AI for Doctors
Clinical decision support, ambient scribing, chart summarization, patient Q&A — the AI tools and LLMs that fit real clinical practice in 2026.
Quick answer
For most clinicians, an ambient scribe (Abridge, Nuance DAX, or Nabla) paired with Claude Sonnet 4 or GPT-4o for chart summary and patient Q&A handles 80% of non-diagnostic load. Expect $100-300/seat/month for HIPAA-compliant commercial tools. Never let the model deliver a diagnosis unreviewed — clinical decision support must stay physician-in-the-loop.
The problem
Physicians spend 2 hours on documentation for every 1 hour of patient care. Charts are long, guidelines shift yearly, and burnout is epidemic. The right AI stack — ambient scribe, chart summarizer, and guideline lookup — buys back 60-90 minutes a day without introducing liability or HIPAA leaks. The wrong stack hallucinates diagnoses or quietly absorbs PHI into third-party training sets.
Core workflows
Ambient scribing
Transcribe and structure visit conversations into SOAP notes, orders, and billing codes. Saves 60-90 minutes per clinician per day.
Chart summarization
Collapse years of EHR history into a 1-page pre-visit summary. Long-context models handle multi-hundred-page charts in one pass.
Clinical guideline lookup
Ground answers in UpToDate, NICE, and specialty-society guidelines with citations. Retrieval-based to prevent hallucinated dosing.
Patient portal triage
Draft replies to inbox messages, route urgent concerns, flag red flags for physician review. Cuts inbox time 40-60%.
Billing + coding
Suggest ICD-10/CPT codes from notes, flag missing documentation, optimize level-of-service capture.
Top tools
- abridge
- nuance-dax
- nabla
- openevidence
- doximity-gpt
- suki-ai
Top models
- claude-sonnet-4
- claude-opus-4
- gpt-4o
- gemini-2-5-pro
FAQs
Is AI scribing actually HIPAA compliant?
Yes — if you use a BAA-signed vendor (Abridge, Nuance, Nabla, Suki). They run on private deployments with audit logs and no training on PHI. Direct-to-OpenAI or -Anthropic consumer APIs are not HIPAA-compliant without enterprise agreements and PHI-handling configurations.
Can AI make a diagnosis?
No — not unsupervised. FDA classifies autonomous diagnostic AI as a Class II/III medical device requiring clearance. What current tools do legally is differential generation and decision support: the physician remains the decision-maker. Tools marketing 'AI diagnosis' without FDA clearance are skating a regulatory edge.
How much time does ambient scribing actually save?
Published studies at Kaiser and Epic-integrated systems show 60-90 minutes per day reclaimed, 30-40% reduction in after-hours documentation, and meaningful burnout score improvements. Specialty varies: primary care sees the biggest gains; procedural specialties see less.
Which LLM is best for clinical use?
Claude Opus 4 leads on long-context chart reasoning; GPT-4o is the safest generalist with broad fine-tuning on medical content; Gemini 2.5 Pro handles massive (2M-token) chart dumps. Use specialized tools with medical guardrails for production — not direct API calls.
What about hallucinated medications or doses?
Real risk. Use tools that force retrieval against formularies (Lexicomp, FDA orange book) and require physician confirmation on any prescription. Never let an LLM freely generate dosing — always have it suggest from a constrained, validated list.
Will patients know AI is involved?
Best practice (and increasingly required by state law) is disclosure. AMA guidance says patients should be informed when AI is used in their care. Most scribes pop a consent at visit start; most portal-triage systems mark AI-drafted replies.
What's the ROI for a practice?
A primary care physician reclaiming 90 minutes/day at $3-5 per minute of billable capacity = $200-400/day. Most practices break even on a $200/month scribe tool within the first week of the month.