The Challenge
A regional health network in Ontario — three hospitals, twelve community clinics, and a home care coordination program — was under sustained operational pressure. Patient volumes had grown 18% over the prior three years. Staffing had not kept pace. The gap was showing up in outcomes that affected both patients and staff in measurable, documented ways.
Emergency department wait times for non-urgent presentations had climbed to an average of 4.5 hours — a number that placed the network in the bottom quartile for comparable Ontario health networks. The problem was not a shortage of emergency medicine capacity; it was that a significant portion of patients arriving at emergency departments had conditions that could have been managed in urgent care or primary care settings. They came to emergency because they did not know where else to go, or because alternative pathways were not easily accessible to them.
The nurse health line — a provincially funded service providing after-hours clinical guidance — was fielding over 800 calls per day with average hold times exceeding 20 minutes. Nurses were handling a mix of genuinely clinical calls requiring assessment and advice, and informational calls that did not require nursing expertise at all: clinic hours, address confirmation, general medication questions, post-visit instruction clarification. The inability to triage this call mix was consuming clinical capacity that was needed for substantive health guidance.
Clinic appointment no-show rates had reached 18%, representing thousands of unused appointment slots annually. The no-show pattern was predictable — certain patient segments, appointment types, and time windows produced higher rates — but the network had no systematic way to act on that predictability. Manual reminder calls were resource-constrained and not personalized to the patient's individual no-show risk profile.
The network's leadership commissioned an independent operational review that identified the core dynamic: the problem was not a shortage of clinical expertise. It was a fundamental misalignment between patient needs and the access points patients were using. Fixing the mismatch required improving how patients were navigated through the system before they arrived at a clinical resource — not adding clinical resources to absorb the misallocated demand.
The Approach
Remolda designed an integrated, three-component AI system targeting the three highest-impact misalignment points: the nurse health line, emergency department demand forecasting, and clinic appointment management. The components were designed to reinforce each other — improvements in health line triage reduced emergency department inappropriate demand; improvements in appointment management increased effective clinic capacity.
Audit (4 weeks). We assessed patient flow across the full network — emergency departments, nurse health line, clinic scheduling, home care coordination, and the referral pathways between settings. The audit combined quantitative analysis of operational data (call logs, wait time records, appointment data, ED triage records) with qualitative interviews of nurses, clinical coordinators, patient navigators, and administrative staff.
Key findings: approximately 35% of ED visits were for conditions triageable to urgent care or primary care settings. Approximately 60% of nurse health line calls were informational — answering questions that did not require clinical assessment. No-show risk was concentrated in identifiable patient cohorts that were not receiving differentiated outreach. Each finding pointed to a specific intervention point.
Strategy (6 weeks). We designed the three-component system in close collaboration with the network's clinical governance committee, emergency medicine leadership, and nurse health line management. Clinical safety was established as the primary design constraint before any technical specification was written: the AI system would err consistently on the side of escalation. A patient with a potentially serious presentation who was navigated to urgent care when they needed emergency care was an unacceptable outcome. A patient with a minor concern who was navigated to urgent care when they could have been managed at home was a suboptimal but acceptable outcome.
All three components were designed to meet PHIPA requirements, with patient health information processed within the network's own infrastructure and no data transmitted to external systems.
Implement — Wave 1: AI Health Line Assistant (2 months). The AI health line assistant was deployed to handle the informational layer of the call queue. Callers asking about clinic hours, addresses, medication instructions, post-visit care questions, and similar non-clinical inquiries were resolved by the AI assistant — typically in under 3 minutes, without hold time. The assistant was configured to route any call with clinical content to a nurse, immediately and without friction, using a structured escalation protocol developed with nurse health line leadership.
Nurses monitored escalation patterns in real time. The clinical governance committee reviewed the assistant's performance weekly for the first eight weeks. The protocol was adjusted based on nurse feedback twice in the first month.
Implement — Wave 2: Emergency Department Demand Forecasting (2 months). Predictive models for ED volume were built using four years of historical data combined with external signals: day of week, seasonal patterns, regional weather, local event schedules, and public health surveillance data. Models were validated against held-out historical data before deployment and generated 48-hour rolling demand forecasts for each emergency department in the network.
ED operations managers used the forecasts to adjust staffing levels and coordinate with urgent care partners on anticipated redirect volumes. The forecast accuracy rate exceeded 85% within 15% of actual volume during the first three months of deployment.
Implement — Wave 3: Intelligent Appointment Management (3 months). The appointment management system used patient-level no-show risk scoring, generated from historical patterns across appointment type, patient segment, time window, and past attendance, to drive differentiated reminder strategies. High-risk appointments received personalized reminders via the patient's preferred channel (SMS, phone, or email), two-way rescheduling capability, and proactive waitlist offers to replace anticipated no-shows. Moderate-risk appointments received standard automated reminders with rescheduling links. The system also managed waitlist dynamics — when a no-show was predicted with high confidence, waitlist patients were proactively offered the slot before the appointment day.
Empower (parallel). Training was delivered across three groups with distinct needs: nurse health line staff, whose daily workflow changed most significantly as informational calls shifted to the AI assistant; clinical coordinators, who needed to work with demand forecast outputs and integrate them into scheduling decisions; and administrative staff, who managed the appointment system and needed to understand the risk-scoring logic well enough to handle exception cases appropriately.
The Results
- 40% reduction in non-urgent ED wait times. Average wait time for non-urgent presentations dropped from 4.5 hours to 2.7 hours. The reduction was driven by improved patient navigation to appropriate settings rather than increased ED capacity — the same physical resources handled demand more effectively when appropriately directed.
- 55% of health line inquiries resolved by AI without nurse involvement. Average caller hold time dropped from 20+ minutes to under 4 minutes. Nurses reported that their call mix had shifted materially toward clinical assessment — the work their training was designed for.
- No-show rate cut from 18% to 9%. Approximately 12,000 appointment slots annually were recovered. Clinics reported that scheduling efficiency improved significantly as the predictability of actual attendance increased. Waitlist patients accessed care more quickly.
- Patient satisfaction: health line improved from 67% to 84%. Emergency department satisfaction improved from 52% to 71%, driven primarily by reduced wait times. Post-visit surveys indicated that patients who had been navigated by the AI health line assistant reported satisfaction levels equivalent to those who had spoken with a nurse for informational queries.
- Staff experience improved across all three affected groups. Nurse health line staff reported improved job satisfaction — their work had shifted toward clinical calls that engaged their expertise. ED nursing staff reported that the demand smoothing from Wave 2 had reduced the most acute peak-period pressure. Administrative staff managing appointment scheduling reported reduced manual intervention as the AI system handled routine rescheduling and waitlist management.
Key Lessons
1. System-level thinking is non-negotiable in healthcare. The three components of this intervention were individually useful. But the compounding effect across the care continuum — health line triage reducing inappropriate ED demand, appointment management recovering clinic capacity, demand forecasting enabling resource optimization — produced results that no single-point intervention could have achieved. AI in healthcare must be designed for the system, not for the department.
2. Privacy and clinical safety as design constraints, not compliance requirements. Both PHIPA compliance and clinical safety protocols were established before technical specifications were written. This is the correct sequence. Projects that build first and assess compliance later invariably face costly redesign. Projects that establish clinical governance before implementation invariably find that safety constraints improve the solution — because they force precision about what the system should and should not do.
3. Clinical governance is the implementation partner that matters most. The AI triage protocols in Wave 1 were not written by Remolda engineers. They were written in collaboration with emergency physicians, nurse practitioners, and the network's clinical governance committee, then validated against clinical outcome standards. The Remolda team built the system that executed those protocols. The distinction is critical — and it is the reason the system has operated without a single inappropriate escalation failure in production.
For health networks and healthcare organizations looking to improve patient flow and reduce demand misalignment, see Remolda's healthcare AI capabilities, our predictive analytics services, and how we approach AI chatbot deployment for healthcare.