also reported that calls arriving with FHIR-integrated context meant less time spent re-establishing basic patient information at the start of every call, letting the actual clinical conversation start sooner.
KEEP READING
Explore more studies on AI, analytics, and enterprise intelligence.
