Important disclaimer

AI tools in healthcare require careful evaluation of clinical validation, regulatory approval (FDA clearance in the US, CE marking in Europe), data privacy compliance (HIPAA, GDPR), and institutional policy. General AI chatbots like ChatGPT and Claude are not validated for clinical decision support and should not be used for direct patient care decisions. The tools in this guide are categorised by appropriate use case.

Clinical documentation (highest-impact category)

Nuance DAX Copilot (Microsoft) — best AI clinical scribe

Nuance DAX Copilot is the leading ambient AI documentation tool — it listens to patient-clinician conversations and automatically generates clinical notes in the physician's EHR. Studies show clinicians using DAX save 2–3 hours of documentation time per day. Integrated with Epic, Cerner, and other major EHR systems. FDA-cleared for clinical documentation use in the US.

Best for: Physicians, NPs, and PAs drowning in documentation burden.
Pricing: Enterprise (health system level) — contact for pricing

Suki AI — best for independent practices

Suki is an AI voice assistant for clinical documentation that works with smaller practices that cannot afford enterprise Nuance deployments. It transcribes patient encounters, structures notes, and integrates with over 25 EHR platforms. Faster to deploy and more accessible for independent and group practices.

Pricing: From $299/provider/mo

Medical research and literature

Elicit — best for literature review

Elicit searches and synthesises academic papers, extracting key data (sample sizes, methodologies, outcomes) across studies. For physicians doing evidence-based literature reviews, Elicit dramatically accelerates the research process. Not suitable for direct clinical decision support — it is a research acceleration tool.

Pricing: Free / $10 Basic / $42 Plus

NotebookLM — best for synthesising clinical guidelines

Upload clinical guidelines, protocols, and reference documents to NotebookLM and query them in natural language. "What does the AHA guideline say about anticoagulation in AF patients over 75?" becomes a seconds-long query rather than a manual search. Keep patient data out of any non-HIPAA-compliant tool.

Pricing: Free

Administrative and operational AI

Otter.ai or Fathom — for administrative meetings

For non-clinical meetings (management, departmental, administrative), AI meeting tools are safe to use and deliver immediate time savings on note-taking and follow-up drafting.

Claude / ChatGPT (with appropriate data handling) — for non-PHI drafting

For drafting patient education materials, administrative policies, grant applications, and research summaries (without patient-identifiable data), general AI assistants are appropriate and highly effective. Use enterprise plans with data privacy guarantees and remove all PHI before inputting any clinical content.

AI tools for lawyers — another high-stakes profession →

Why healthcare needs a deliberate AI stack in 2026

Healthcare is simultaneously the highest-value and the highest-risk AI adoption environment. The upside is enormous — ambient scribes are recovering 2-4 hours a day of physician time, AI-assisted imaging is catching findings human radiologists miss, and patient messaging tools are saving clinicians from EHR inbox burnout. But the risks are also real: HIPAA violations, biased model outputs, deskilling, and errors that harm patients. The dividing line between good and bad AI deployment in healthcare is whether the tool is HIPAA-compliant, clinically validated where appropriate (some require FDA clearance), and integrated so that it augments clinician judgement rather than replacing it.

What changed in 2026: ambient clinical documentation is now mainstream across large health systems, FDA has cleared hundreds of AI-based medical devices, and most major EHRs (Epic, Cerner/Oracle Health) have baked generative features into their workflow. The shift in focus has moved from "should we pilot AI" to "how do we roll it out to 10,000 clinicians without burning them out on yet another tool." Below are the five categories that matter and the HIPAA-aware playbook for each.

The five categories of a healthcare AI stack

1. Clinical documentation (ambient scribing): capture notes during visits. 2. Medical imaging and diagnostics: FDA-cleared AI for radiology and pathology. 3. Patient communication and inbox: drafting replies, triage, scheduling. 4. Medical research and literature: summarising papers, staying current. 5. Administrative and operations: coding, billing, prior authorisation.

Clinical documentation (ambient scribes)

Abridge (Enterprise — contact for quote): One of the most widely adopted ambient clinical documentation platforms in 2026, rolled out across Mayo Clinic, Kaiser, and many Epic customers. Records the visit with patient consent and produces a structured SOAP note pushed back to the EHR. Best for: hospitals and large health systems wanting EHR-integrated documentation. Limitations: enterprise only; implementation depends on your EHR.

Suki AI (Enterprise — contact for quote): Voice-first AI assistant that writes notes, updates orders, and interacts with the EHR. Embedded across major EHRs. Best for: health systems and mid-size group practices. Limitations: enterprise pricing.

Nuance DAX (Microsoft) (Enterprise — contact for quote): The Microsoft/Nuance ambient documentation platform, deeply integrated with Epic and the Microsoft Cloud for Healthcare. Used by hundreds of health systems. Best for: Epic-heavy systems wanting the most mature integration. Limitations: enterprise contracts; legacy pricing structure.

Augmedix (Enterprise — contact for quote): Another enterprise ambient scribe used widely in the US, with both fully automated and human-in-the-loop options. Best for: health systems that want a mix of AI and human review. Limitations: enterprise pricing.

Whisperflow or MacWhisper (various, $10-$40): For solo practitioners and small clinics without an enterprise ambient contract, local voice-to-text using Whisper-based tools can work — but only for non-PHI workflows like dictating personal notes. Never feed PHI into consumer Whisper cloud services.

Medical imaging and diagnostics (FDA-cleared)

Aidoc (Enterprise — contact for quote): FDA-cleared AI for radiology that flags acute findings — pulmonary embolism, intracranial haemorrhage, aortic dissection — in near real-time. Used widely in US hospitals. Best for: emergency and radiology departments. Limitations: requires integration with PACS and radiology workflow.

Imagen Technologies (Enterprise — contact for quote): AI imaging for primary care and orthopaedic practices — interpreting X-rays with FDA clearance in specific use cases. Best for: primary care, urgent care, and ortho groups. Limitations: narrower scope than hospital-grade platforms.

PathologyWatch (Enterprise — contact for quote): Digital pathology with AI-assisted case review. Best for: pathology practices moving to digital workflow. Limitations: requires slide digitisation infrastructure.

Patient communication and inbox AI

Microsoft 365 Copilot for Healthcare ($30/user/mo add-on): For administrative staff and non-PHI-facing workflows, Copilot inside Outlook, Word, and Teams — with Microsoft's BAA in place — is the safest general-purpose AI for healthcare orgs. Best for: administrative staff, operations, and non-clinical writing. Limitations: must confirm your license includes a BAA before any PHI exposure.

Epic-embedded GPT (via Epic's own AI partnerships): Many health systems now use Epic's built-in GPT integration for inbox replies — the AI drafts a response to patient MyChart messages, the physician reviews and sends. Measurable ROI on clinician burnout. Best for: Epic customers. Limitations: must be enabled by your Epic admin.

Claude with enterprise BAA: For health systems that have a signed BAA with Anthropic (available on Enterprise plans), Claude is a strong option for drafting non-diagnostic clinical communications. Always route outputs through a clinician before sending. Best for: administrative and communication use cases with appropriate contracts. Limitations: free and consumer tiers do not include BAAs — never use those for PHI.

Medical research and literature

OpenEvidence / ChatGPT Enterprise (Enterprise): For physicians doing point-of-care research, OpenEvidence (built on GPT-4) and similar tools provide citation-grounded answers from peer-reviewed literature. Safer than general ChatGPT for clinical questions. Best for: physicians doing clinical literature review. Limitations: still verify against primary sources for critical decisions.

Consensus Try Consensus → (Free, Premium $11.99/mo, Enterprise): AI search grounded in peer-reviewed papers — useful for quickly checking what the literature says on a topic. Best for: clinicians and researchers doing exploratory literature search. Limitations: coverage depends on database; check primary sources for decisions.

Elicit (Free, Plus $12/mo, Pro $42/mo, Enterprise custom): AI research assistant for literature review with extracted findings and methodology comparison. Best for: clinical researchers and evidence-based medicine work. Limitations: not a substitute for systematic review.

Perplexity Pro ($20/mo): For fast, cited answers on general medical or regulatory questions. Best for: staying current and quick fact-checking. Limitations: always verify against primary sources for clinical decisions.

Administrative and operations AI

Merative (formerly IBM Watson Health) (Enterprise): AI for medical coding, revenue cycle, and population health analytics. Best for: hospital revenue cycle and coding teams. Limitations: enterprise-scale investment.

Ultimate AI / Notable (Enterprise): AI for prior authorisation, scheduling, and administrative automation. Best for: administrative and operations teams. Limitations: enterprise implementation.

HIPAA, BAAs, and compliance: the non-negotiable rules

Healthcare AI has one non-negotiable rule: never put PHI (Protected Health Information) into any tool that doesn't have a Business Associate Agreement (BAA) in place. A BAA is a contract under HIPAA that obligates the vendor to handle PHI to HIPAA standards. Consumer ChatGPT, free Claude, free Gemini, and most consumer AI tools do not come with a BAA — using them with PHI is a HIPAA violation per se, regardless of outcome. Enterprise tiers that can provide a BAA include Microsoft 365 Copilot (via Microsoft Cloud for Healthcare), Google Cloud Healthcare API, AWS Bedrock with healthcare contracts, ChatGPT Enterprise (with specific healthcare terms), and Claude Enterprise. Additionally: FDA clearance is required for any tool that makes diagnostic claims — "AI detects pulmonary embolism" is a medical device. Validate every clinical-facing tool against the FDA's list of cleared AI/ML-enabled medical devices, and confirm your institution's risk management and compliance teams have approved the specific use case before clinical rollout.

How to build your healthcare AI stack: solo, clinic, enterprise

Solo practitioner ($100-$300/mo plus admin time): A lightweight stack of Perplexity Pro ($20) + Elicit Plus ($12) + ChatGPT Enterprise or Claude Enterprise (requires contract) for non-PHI research + a single ambient scribe contract (enterprise pricing). Solos typically partner with their EHR vendor for clinical AI rather than building their own stack.

Clinic or group practice (100-1,000 clinicians): One ambient scribe platform (Abridge, Suki, DAX, or Augmedix) + Microsoft 365 Copilot with healthcare BAA + a research tool (Elicit, Consensus, Perplexity Enterprise) + Epic-embedded GPT for inbox. Expect $200-$500/clinician/month in combined AI investment, offset by burnout and time savings.

Enterprise health system (1,000+ clinicians): Ambient scribe at scale + FDA-cleared imaging platform (Aidoc or equivalent) + population health/revenue cycle AI (Merative or similar) + in-house GenAI governance committee + a dedicated Chief AI/Informatics Officer. The pattern is tight governance, limited vendor count, and aggressive training programs.

Common mistakes healthcare orgs make with AI

1. Using consumer AI with PHI. One casual ChatGPT paste of a patient note is a HIPAA violation. Enforce policy with training and technical controls. 2. Skipping clinician adoption. Ambient scribes only save time if clinicians actually use them. Plan for 3-6 months of change management. 3. Ignoring bias and validation. AI trained on non-representative populations can underperform for specific patient groups. Audit for equity. 4. Over-automating triage and inbox. Auto-routing patient messages without clinician oversight is a patient-safety risk. Always keep humans in the loop on clinical decisions. 5. Buying tools without integration. An AI that lives outside the EHR workflow won't get used. Prioritise tools that embed directly in Epic/Cerner/etc.

A day in the life: an AI-enabled healthcare practice

8am: Clinicians start rounds. Abridge runs in the background of every room visit, producing SOAP notes pushed to Epic within minutes of the encounter. 10am: Radiology reviews the overnight batch — Aidoc has pre-flagged 3 acute findings, which the radiologist confirms and prioritises. 11am: A primary care physician sits down to MyChart — Epic's embedded GPT has pre-drafted 15 patient replies based on the message content and patient history; she reviews, edits, and sends in 20 minutes what used to take an hour. 1pm: Clinical researcher uses Elicit to scan 80 new papers and extract key findings for a department meeting. 3pm: Admin team uses Microsoft 365 Copilot (with BAA in place) to draft non-PHI communications, policy documents, and HR materials. 5pm: Clinicians leave on time — for the first time in years. The 2026 healthcare AI stack isn't about replacing clinicians; it's about giving them back their afternoons.

Frequently asked questions

Keep reading → Compare in depth: claude vs gemini.