How AI Can Fix Specialty Margins, Reroute Referrals, and Reset Vendor Deals - Insights from Sean Cassidy, CEO of Lucem Health
Hospitals are packed. Specialists are booked out. And too many low-acuity cases are clogging high-value service lines. Sean Cassidy, CEO and cofounder of Lucem Health (founded with the Mayo Clinic Platform), argues that AI can front-load care with smarter identification, earlier diagnosis, and tighter routing - so specialists spend their time on the cases that truly need them.
The core problem: specialty time is going to the wrong patients
Specialty clinics are seeing a wave of patients who could be managed in primary care. It burns capacity, stalls access, and pushes specialists into work below the top of their license. That's bad for margins and worse for outcomes.
Cassidy's view is simple: use AI to detect disease risk signals in the EHR and build an upstream process that routes people into the right pathway based on stage, progression, and acuity. Do the heavy lifting before they hit the specialty clinic.
What "front-ending" care with AI looks like
- Surface hidden risk: Mine structured and unstructured EHR data for subtle patterns - labs, vitals, meds, notes - that indicate disease risk earlier.
- Proactive outreach: Don't wait for the patient to call. Trigger outreach with clear next steps and easy scheduling.
- Accelerated diagnosis: Move qualified patients quickly to confirmatory tests and specialty review where warranted.
- Right-sized pathways: Route lower-acuity patients to primary care programs; reserve specialist time for complex cases.
The result: higher-yield visits, earlier treatment, and improved margins. Fee-for-service lines see more complex, appropriate care. In capitated or risk-bearing models, earlier intervention reduces total cost over time.
Flipping the referral dynamic with AI-driven identification
Today's referral flow is mostly reactive. Patients wait, worry, then call. Even proactive outreach (like screening reminders) gets ignored too often. That delay leads to later-stage disease, harder care coordination, and higher cost.
An AI layer between primary and specialty care changes that. It identifies likely disease earlier, triggers outreach, and sends patients to the setting that can confirm and act - without bouncing them around. Specialists see fewer low-acuity cases. PCPs get clarity and support, not an extra burden.
- PCP experience improves: Clear criteria, fewer back-and-forths, and faster access for the right patients.
- Specialty throughput increases: Slots go to cases with higher clinical need and financial impact.
- Patients move faster to answers: Less friction, fewer touchpoints, earlier treatment.
Why this boosts margins (and outcomes)
- Case mix shifts up: More complex cases per specialist hour.
- Stage shift happens: Earlier diagnoses decrease downstream spend and complications.
- Operational waste drops: Fewer inappropriate referrals, fewer redundant visits.
Shared-risk AI models: how vendor deals will change by 2026
For years, health systems paid for software access and usage - and held all the risk. Cassidy expects that to flip. Health systems will tie payment to measurable outcomes: clinical yield, throughput, cost savings, or margin impact.
Vendors will need to prove value, not pitch features. Those that deliver repeatable results will become long-term partners. Those that can't, fade out. This also lowers adoption barriers for providers: fund innovation directly from incremental revenue or realized savings, not big upfront checks.
- What to ask for: Outcome-based pricing, shared upside, clear SLAs on PPV/sensitivity, transparent reporting.
- What to avoid: Pure seat licenses or volume-based fees with no accountability to results.
- What to expect: By the end of 2026, shared-risk relationships become the default for serious AI deployments.
Implementation playbook (start here)
- Pick 2-3 high-value lines: Cardiology, hepatology, GI are common wins. Define target diseases and inclusion criteria.
- Stand up an AI "front door": Ingest EHR data, validate models prospectively, and set triage thresholds with clinical leaders.
- Build the care pathways: Primary-care-managed track, accelerated diagnostics track, and specialty fast lane.
- Operationalize outreach: Multichannel messaging, scheduling links, and diagnostics access within days, not weeks.
- Reserve specialty capacity: Hold blocks for AI-identified patients with high-acuity flags.
- Align PCPs: Clear handoffs, feedback loops, and simple notes/templates to reduce extra work.
- Governance and safety: Bias checks, calibration by subpopulation, human-in-the-loop confirmation, and ongoing monitoring.
- Measure and tune monthly: Stage shift, time-to-diagnosis, specialty no-shows, margin per slot, and PMPM in risk lives.
Metrics that matter to the C-suite
- Clinical: Earlier-stage diagnoses, adherence to guideline-concordant care, readmissions avoided.
- Operational: Specialist throughput, referral quality, diagnostic turnaround time, care pathway cycle time.
- Financial: Gross margin per specialty slot, incremental revenue in FFS lines, PMPM savings for capitated populations.
Guardrails you shouldn't skip
- Equity: Validate performance across age, race, language, and SDOH cohorts; adjust thresholds to avoid disparities.
- Privacy and consent: HIPAA compliance, clear patient communication, and auditable access controls.
- Safety net: Human review for high-consequence decisions, rapid escalation routes, and model drift monitoring.
Bottom line
AI can increase specialty margins and improve outcomes when it sits upstream of referrals, finds risk early, and routes patients to the right place the first time. Pair that with shared-risk vendor contracts and you get accountability, faster ROI, and less financial exposure. The playbook is straightforward - pick the lines, set the pathways, measure relentlessly.
Resources
- Mayo Clinic Platform - background on clinical AI collaboration and deployment.
- Complete AI Training: Courses by Job - practical training paths for clinical, ops, and analytics teams.
Your membership also unlocks:
