AI tools for doctors in 2026: an honest guide by category

See also our free medical calculators, each with its primary references.

Physicians are adopting AI quickly. In an AMA survey of nearly 1,200 physicians, 66% reported using health care AI in 2024, up from 38% in 2023.^[1] The pressure behind that is not novelty. In a time-and-motion study across four specialties, physicians spent 27.0% of their time on direct clinical face time and 49.2% on EHR and desk work, with diary data from a subset showing 1 to 2 hours of after-hours work each night.^[2] In primary care, EHR event logs showed 5.9 hours of EHR work in an 11.4-hour workday, 1.4 hours of it after clinic.^[3]

So the honest guide starts from the burden, not the tools. AI tools for doctors fall into five categories, and each carries a different part of the load.

1. Ambient scribes: the documentation load

What they do. Listen to the visit and draft the note. Examples include Microsoft Dragon Copilot, which ambiently captures multi-party conversations at the point of care; Abridge, which turns clinical conversations into documentation; Freed, with plans starting at $39 a month; Heidi Health, which offers a free seat; and Nabla.^[4]

The evidence. This is now the best-studied category. In a randomized trial of 238 outpatient physicians, one scribe (Nabla) reduced time-in-note by 9.5% versus usual care, while the other (DAX Copilot) showed no significant change; clinicians reported clinically significant inaccuracies "occasionally" with both.^[5] In a pragmatic randomized trial of 66 practitioners, ambient AI reduced work exhaustion and time on notes (by 0.36 hours a day) but did not significantly increase professional fulfillment.^[6] A single-group study at one academic health system found 20.4% less time in notes per appointment and 30.0% less after-hours work.^[7]

The limit. A scribe documents what was said. It does not decide what to do next, and its drafts need review, because inaccuracies do occur.

2. AI reference and Q&A: the knowledge load

What they do. Answer a clinical question in natural language with citations. OpenEvidence is free for verified US healthcare professionals and has content agreements with publishers including NEJM, JAMA, Nature, NCCN and Cochrane.^[8] UpToDate Expert AI is generative AI built on UpToDate content and, according to Wolters Kluwer, is now included in all US UpToDate subscriptions.^[9] ClinicalKey AI from Elsevier combines its evidence content with conversational search, with each response citing published evidence; individual pricing starts at $349 a year.^[10]

The evidence and the limit. These tools are only as good as their grounding. General-purpose chatbots, used without grounding, have been shown to fabricate references: in one study 47% of ChatGPT-3.5 references were fabricated,^[11] and in another 55% of GPT-3.5 citations versus 18% of GPT-4 citations were fabricated.^[12] That is the reason grounded, cited tools exist, and the reason to click through to the source before acting on any answer.

3. Differential diagnosis: the diagnostic load

What they do. Take findings and return candidate diagnoses. Isabel's DDx Companion covers more than 10,000 conditions and costs $149 a year for individuals.^[13] Glass Health combines AI decision support with ambient scribing, with a free tier.^[14]

The evidence and the limit. On 70 difficult published cases, GPT-4 included the final diagnosis in its differential 64% of the time.^[15] But in a randomized trial, giving physicians an LLM did not significantly improve their diagnostic reasoning compared with conventional resources.^[16] A differential is a list of possibilities, not a plan. See our longer piece on differential diagnosis generators for detail.

4. Decision support that returns next steps: the decision load

What it does. Reads the note or chart and returns what to do next, as structured items a chart can act on: labs, imaging, referrals, escalation criteria. This is the category Pogosh is in. Pogosh returns labs with LOINC codes, imaging with RadLex codes, specialist routing adapted to the setting and the patient's age and sex, escalation criteria, a separate strict emergency check, and a note draft that does not invent negatives. It accounts for what the note says was already done. It is a REST API with a HIPAA business associate agreement on AWS, and an OpenEMR module; pricing starts at $149 a month.

The limit. Pogosh is not FDA-cleared, has no published independent validation yet, and is not an evidence library. Everything it returns is advisory and can be overridden. If you need to read and cite the evidence, the reference tools above are better at that.

5. Calculators: the arithmetic load

What they do. Compute validated scores. MDCalc is free for clinicians on the web and in its apps, with content written by physicians or pharmacists, and offers an EHR integration for health systems.^[17] It is the established free reference, and for looking up a score by name it is hard to beat.

Where AI changes it. The arithmetic was never the hard part. Remembering that a score applies, at the right moment, with the right version, is. Pogosh offers relevant calculators automatically from the note (Wells and PERC for suspected PE, HEART for chest pain, MELD 3.0 for cirrhosis), with each one's references, and never runs them on its own. You can browse them all on our free calculators page.

How to choose

Most clinicians will end up using more than one. The sensible test for any of them is your own cases, with a skeptical eye on every output, and a BAA before any patient identifiers go in.