AI in clinical practice in 2026 splits cleanly into two categories: FDA-regulated diagnostic devices (radiology AI, ECG analysis, retinal imaging) and clinician-assistive tools (scribes, decision support, knowledge retrieval). The first is regulated like a device; the second is a productivity tool. Both matter; they have different rules and different ROI.
What changed in 2026
- AI scribes (ambient documentation) hit ~80% adoption at large health systems for outpatient encounters. Time saved is real and measurable.
- FDA-cleared diagnostic AI crossed 1000+ approved devices, with deep integration in radiology workflows.
- Clinical reasoning tools matured but remain assistive, not autonomous — Glass Health and similar generate differentials and plans for clinician review.
AI scribes — the practical workhorse
Tools like Heidi, Abridge, Suki, Nabla listen to the encounter, transcribe, and generate structured notes. Time saved per clinician: 1-2 hours/day in ambulatory settings.
Heidi: Australia-origin, strong global market presence, multilingual.
Abridge: US enterprise leader; deeply integrated with Epic; deployed at major academic medical centers.
Suki: voice-first; works across specialties; broad EHR integration.
Nabla: strong in primary care; European GDPR-strong story.
Cost: $99-300/clinician/month; enterprise pricing scales with volume.
ROI: if clinicians save 1.5 hours/day at $200/hour effective rate, that's ~$60K/year of capacity per clinician. The tool pays back ~50x.
Glass Health and clinical reasoning
Glass Health is the leading "AI co-pilot" for clinical reasoning. Input symptoms, history, exam findings; get a differential diagnosis, recommended workup, and treatment plan — for clinician review.
Cost: $30/clinician/month individual; enterprise custom.
Best at: complex cases, second opinions, training residents, augmenting decision-making in time-pressed environments.
Sharp edge: outputs are suggestions, not directives. Used as decision support, not automation.
FDA-cleared imaging AI
Hundreds of FDA-cleared algorithms now run inside radiology PACS workflows: detecting lung nodules on CT, brain hemorrhage on head CT, atrial fibrillation on ECG. Most large hospital systems run multiple in parallel.
Examples: Aidoc (acute findings triage), Viz.ai (stroke detection + workflow), HeartFlow (coronary artery analysis).
Integration: auto-flag findings; route urgent results; generate preliminary reads. Don't replace the radiologist — augment.
Cost: Subscription, typically $5-50K/year per algorithm depending on volume; enterprise bundles available.
What's regulated
FDA-cleared (device): any tool that diagnoses or monitors disease. Subject to 510(k) or De Novo pathway.
Not regulated as device: documentation tools (scribes), administrative tools (coding, scheduling), patient communication tools — assuming they don't cross into diagnosis or treatment recommendations.
Gray area: clinical decision support tools that "recommend" diagnosis or treatment. FDA's 2024 guidance clarified some — generally exempt if a clinician can "independently review" the recommendation. Glass Health and similar fit this exemption.
What works for which clinical setting
| Setting |
Best AI tool category |
| Primary care ambulatory |
Scribe + clinical reasoning support |
| Hospital inpatient |
Scribe + radiology AI + sepsis early-warning |
| Emergency department |
Triage AI + imaging AI + scribe |
| Specialty clinic |
Scribe + specialty-specific decision support |
| Telehealth |
Scribe (real-time captioning + summary) |
Common implementation mistakes
Skipping the workflow integration question. AI scribes that don't push to your EHR are inferior to ones that do. Integration depth matters more than transcription accuracy past ~95%.
Underweighting clinician change management. Scribes save time; clinicians who don't trust them undermine ROI. Pilot with champions before mass rollout.
Missing privacy details. All HIPAA-eligible tools sign BAAs; verify recording consent flows for your state.
When AI medical tools shine
High-volume documentation. Scribes pay back fastest at high-volume practices.
Pattern detection in imaging. Radiology AI catches things; helpful adjunct.
Knowledge retrieval. Real-time clinical knowledge at the point of care.
Coding and billing. Auto-coding from notes with audit trails.
When they fall short
Rare diseases. AI is trained on common cases; rare presentations get missed or misdiagnosed.
Ambiguous symptoms. "Tired" + "abdominal pain" + "stress at work" — AI flounders where experienced clinicians don't.
Patient communication. AI scribes capture content; the human judgment about how to communicate to the patient remains the clinician's.
FAQ
Are AI scribes safe with HIPAA?
All major vendors sign BAAs. Verify state consent rules — some require explicit recording consent.
Will AI replace doctors?
Some tasks, yes (documentation, image triage). The doctor-as-decision-maker is more secure than the doctor-as-documenter. Specialty roles change differently — radiology will see more workflow change than psychiatry.
Can AI prescribe?
No. Prescription requires a licensed prescriber. AI can suggest; humans decide.
Where to go next
For related deep dives see AI for doctors in 2026, AI research paper tools in 2026, and AI legal research tools in 2026.