Two very different stories about AI in healthcare are running at the same time. In one, doctors at large health systems use AI scribes in millions of visits, finish their notes before they leave the exam room, and report less burnout. In the other, a 60-year-old man asks a chatbot how to cut salt from his diet, replaces table salt with sodium bromide, and is hospitalized with hallucinations. His doctors cannot even recover the conversation to see what it told him. Both stories are true, and the difference between them is not the model. It is how the AI was put to work.
Where AI is actually working in healthcare
Documentation. Ambient AI scribes are the clearest success so far. The AMA's 2026 survey of 1,692 physicians found 81% now use AI in their practices, more than double the 38% of 2023. At Kaiser Permanente's Permanente Medical Group, 7,260 physicians used AI scribes in about 2.5 million patient encounters between October 2023 and December 2024. UCSF's Robert Wachter has written that about 70% of UCSF doctors with access now use one. The results are measurable. In a multicenter study of 263 clinicians published in JAMA Network Open in October 2025, burnout fell from 51.9% to 38.8% after 30 days with an ambient scribe. A UCSF study in the same journal in January 2026 found adopters produced 1.81 more RVUs and about 0.8 more patient visits per week, with no difference in claim denials.
The front office. In MGMA's February 2026 poll, practices named scheduling (31%) and phone calls (27%) as their top AI and automation priorities. That makes sense. The phone is where a short-staffed practice breaks first, and the same people answering it are usually the ones doing the authorizations.
Prior authorization. The AMA puts the load at about 40 requests and 13 hours of physician and staff time per physician per week. 40% of physicians employ staff who work on nothing else. Since January 1, 2026, the CMS Interoperability and Prior Authorization rule requires Medicare Advantage, Medicaid and CHIP plans to decide standard requests within 7 calendar days and urgent ones within 72 hours, and to give a specific reason for every denial. Payers are speeding up, and many are using AI to do it. 61% of physicians told the AMA they worry payer AI is increasing denials. Practices that still assemble requests by hand are now up against automated systems on the other side.
Where it is breaking
Patients are already using AI, mostly without guardrails. KFF's March 2026 poll found about a third of adults have used AI chatbots for health information or advice, and younger and lower-income adults were more likely to say they did so because care was hard to afford or reach. Only about a third of those users trust the accuracy of what they get, and about three in four Americans are concerned about the privacy of medical information they give to AI. A West Health-Gallup survey released in April 2026 found that 14% of recent AI health users skipped a provider visit they otherwise would have made because of what AI told them, roughly 14 million adults. Only 4% strongly trust the accuracy of AI health information, and 11% of recent users said AI had given them advice they believed was unsafe.
No record, no accountability. The bromide case comes from a new paper in Nature Health by researchers at Binghamton, Stanford, Texas A&M and Indiana University. Their point is that harm from AI health advice is mostly invisible because nobody keeps the record. They call for conversation logs patients can share with their doctors, a way to flag and report harmful answers, and independent monitoring. Those are good requirements for any AI that talks to patients, not just consumer chatbots.
The small-practice gap. Most of the success stories come from large systems with IT departments, enterprise contracts and research teams to measure the results. Price is part of the reason. Physicians quoted in Medical Economics' 2026 coverage of AI scribes describe tools costing around $500 per provider per month, and a small practice cannot afford a separate subscription for every problem.
The question for a practice is no longer whether to use AI. Its patients and its payers already do. The question is whether its own AI is built so you can see, check and stand behind what it does.
The rules are arriving state by state
Congress has not passed a law aimed at AI in healthcare, but the rules are coming anyway. The FDA finalized updated clinical decision support software guidance in January 2026, and it makes clear that software giving recommendations directly to patients rather than clinicians is still a regulated device. California now lets its licensing boards act against AI that implies it holds a healthcare license (AB 489). Under AB 3030, practices that use generative AI to send patients clinical information must include a disclaimer and a way to reach a human, unless a licensed provider reviews the message first. At least seven states passed laws in 2026 limiting how insurers use AI in coverage decisions. Iowa, for example, bars AI from being the sole basis for a denial. Under HIPAA, any AI vendor that touches patient data is a business associate and needs a business associate agreement. HHS has proposed a HIPAA Security Rule update that would require written verification of each business associate's safeguards at least every 12 months. Meanwhile healthcare breaches still cost more than any other industry, at an average of $6.64 million in IBM's 2026 report.
The direction is consistent. Regulators want disclosure, a human in the loop for clinical decisions, records that can be audited, and vendors who can prove how they protect data.
What a practice should do next
Start with administrative work, not clinical judgment. Phones, scheduling, reminders, recalls, intake forms, prior authorization paperwork and visit notes are where AI has a proven return. The clinical decision and the signature stay with the provider.
Make every AI conversation recordable and reviewable. If an agent talks to a patient, the practice should be able to pull up exactly what was said, in minutes. Someone should review a sample every week.
Ground answers in your own approved content. An agent should answer from your policies, schedules, fee schedules and templates. When a question falls outside them, especially anything clinical, it should hand off to a person, not guess.
Know where patient data goes. Ask every AI vendor which models see patient data, where they run, whether a business associate agreement covers each one, and whether the data is used for training.
Buy one platform, not ten point tools. Every separate AI subscription is another vendor to vet, another BAA, another place data lives, and another login to manage when staff turn over.
How Acclivity Labs helps
Acclivity Labs builds and runs AI agents for small and mid-sized organizations, including medical and dental practices. We do not sell a single bot. It is one private platform where the agents share one CRM, one approval model and one audit trail:
Front office. A receptionist and voice agent that answers calls and portal messages, books and reschedules from your rules, chases no-shows and recalls, and hands any clinical question to your staff. Our Scribe agent writes visit and meeting notes, and eSign handles consents and intake forms.
Practice operations. Prior authorization requests assembled from the chart and the payer's checklist, status tracked, appeals drafted for the provider to approve. The agent also handles payer-line holds and reference numbers, referral coordination and patient follow-up from your templates.
Back office, security and growth. Bookkeeping and payer remittance follow-up. The HIPAA risk analysis, multi-factor authentication, tested backups and vendor records regulators expect, kept current with our affiliated managed IT partner. Search and AI-answer visibility from Acclivity Digital Marketing, our affiliated agency, so patients who ask an AI about your specialty and city find you.
Built for accountability. Every conversation is logged where you can search it. Automated messages are checked against your records before they go out. Real conversations are sampled and reviewed every day. Our agents run on our own Nvidia DGX hardware using open-weight models, so patient data does not have to leave private infrastructure. Healthcare work that involves protected health information starts only after a business associate agreement is in place.
The practices that come out ahead with AI will not be the ones with the most tools. They will be the ones whose AI does the paperwork, answers the phone, and leaves a record they can stand behind.