AI for Healthcare: What Actually Works in Production

Healthcare is one of the highest-value verticals for AI. It is also one of the most unforgiving. Here is what I have seen work in real deployments and what tends to fail.

What Works Right Now

Clinical document RAG and KAG. The most proven healthcare AI use case. Physicians spend 35-40% of their time on documentation and record review. RAG-based systems can cut this significantly by surfacing relevant prior notes, lab results, and referral history at the point of care. For complex cases involving contradictory information across multiple visits, KAG (Knowledge-Augmented Generation with graph databases) outperforms flat RAG significantly.

Patient intake automation. LLM agents handling initial patient intake — collecting symptoms, history, insurance information — work well when scoped tightly. The key is keeping the LLM in a structured extraction role rather than a diagnostic role. Extract, do not diagnose.

Scheduling and follow-up via voice AI. Autonomous voice agents (built on Twilio and ElevenLabs) handling appointment reminders, rescheduling, and post-discharge follow-up calls reduce no-show rates and free up front-desk staff. I have built these for clients processing 10,000+ calls per month at a cost of under $0.05 per call.

Contract and compliance review. Healthcare contracts, BAAs, and insurance agreements are a natural fit for LLM-powered review. Reducing review time by 60% on routine documents is consistently achievable. This is one of the clearest ROI stories in the vertical.

The Compliance Layer is Non-Negotiable

Every healthcare AI system needs a PII masking and audit layer before data touches any external model. This is not optional. PHI must be identified, masked or pseudonymized, and the masking must be logged and reversible for authorized users. I build this as a dedicated pipeline stage rather than relying on prompt instructions — prompts can be bypassed, a masking layer cannot.

What Tends to Fail

Diagnostic AI deployed without a human-in-the-loop. Summarization systems that hallucinate clinical details. Chatbots given too broad a scope that end up giving medical advice. Any system where the failure mode is a missed diagnosis or a wrong medication dose is a system that needs human review at the critical juncture — AI should assist, not decide.