Nowhere is the gap between AI demo and AI production wider than in healthcare. The demos are magical, the pilots are common, the production deployments are rare. That's not because healthcare is uniquely hard for AI — it's because healthcare has uniquely low tolerance for hand-waving.
Here are five healthcare AI use cases we've seen ship to real production — and five that consistently die in pilot.
SHIPS — 1. Clinical documentation (ambient scribes)
The single biggest healthcare AI win of the last three years. The clinician wears a mic, the ambient LLM transcribes and structures the encounter, the clinician reviews and signs. Products like Abridge, Nabla, Suki and DAX Copilot have deployed to thousands of clinicians.
Why it ships: the value prop is unambiguous (30-90 minutes back per clinician per day), the compliance path is well-trodden (BAA + audit + human sign-off), and the failure mode is graceful (human reviews before signing).
SHIPS — 2. Prior authorization automation
Prior auth is the single most soul-crushing task in US healthcare admin. Modern implementations use LLMs to draft prior-auth submissions from clinical notes, then a human reviewer approves before the X12 278 goes on the wire.
Why it ships: extraction from unstructured clinical notes is exactly what LLMs are good at, the human-in-the-loop step keeps the safety story intact, and the ROI is measured in $/hour of nursing time saved. Companies like Cohere Health and Availity are deployed at scale.
SHIPS — 3. Claim status Q&A
'Has claim 12345 been paid yet? If not, why?' A grounded RAG layer over the X12 837/835/276/277 transaction stream + payer implementation guides answers this in plain English. The user gets citations back to the underlying documents; the ops team gets 4 hours a day back.
This is the same architecture as our Document Intelligence product, applied to healthcare claims instead of contracts. Sovereign, cited, cost-transparent.
SHIPS — 4. Radiology second-read
Not the sci-fi 'AI replaces radiologists'. The reality: computer-vision models flag likely findings on chest X-rays and CTs, a radiologist reviews and either agrees or overrides. FDA-cleared products (Aidoc, Rad AI, Annalise.ai) have real deployments.
Why it ships: narrow scope, clear ground truth (biopsy or follow-up imaging), regulator-cleared, workflow-integrated. The killer app is 'triage' — surfacing likely acute cases to the top of the worklist.
SHIPS — 5. Patient-facing symptom triage
NHS 111 online, Ada Health, Buoy, Symptomate. Not diagnosis — triage. 'Given what you're describing, you should book a same-day GP visit; call 999 if X, Y or Z happens meanwhile'. Deployed to tens of millions of users.
Why it ships: the value is genuine (self-service triage reduces call-centre load), the risk is manageable (always safer-erring), and the human handoff is graceful (escalate to human).
DIES IN PILOT — 1. AI-generated clinical decision support
The 'AI recommends the next diagnostic test / drug' feature. Dies because: no clinician wants to be second-guessed by a black box on a legal-liability-carrying decision, no regulatory path is clean, and the value prop is 'saves the clinician 30 seconds sometimes' when they'd rather save 30 minutes of documentation.
DIES IN PILOT — 2. Fully autonomous scheduling agents
The idea that an AI agent will negotiate appointments across multiple providers autonomously. Consistently dies on integration complexity (every provider's scheduling system is different), consent flows (patients want to see and confirm), and edge cases (insurance verification, prep instructions).
DIES IN PILOT — 3. Predictive readmission modelling
Predict which patients will be readmitted within 30 days. Everyone builds one. Almost nobody deploys it meaningfully. Dies because: predictions get made too late to change behaviour, care teams don't have capacity to act on the prediction, and the model's precision at useful recall is usually just okay.
DIES IN PILOT — 4. Freeform patient chatbots on general LLMs
'Chat with our AI nurse about anything'. Dies because: hallucination + patient safety is a catastrophic combination, the compliance story is impossible to defend, and every deployment gets one bad interaction that makes headlines. See ChatGPT + eating-disorder helpline. Do not do this.
DIES IN PILOT — 5. Fully generative medical imaging
AI generates synthetic imaging for training or teaching. Interesting research, no clear product line. Dies because: nobody trusts synthetic imaging for real clinical decisions, and using it for training runs into the 'was our model trained on hallucinated data' problem.
The pattern
AI ships in healthcare when: (a) there's a clear human review step, (b) the model is narrow-scope with clear ground truth, (c) the workflow already exists and AI accelerates a known task. AI dies in pilot when it aspires to replace clinical judgment or handle open-ended patient interactions without guardrails.
The Origami approach to healthcare AI: pick the boring, high-ROI use cases, ship with evals + human-in-the-loop + audit logs from day one, and let clinicians be the ones who sign anything that reaches a patient chart.