Healthcare has a paradox: it's one of the most information-intensive industries on earth, and its most expensive professionals spend hours a day doing data entry. For every hour of patient care, clinicians log significant time in the EHR. Staff burn out on prior-auth phone trees, patients wait on hold, and coding backlogs quietly delay revenue by weeks.
That's why the healthcare AI projects actually delivering value in 2026 are mostly administrative, not diagnostic. The bar for "the model suggests a treatment" is — rightly — enormous. The bar for "the model drafts the paperwork a human reviews" is achievable today, and the impact is immediate. Here's where we see it working.
Ambient clinical documentation
The flagship use case: the visit is recorded (with consent), transcribed, and drafted into a structured clinical note — history, exam, assessment, plan — before the clinician leaves the room. The clinician edits and signs; nothing enters the chart without their review. Health systems report meaningful reductions in after-hours "pajama time" charting, and the engineering lessons are transferable: speaker attribution matters, specialty-specific templates matter, and the note must cite the transcript so reviewers can verify any sentence in one click.
Prior authorization: the two-sided document war
Prior auth is a document exchange between provider and payer, and both sides are automating it. On the provider side, an agent assembles the request — pulling the diagnosis, relevant history and payer-specific criteria into a complete submission — and drafts appeals for denials, quoting the payer's own policy language back at it. The design principle mirrors our production-agent playbook: a bounded job, verified retrieval from the actual policy documents, and a human sign-off before anything is submitted.
Intake and triage: the front door that answers
Patient messages, referral faxes (still faxes!), and portal requests pile into queues that nurses work down by hand. Grounded agents now classify and route this traffic: extracting the clinical ask, checking scheduling rules, drafting responses for routine requests, and escalating anything with clinical red flags straight to a human — with conservative, explicitly enumerated escalation triggers. The goal isn't replacing clinical judgment; it's making sure clinical judgment is spent on the messages that need it.
Revenue cycle: coding, claims and denials
Medical coding is a language task with a rulebook — a natural LLM fit, with a catch: coding errors are compliance events. The pattern that works is assist-then-audit: the model proposes codes with evidence linked to the note, coders confirm, and a sampling program continuously measures model-versus-human accuracy by code family. The same machinery works denials from the other end, matching denial reasons against documentation and drafting appeal letters. Days-in-AR is a number CFOs watch; this moves it.
The guardrails that make it deployable
Everything above touches PHI, so the platform work is non-negotiable: BAAs with every model provider in the chain, PHI minimization and redaction before anything leaves your boundary, role-based access mirrored from the EHR, immutable audit logs, and zero data retention on inference. Add domain-specific evaluation — hallucinated medications, wrong-patient references and omitted negatives are the failure modes that matter, and generic benchmarks won't catch them. And every workflow keeps a licensed human as the final signatory. That's not a limitation to engineer away; in healthcare, it's the product.
The short version
Healthcare AI in 2026 pays off where the work is linguistic and administrative: documentation, prior auth, intake, coding. Win there with citations, PHI discipline and human sign-off, and you build both the trust and the infrastructure for the harder clinical use cases later.
If your clinicians are drowning in paperwork or your revenue cycle has a backlog with a name, we should talk.