AI in Healthcare That Actually Moves the Needle
This episode of the MedCity Pivot Podcast, presented by Abarca Health, cuts through hype and gets specific about where AI helps and where it can hurt. The conversation features Serge Perras, EVP and CTO at Abarca Health; Ton Roelandse, Chief Technology and Innovation Officer at Trexin Consulting; and Bertil Chappuis, Co-founder and CEO of Xtillion.
Below is a concise field guide based on the discussion: practical uses you can deploy now, and the guardrails you need to prevent medical errors.
Where AI Delivers Real Value Today
- Medication safety and decision support: Flag high-risk drug interactions, dose adjustments for renal or hepatic function, and duplicate therapies. Keep the final call with the clinician.
- Prior authorization automation: Draft evidence summaries, map policy criteria, and route exceptions to experts. AI prepares; humans approve.
- Pharmacy benefits (PBM): Predict adherence risk, suggest lower-cost clinically equivalent options, and surface real-time benefit checks at the point of prescribing.
- Claims and revenue cycle: Coding suggestions with confidence thresholds, automated documentation review, and fraud/waste/abuse signals for audit.
- Clinical documentation and patient comms: Ambient scribing, after-visit summaries, and multilingual discharge instructions with clinician verification.
- Care management: Risk stratification for outreach, social needs flags from notes, and scheduling prompts that reduce no-shows.
Guardrails That Prevent Medical Errors
- Human-in-the-loop by policy: Define decision rights, escalation paths, and which outputs are "assistive only." No autonomous care decisions.
- Data governance and PHI controls: Minimum necessary access, de-identification for model training, encryption, and full audit logging.
- Model validation before go-live: Prospective testing against real workflows, subgroup performance checks for bias, and predefined acceptance thresholds.
- Monitoring and incident response: Drift detection, quality dashboards, rollback/kill switch, and a clear playbook for issues.
- Transparency: Label AI-generated notes and recommendations. Keep model cards and documentation current. Tell patients when AI assists their care.
- Change management and training: Short, role-based training, usage guidelines, and competency checks. Reduce alert fatigue with tight thresholds.
- Security and vendor oversight: Require SOC 2 or equivalent, pen tests, SBOMs, and clear data use terms. Tighten BAAs and third-party risk reviews.
- Regulatory alignment: Map use cases to FDA and ONC expectations for decision support and SaMD. Useful references: FDA AI/ML in SaMD and NIST AI Risk Management Framework.
How the Guests Frame the Opportunity
Serge Perras (Abarca Health): Focus on friction reduction across the pharmacy experience-cleaner claims, smarter formulary decisions, real-time price transparency, and adherence prediction that triggers the right intervention at the right time.
Ton Roelandse (Trexin Consulting): Start small inside a single workflow slice. Prove ROI with clean data and tight metrics, then scale. Technology is the easy part-governance and change management make or break results.
Bertil Chappuis (Xtillion): Build a modular AI stack with privacy-first design. Use retrieval-augmented generation tied to vetted clinical sources, enforce strict prompting rules, and log everything. Engineering for reliability matters more than fancy features.
A 90-Day Action Plan for Healthcare Teams
- Pick one use case: Example: prior auth draft letters for a single specialty. Define 2-3 success metrics (turnaround time, approval rate, clinician time saved).
- Form a cross-functional squad: Clinician lead, compliance, pharmacy or revenue cycle lead, data engineer, and an operations owner.
- Write the rules: What the AI may do, what it may not do, who approves outputs, and how exceptions get handled.
- Pilot in a sandbox: Use de-identified data, test against a gold-standard set, then run a controlled live trial.
- Train and tune: Short role-based training, feedback loops, and alert threshold tuning to avoid noise.
- Review and decide: If metrics clear the bar and no safety signals appear, scale. If not, stop and reassess.
There's real value on the table, but only if you pair clear use cases with firm guardrails. The full conversation offers more detail and practical examples from PBM, consulting, and engineering leaders.
If your team needs structured upskilling on AI by job role, explore curated options here: AI courses by job.
Your membership also unlocks:
