State of Agentic AI EHR 2026: What Works and What's Risky

An agentic AI EHR is an electronic health record (EHR), the software that holds patient charts, with an AI agent inside it or connected to it. "Agentic" means the AI can plan steps and act, not only answer a question or draft a note. In 2026, EHR and AI companies launched such agents. This report shows what was on sale in October 2026, what studies found and what it costs. It also covers the risks and what a small clinic can do this quarter.

Quick Answer

Agentic AI means software that plans steps and uses tools to act, not just to answer or draft. In October 2026, AI scribes, which listen to a visit and draft the note, are on sale and show modest, uneven gains. Of the ten products we list below, six have shipped. Epic AI Charting, for example, queues orders as well as drafting notes. The other four are early or only announced, and the ChatGPT link to Epic is read-only.

Rules differ by state, and the clinician stays responsible for the signed note. So a clinic can buy AI notes today. A few agents that act are also on sale, but the proof for them is thin. Treat any agent that changes a chart as a pilot, meaning a small trial run.

Key numbers

• Use is common, but the count is broad. In early 2026, 81% of physicians surveyed by the American Medical Association (AMA) said they use AI at work (AMA survey via Fierce Healthcare, 2026). That is up from 38% in 2023 (AMA survey via the Medical Group Management Association, 2026). The count includes any professional use, such as summarizing research, so it overstates scribe use. In Doximity's surveys, 29% named voice-based documentation, such as AI scribes, up from 20% (of physicians surveyed in the latest wave) (Doximity, 2025 to 2026).
• On sale now. In August 2026, athenahealth said its built-in AI scribe, athenaAmbient, is broadly available to all clinicians on its athenaOne network at no additional cost. The network has more than 170,000 clinicians, a vendor-reported figure (athenahealth, 2026).
• Not on sale yet. Epic's Agent Factory, a tool for health systems to build their own agents, starts early-adopter training in October 2026, with broader availability planned for 2027 (Epic, 2026).
• Time saved is modest. Across five health systems (2023 to 2025), adopters had 16.0 fewer documentation minutes per 8 scheduled patient hours, and after-hours EHR time did not change clearly (JAMA, 2023 to 2025). In a 2024 to 2025 trial at the University of California, Los Angeles (UCLA), the researchers assigned clinicians by chance. One scribe cut note time by 9.5%, and the other showed no clear change (NEJM AI, 2024 to 2025).
• Errors are real. In a 2024 pilot, 18% of 356 reviewed AI-drafted notes left out something that was said, and 11.5% held a made-up statement. Errors rated serious or imminent risk were in 5.3% of notes (JMIR Medical Informatics, 2024).
• Agents are not proven. In a 2026 Stanford simulation, the best agent finished only 36.3% of paperwork tasks from start to finish (arXiv, 2026).
• Rules differ by state. In 2026 through July 30, 33 of over 280 state health-AI bills had been signed, and none of about 20 liability bills had passed (Manatt Health, 2026). So few proposed rules had become law by that date, and ordinary malpractice law likely applies.
• AI notes inside an EHR. In our calculator, 23 therapy EHRs publish a price, not counting ClinikEHR, the EHR we build. Of those 23, 16 include AI notes at no extra cost or sell them for a stated price. Thirteen charge an add-on, with a median, or middle value, of $35 a month across 13 vendors and a range of $15 to $175 (EHR cost calculator, 2026). Three include them at no extra cost, and Noterro includes only a capped allowance. ClinikEHR also has a capped allowance: 300 audio minutes a month on Essential and 900 on Team.

At a glance

AI notes are on sale with modest gains; agents are still early

• 81%

  Physicians who use AI at work

  AMA survey, early 2026; more than double 2023

• 16.0 min

  Fewer documentation minutes per 8 hours

  Per 8 scheduled patient hours; five health systems, 2023 to 2025

• 18%

  Reviewed notes with an omission

  University of California, Davis pilot, 2024; 356 notes

• 36.3%

  Paperwork tasks the best agent finished

  Stanford simulation, 2026

• $35

  Median AI-notes add-on, solo therapist

  13 EHRs that charge an add-on, with published prices; ClinikEHR is not in the pool because its AI notes are a capped allowance; read September 26, 2026

• 33

  State health-AI bills signed

  Of over 280 introduced in 2026; the 33 are in 22 states, as of July 30, 2026

How We Gathered This Data

We read federal documents and rules, peer-reviewed studies, preprints, professional-body reports, surveys, state laws, court dockets and vendors' own pages. We read them in September and October 2026. A preprint is a study posted before outside experts review it. Every number carries its data year, and we say when a vendor reported its own result.

We left out social posts, pages we could not open and claims we could not tie to a page, because we could not check them. We also left out duplicate sources and rules that a newer rule replaced. We left out developer-platform prices a clinic would not pay. The one exception is Vapi's price, shown only to illustrate how a privacy agreement can cost more than usage.

A few company pages are named in words, not linked. Our ChatGPT guide links the company's Epic and announcement pages, and our Claude guide links the Claude launch announcement. The ChatGPT help center pages we cite on contracts required by HIPAA, the federal health privacy law, are named by topic.

Our EHR cost calculator holds 123 products: 114 full EHRs and 9 tools that are not EHRs. Seven of the 114 are free software projects a practice installs itself, so we leave them out, which leaves 107 full EHRs. We read each price on the vendor's own pricing page on September 26, 2026.

"Published" means the vendor's own page gives a price for at least some practice sizes. "Complete" means nothing on the vendor's pricing page is missing, unpriced or unlisted for the features we picked. Quote-only vendors, which give a price only on request, sit outside every figure, so the true market median is not known. We quote a median, the middle value, only when 5 or more vendors remain without ClinikEHR. We always give n, the number of vendors.

Each scenario assumes one note per session for per-note vendors and one insurance claim per billed session. Claims equal sessions times the insured share. That is 77 a month for the solo therapist (110 times 70%), 385 for the group (550 times 70%) and 720 for primary care (900 times 80%).

"Monthly" is the steady cost after first-year discounts, before card fees; "first year" adds set-up fees. SimplePractice claim rates are those from October 1, 2026. ClinikEHR's quote includes features its owner confirmed but its pricing page does not list.

We build ClinikEHR, and it is one of the 107 EHRs in the data, so where it changes an answer we show results with and without it. This report is mainly about the United States, because that is where the public data we found come from.

What is agentic AI in an EHR, in plain words?

Short answer: an AI agent is software built on a language model that plans steps and uses tools to act in real systems, not only write text. A tool here is a program the agent can call, such as a chart search or a scheduling program. A language model is the kind of AI, like the one behind ChatGPT, that reads and writes text. An agentic AI EHR is an EHR with such an agent inside it or connected to it. Health research has no standard definition.

How is an AI agent different from a chatbot or an AI scribe?

Think of three rungs. A chatbot answers in words and, by itself, changes nothing. An AI scribe listens to a visit and drafts a note that a clinician reads, fixes and signs. An agent plans steps and uses tools to act in a real system. A U.S. government notice, which treats "AI agents" and "agentic AI" as one thing, puts it this way.

AI agent systems are capable of planning and taking autonomous actions that impact real-world systems or environments.

U.S. Department of Commerce, Federal Register notice on AI agent security, January 8, 2026.

The notice adds that such systems can run with little to no human oversight. It is a security document, not a clinical definition. The chart below shows the three rungs.

How it works

Each step up from a chatbot gives the software more power to act

Three rungs: answer, draft, act

1. Answers

   Chatbot

   Answers a question in words. On its own it changes nothing in a chart.

   adds listening
2. Drafts

   AI scribe

   Listens to a visit and drafts a note. A clinician reads, fixes and signs it.

   adds tools
3. Acts

   AI agent

   Plans steps and uses tools to prepare or take an action in a real system.

Research on agents is thin. A 2026 scoping review, which maps existing studies, kept only 7 of 984 records, and the authors say the field lacks clear definitions (npj Digital Medicine, 2026). Its search ran to April 29, 2025, before every 2026 product here. The evidence section gives the other details.

Where does a scribe end and an agent begin?

The line is blurring. Epic's AI Charting drafts the note and also queues up orders from the conversation (Epic, 2026). Ambience Healthcare describes a system that writes the note, updates the problem list, suggests billing codes and holds orders for the clinician's signature (Ambience Healthcare, 2026). Its post gives no release date.

The test: ask what the tool can change without you. The more it can change, the more review it needs.

What have EHR and AI companies actually shipped by October 2026?

Three labels help. "Shipped" means on sale. "Early or limited" means a pilot, an alpha version (an early test), early adopters, or approved organizations only. "Announced" means a dated plan that is not confirmed live.

Short answer: six of the ten products have shipped, four are early or only announced, and every figure in the table comes from the vendor. Here is where AI agents in healthcare stand in 2026. Read the verb in launch news, because "now available" and "planned for 2027" are very different promises.

Which AI scribes and agents can a clinic use today?

The table lists ten products and the status of each on October 6, 2026. Every figure in it is reported by the vendor.

Six products have shipped. Two reach small practices through an EHR a clinic may already use: athenaAmbient inside athenaOne, which serves ambulatory practices, and healow Genie with eClinicalWorks. The Oracle nurse agent is for hospital inpatient units. Of the six, only athena's page in our sources states a price, which is no additional cost.

What is still announced, early or limited?

Four products are not simply on sale. ChatGPT for Healthcare with Epic is for approved organizations only. Epic's Agent Factory is for health systems. The 20 hours a week saved by one ECU Health agent, itself a health system, comes from Epic's own statement. The front-office agents from athenahealth have staged dates, and its August 13 release is about the clinical visit and its scribe, not those agents.

Ambience Chorus has no release date or price. Treat "alpha," "coming soon" and "planned" as a roadmap, and ask for a date in writing.

What does the evidence say about AI scribes and agents?

Short answer: for scribes, the best studies show modest, uneven time savings, a drop in self-reported burnout, and notes that still contain mistakes. For agents that act, the evidence comes from simulations and early studies, and we found no trial of an agent that acts in a live practice. An ambient scribe is an AI scribe that listens to the visit as it happens. A randomized trial assigns people by chance to use a tool or not, which makes the comparison fair. A confidence range is the span where the true effect probably sits.

Do AI scribes save time and ease burnout?

The table shows five studies.

Time first. In the UCLA trial, one scribe cut time in each note by 9.5% and the other by 1.7%, too small to tell from no change. So results depend on the product. The Wisconsin trial found 0.36 fewer hours a day on notes, about 22 minutes (0.36 times 60 is 21.6).

In the largest comparison, 16.0 fewer documentation minutes per 8 scheduled patient hours is 2 minutes per scheduled hour (16.0 divided by 8). As an example, not a study result, a clinician with 24 scheduled patient hours a week would see about 48 fewer minutes at that rate (2 times 24). That study was not randomized, and clinicians at 4 of 5 sites chose to join. PHTI called the evidence on time saved limited.

Now burnout: in the 30-day survey, the share of clinicians reporting burnout fell from 51.9% to 38.8%, a drop of 13.1 percentage points (51.9 minus 38.8). But about 40% of enrollees did not finish both surveys, and there was no control group. In the studies above the gains were minutes a day. In the Atrium Health study cited by PHTI, though, 18% of participants saved over an hour a day in the EHR. Test any product in your own clinic.

How often are AI-drafted notes wrong?

A hallucination is a statement the visit did not contain. An omission is something said and left out. The table shows four sources.

In the UC Davis pilot, 337 of 356 reviewed notes (94.7%) had no errors that the reviewers rated significant. The other 19 notes (5.3%) had errors rated serious or imminent risk. Check it: 337 plus 19 is 356. These rates count notes, not sentences, so do not compare them with sentence-level rates. The authors wrote:

careful clinician review of notes remains imperative

Taylor et al., University of California, Davis, JMIR Medical Informatics, April 2026.

The other sources point the same way. In the JAMIA comparison, the scribe that physicians rated best had more audit-confirmed errors than another, so a note can read well and still be wrong. Physicians also flagged more hallucinations than the audit confirmed, so their ratings did not track the audit either way. In the ECRI survey, 9% of 124 respondents is about 11 people (0.09 times 124) who said an AI error reached a patient or affected a care decision. Read and fix every AI-drafted note before you sign.

Do AI scribes change billing?

At the University of California, San Francisco (UCSF), using visit data from January 1, 2023, through April 1, 2025, 698 attending physicians who adopted a scribe were compared with 867 who did not, 1,565 in all (698 plus 867). Adopters billed 1.81 more relative value units (RVUs) a week. An RVU is the score Medicare uses to set payment. Adopters also saw 0.80 more encounters a week, and the share of claims with a denial did not differ. The authors put 1.81 RVUs a week at $3,044 per physician per year (JAMA Network Open, 2023 to 2025).

That is revenue, not profit, from one health system. The authors say future work should check for upcoding, which is billing at a higher level than the care supports. A scribe that lists at $39 a month costs $468 a year (39 times 12). One list price and one hospital's revenue do not compare like for like.

Before you count on extra billing, track your own billing level for 30 to 60 days before and after a pilot. Then have someone check that any higher level matches the care given.

Do AI agents that act on their own work yet?

Not reliably, and we found no trial of an agent that acts in a live practice. The San Diego study tested draft replies that a physician sends. The table shows four tests.

In this test, finishing a whole task was harder than finishing its pieces. The best Stanford agent was a computer-use agent (CUA), which works a screen the way a person does.

the best-performing agent (Claude Opus 4.6 CUA) achieves only 36.3% task success

Bedi et al., Stanford University and Stanford Health Care, arXiv preprint, April 10, 2026.

A different agent scored 82.8% on subtasks, the pieces of a task, so the gap is 46.5 points (82.8 minus 36.3). On simulated tests, scores rose quickly: MedAgentBench v2's 91.0% on the original tasks is 21.33 points above the 2025 original's 69.67%, though the new paper also changed the agent's design, and it scored 88.67% on 300 new tasks. Its authors call for human review of consequential actions such as medication orders. Treat agents as helpers for single, checkable steps.

How widespread is AI scribe adoption, and what do AI tools cost?

Short answer: four in five surveyed physicians say they use AI at work, but the survey counts any professional use. In Doximity's surveys, 29% named voice-based documentation, which includes AI scribes (of physicians surveyed in the latest wave). We found no source with a rate for small practices. Stand-alone scribes in our examples list from $39 to $180 per clinician a month, and four enterprise scribes and one EHR list no price.

How many clinicians and practices use AI?

The table gives seven surveys. They ask different questions, so we do not add them.

The biggest hole is small practices. The Centers for Disease Control and Prevention (CDC) ran a 2024 survey of office-based physicians. In it, 20.7% of physicians worked solo, 19.2% in groups of 2 to 3 and 28.8% in groups of 4 to 10. That adds to 68.7% (our sum), about 7 in 10. The survey's published national estimates list no AI question, so it gives no rate for small practices (CDC, 2024).

Buying is not adopting. Large health systems told PHTI that typically 20% to 50% of clinicians use a scribe offered to everyone (PHTI, 2025). In Doximity's survey, 71% worried about accuracy, and only 8% called their institution's AI process clear. In the AMA survey, 41% expect AI to harm patient privacy and 13% expect it to help. A short written policy on who may use AI, for what, and who reviews the output is one way to start closing that gap.

What do AI scribes and agents cost on their own?

The table shows list prices read on each vendor's own page on October 6, 2026. Dollar signs are assumed to be US dollars.

Pricing units differ: per tier (Freed), per seat with an agent allowance (Heidi) and add-on tiers (Upheal). Phone agents use three pricing units. Goodcall charges per agent, $79 to $249 a month. Dialzara sells included minutes, from 60 minutes for $29 to 1,000 minutes for $349 a month. Smith.ai charges per call: $300 a month covers 30 calls and $2,100 covers 300 (Goodcall, 2026; Dialzara, 2026; Smith.ai, 2026).

None of the three pricing pages mentions HIPAA, the federal health privacy law, or a business associate agreement (BAA). A BAA is the contract HIPAA requires with a vendor that handles patient information. Ask each vendor before you use it with patient calls.

Vapi, a developer platform, lists hosting at $0.05 a minute, which is $50 for 1,000 minutes (1,000 times 0.05). It sells its HIPAA agreement for $2,000 a month, 40 times that $50 (2,000 divided by 50) (Vapi, 2026). Speech and phone costs come on top, so this is not a clinic's receptionist cost.

What do AI notes cost inside an EHR?

The last section priced stand-alone tools. Our calculator prices a whole practice, so the AI-notes line is the change in the monthly bill, for vendors with published prices only. Of the 107 full EHRs, 44 publish a price, which is 41% (44 divided by 107). Therapy EHRs publish 24 of 33 (73%), and medical EHRs 12 of 32 (38%).

Source: EHR cost calculator; full EHRs only; prices read September 26, 2026.

The three groups add up: 44 plus 57 plus 6 is 107. For 63 of the 107 EHRs (57 plus 6), a clinic cannot see a price without talking to the vendor. So ask for a written quote that lists AI notes.

How many EHRs include AI notes, charge extra or stay silent?

A feature the pricing page does not mention is "not listed," not "not offered." Of the 23 therapy EHRs with a published price, not counting ClinikEHR, 16 include AI notes at no extra cost or sell them for a stated price. Noterro includes only a capped allowance, and so does ClinikEHR if it is counted. The table's group-of-5 row has 23 EHRs, one fewer than the solo row's 24, because WriteUpp has no published price at 3 or more clinicians.

Three include them at no extra cost: Carepatron, Halaxy (video visits only) and Blueprint. Noterro caps it at 300 credits, about 18 minutes of recording. ClinikEHR caps its AI Note Taker at 300 pooled audio minutes a month on Essential and 900 on Team. Thirteen charge an add-on or a higher plan, 4 give no price, and 2 do not list it.

Source: EHR cost calculator; n is full EHRs with a published price; prices read September 26, 2026. ClinikEHR's AI Note Taker is a capped allowance, not an unlimited inclusion.

An add-on is the change in a vendor's monthly bill when AI notes are added. For a vendor that needs a higher plan, it includes the step up. ClinikEHR is not in this pool because its AI notes come with a capped monthly allowance, not an add-on price.

Source: EHR cost calculator; monthly change in the bill; prices read September 26, 2026.

The solo therapist add-on has a median of $35 a month across 13 vendors with published prices in our calculator, and prices run from $15 to $175. The group-of-5 median is $188 across 12, with prices from $75 to $625. The medical pool has 4 vendors, below our 5-vendor rule, so the table names them. In therapy, 16 of 24 EHRs with a published price include AI notes at no extra cost or charge a stated add-on. Two more, Noterro and ClinikEHR, include only a capped allowance, and 6 leave it unpriced or unlisted (16 plus 2 plus 6 is 24).

The chart below shows how far apart the add-on prices are. The line runs from the lowest to the highest price, the shaded box holds the middle half of prices, and the tick marks the median.

The data

The AI-notes add-on runs from $15 to $625 a month, depending on the practice

Change in the monthly bill when AI notes are added, for EHRs that charge extra

• Solo therapist13 products · Prices read September 26, 2026
  $35 middle$15 to $175
• Therapy group of 512 products · Prices read September 26, 2026
  $188 middle$75 to $625

$0$200$400$600

lowest to highestthe middle halfthe median (half cost less, half cost more)

What does a solo therapist pay for AI notes?

The persona is a solo therapist: 1 clinician, 40 clients, about 110 sessions a month and 70% billed to insurance, which is 77 claims (110 times 0.70). The features are telehealth, claims, portal, online booking, text reminders, e-signature and intake forms.

The AI scenario adds AI notes. Eight vendors have nothing missing, unpriced or unlisted in both scenarios, so the medians are close to like for like, except for ClinikEHR's capped AI allowance. With AI notes, the cheapest are Carepatron at $58, TheraNest at $108 and Zanda at $108. The most expensive are Healthie at $254, TherapyAppointment at $150 and SimplePractice at $146.

Source: EHR cost calculator; solo therapist, 77 claims a month; n 7 and 8; prices read September 26, 2026.

Without ClinikEHR the median is $86 a month across 7 vendors with published prices in our calculator; with it, $87 across 8. With AI notes it is $120 without ClinikEHR and $125 with it. Adding AI notes moves the middle monthly price up by $34 without ClinikEHR ($120 minus $86) and by $38 with it ($125 minus $87). Over the first year the middle price rises by $418 ($1,445 minus $1,027) and $457 ($1,513 minus $1,056). These gaps compare two medians, so they differ from the $35 add-on median.

ClinikEHR's quote is $130 a month in both scenarios. For every vendor the calculator adds per-user fees, claims fees and needed add-ons, so that is above the $99.90 Essential price. Counting ClinikEHR moves the median with AI notes from $120 to $125.

ClinikEHR's $130 assumes its AI Note Taker stays within the pooled allowance of 300 audio minutes a month on Essential. If every one of the 110 sessions were recorded, that is 2.7 minutes a session (300 divided by 110). The calculator does not price minutes beyond the allowance, so ClinikEHR's AI figures are not like for like with vendors that price every note.

Card fees are not in these monthly figures. ClinikEHR has the highest online card fee in our calculator data: 5.10% on a $100 payment (a 1.9% platform fee plus Stripe's 2.9% + $0.30), against 2.60% to 4.50% for the 14 other therapy EHRs that charge the practice a published rate.

What does a five-clinician therapy group pay?

The persona is a group of 5 clinicians and 1 admin, with 200 clients and about 550 sessions a month. With 70% billed to insurance, that is 385 claims (550 times 0.70). Features: telehealth, claims, portal, online booking, text reminders, e-signature, intake forms, outcome measures and group sessions. Only 5 of 23 therapy EHRs with a published price are complete here, 4 without ClinikEHR, below our 5-vendor rule, so we name vendors and give no median.

Source: EHR cost calculator; group of 5, 385 claims a month; n 5 with ClinikEHR, 4 without; prices read September 26, 2026.

With AI notes, the five range from $384 (ClinikEHR) to $848 (Healthie). Without them, they range from $318 (Zanda) to $551 (SimplePractice). For add-ons use the pool figures: Zanda $95, TheraPlatform $150, Healthie $469. We do not subtract the monthly figures, because rounding differs.

ClinikEHR's $384 assumes the notes stay within the Team allowance of 900 audio minutes a month, pooled across the clinic. If every one of the 550 sessions were recorded, that is 1.6 minutes a session (900 divided by 550), so the row is not like for like.

What does a three-clinician primary care practice pay?

The persona is 3 prescribing clinicians and 2 admin staff, with 900 patients a month, 80% billed to insurance (720 claims, 900 times 0.80) and 60 lab orders. Features: e-prescribing, telehealth, claims, portal, online booking, text reminders, e-signature and labs. Only 12 of the 32 medical EHRs publish a price, so 20 give no published price (32 minus 12). Of the 12 that do, only 2 are complete here: Healthie and ClinikEHR. The other 10 are missing at least one feature we priced.

Source: EHR cost calculator; primary care, 720 claims a month; n 2 with ClinikEHR, 1 without; prices read September 26, 2026.

Two vendors cannot give a median. ClinikEHR is the higher of the two without AI notes and the lower with them. Check Healthie: $1,166 minus $400 is $766, matching its add-on. Ask for quotes and compare the same features. ClinikEHR's $940 likewise assumes its pooled 900 audio minutes a month, which is 1 minute a visit across 900 visits (900 divided by 900).

How do AI agents connect to an EHR?

Short answer: through a standard door into the record, called a FHIR API, and a standard plug for AI tools, called MCP (Model Context Protocol). An API (application programming interface) is a set of rules that lets one program ask another for data. Permission settings say read or write, and ideally a person approves changes.

What do FHIR, SMART and an API do?

FHIR, short for Fast Healthcare Interoperability Resources, is the common data format and web interface for health records. HL7, the standards group behind it, lists Release 5 (R5), dated March 26, 2023, as its newest published version (HL7, 2023). We did not check for a newer one. But U.S. certification rules, the federal tests a product must pass, still name Release 4.0.1 (R4) as of October 2026 (eCFR, 2026). So R4 is the version that certified EHRs must support.

SMART on FHIR is a set of rules for how apps connect to FHIR systems. Version 2 says what an app may do to each record type. It uses five letters: c create, r read, u update, d delete and s search. A read-only app asks for r and s, as in patient/Observation.rs, while patient/Observation.cu allows creating and updating. A "scope" is only a request, and the EHR and clinic decide what is granted (HL7, 2024).

Our calculator's "API" feature means an API for connecting other systems. That can be FHIR or REST, a common style of web API, so it is not a count of FHIR support. In the table, the two "of which" columns split the first count. The other columns do not overlap (therapy: 6 plus 1 plus 4 plus 2 plus 11 is 24).

Source: EHR cost calculator; n is full EHRs with a published price; prices read September 26, 2026.

Among the 24 therapy EHRs, 6 offer or include an API and 11 do not list one. Among the 12 medical EHRs, 6 offer or include one and 2 do not list one. Not listed is not the same as not offered, so ask any EHR vendor in writing whether it has an API and which AI tools may use it.

What are MCP and A2A?

MCP, short for Model Context Protocol, is an open standard, a common plug that lets an AI app call outside tools and read outside data. Its official site marks the July 28, 2026 revision as current (Model Context Protocol, 2026). In December 2025 the Linux Foundation formed the Agentic AI Foundation as a neutral home for MCP. More than 10,000 MCP servers had been published by then, and the announcement gives no count of clinical ones (Linux Foundation, 2025).

A2A, short for Agent2Agent, is the open standard for how separate agents talk to each other across organizations. Version 1.0 shipped in March 2026 (A2A project, 2026), and over 150 organizations supported it by April 9, 2026, counting supporters, not deployments (Linux Foundation, 2026). MCP is the plug between an agent and a tool, and A2A is the phone line between two agents. You do not need to run either one. Ask any vendor that sells you an agent which standard it uses and what the agent is allowed to change.

What does the ChatGPT Epic integration show about read-only access?

The ChatGPT help center says the Epic plugin is read-only: it cannot update records, place orders, message patients or override chart permissions. It needs an administrator on both the ChatGPT and Epic sides, and a business associate agreement. It is for approved organization workspaces, which are accounts an organization sets up for its staff, not for individual accounts, including the free clinician account. The page says read-only for now, so this can change. Our ChatGPT guide links to the company's pages.

The company's own rating was 99.1% of responses rated safe across 4,363 physician ratings, which leaves 0.9% not rated safe (100 minus 99.1). That is an in-house evaluation, not an independent study. TechCrunch reported that the company called the access read-only (TechCrunch, 2026).

The design rule comes from the MCP standard.

For trust & safety and security, there SHOULD always be a human in the loop with the ability to deny tool invocations.

Model Context Protocol, Tools specification, revision of July 28, 2026.

"SHOULD" is a recommendation, and the project says it cannot enforce it. Authorization, meaning login and permissions, is optional too.

Over the web, MCP uses OAuth 2.1. That is a draft internet standard that lets an app get limited permission without a password. Its scopes can split reading from writing (Model Context Protocol, 2026). So a locked-down connection depends on the builder, not the standard. The chart below shows the order of the pieces.

How it works

How an AI agent reaches your chart, and where a person fits

Four parts between an agent and the record

1. Step 1

   The agent and its connector (MCP)

   A language model with tools asks to read chart data or make a change. MCP is the common plug that carries the request.

   asks to act
2. Step 2

   A person can approve or deny

   MCP says a human should be able to deny a tool call before it runs. That is a recommendation, not a requirement.

   if approved
3. Step 3

   The permission check

   Scopes say what an app may do: read and search only, or also create and update. The EHR decides what is granted.

   if allowed
4. Step 4

   The EHR door (FHIR API)

   The EHR answers through its FHIR API, or makes a change if it is allowed. US certification rules name FHIR Release 4.0.1.

What are the risks of AI in an EHR, and what do the rules say?

Short answer: the main risks of AI in an EHR are wrong or incomplete notes, over-trusting a draft, privacy, higher billing levels and unclear responsibility. HIPAA applies to AI vendors as it does to any vendor that handles patient data, and we found no HIPAA guidance written for AI. Six states had enacted seven laws on AI in clinical care by July 30, 2026, mostly limits on AI in therapy. The clinician stays responsible. This is general information, not legal advice.

Which AI tools will sign a HIPAA business associate agreement?

The Office for Civil Rights (OCR) at the U.S. Department of Health and Human Services (HHS) enforces HIPAA. Its cloud guidance says a vendor that stores a clinic's electronic patient data is a business associate, even if the data is encrypted. Using one without a signed BAA is itself a HIPAA violation (HHS OCR, 2022). That page covers cloud services, so applying it to AI is reasoning by analogy. OCR does not certify products, so no AI tool is "HIPAA certified" by the government.

The contract should limit the vendor's uses, require safeguards and incident reports, bind subcontractors, and cover return or destruction of data. HHS's sample also says to state whether the vendor may de-identify the data, meaning remove details that point to a person (HHS OCR, last reviewed 2017). It does not mention AI training, so whether a vendor may train on your data is our inference, and a lawyer should confirm it. We found no OCR HIPAA guidance written for AI (HHS OCR, 2026).

On October 1, 2026, the Coalition for Health AI (CHAI) released an intake playbook. CHAI released it as Version 1.0 and says it will be refined through a pilot phase. It holds a vendor checklist of 57 yes-or-no checks in 11 topic areas, and the PDF is dated September 2026. The checks cover a BAA, use of customer data for training, and security certifications such as SOC 2, an outside audit of security (CHAI, 2026).

The table shows what each vendor's own pages said on October 6, 2026.

No BAA, no patient information. Ask for the agreement in writing, and ask which plan it covers.

What do federal agencies and states require?

Federal AI rules that reach clinics are few, and one is only a proposal. State laws differ, and the law firm Manatt counted six states with a clinical-care AI law by July 30, 2026. ASTP/ONC, the Assistant Secretary for Technology Policy and Office of the National Coordinator for Health IT, is the part of HHS that runs EHR certification. HTI stands for Health Data, Technology, and Interoperability, which means getting different health systems to share data. The table gives each rule's status on October 6, 2026.

In 2026 through July 30, Manatt, a law firm, counted over 280 health-AI bills introduced and 33 signed, in 22 states. Thirty-three is under 12% of 280, so fewer than 12 in 100 had become law. For AI in clinical care, 27 states introduced 46 bills and six enacted seven (Manatt Health, 2026).

It is unclear whether Colorado's law covers a scribe, so ask counsel. Colorado also passed HB26-1195, which requires written consent before AI records a psychotherapy session, from August 12, 2026. It applies to anyone licensed, registered or certified to provide psychotherapy in Colorado, or otherwise lawfully permitted to. The timeline below shows where five rules stand.

Timeline

Three AI rules are in force, one is only proposed and one starts in 2027

1. March 11, 2024In force

   HTI-1: federal AI transparency rules

   Published January 9, 2024. The duty falls on developers' certified products, not on every AI tool a clinic buys.

2. September 1, 2025In force

   Texas: review AI records, tell patients

   Practitioners who use AI for diagnostic purposes must review all AI-created records and tell patients.

3. December 29, 2025Proposed

   HTI-5 would drop the AI model-card rules

   Comments closed February 27, 2026. No final rule was published by October 6, 2026; it reached White House review on July 28, 2026.

4. August 12, 2026In force

   Colorado: written consent to record therapy

   HB26-1195 applies to anyone licensed, registered or certified to provide psychotherapy in Colorado, or otherwise lawfully permitted to.

5. January 1, 2027Coming

   Colorado SB 26-189: main duties start

   HIPAA covered providers in Colorado must give patients a general notice of advanced technologies. Attorney general rules are still proposed, and a court order of April 27, 2026, entered on the parties' stipulation, bars the attorney general from starting enforcement until 14 days after the court rules on xAI's forthcoming preliminary-injunction motion.

Do you need patient consent to record a visit?

Recording law is separate from AI law. In an all-party-consent state, everyone in the conversation must agree to a recording. California's Penal Code section 632 makes it an offense to record a confidential communication without all parties' consent (FindLaw, 2026). Federal law lets a recording go ahead with one party's consent (Cornell Law School, 2026). States can ask for more, and a clinic should follow the stricter rule.

A journalists' guide, the Reporters Committee's state pages, covers recording law. The guide says about 11 states primarily require every party's consent, including California, with Michigan's rule applying at least to recordings by an outsider to the conversation, and four more require it for either in-person conversations or phone calls. The pages were last updated between October 2019 and October 2025, so ask counsel (Reporters Committee, 2019 to 2025).

Colorado's HB26-1195 requires written, informed consent before AI records or transcribes a psychotherapy session, from August 12, 2026. The consent covers the first use of the AI system and is needed again only if the purpose or manner of use materially changes (Colorado General Assembly, 2026).

Patients' answers depend on the wording. In a survey of 103 patients at NYU Langone, 84 (81.6%) consented when given basic information about the tool. Only 57 (55.3%) consented when told about its AI features, data storage and corporate involvement. That is a gap of 26.3 percentage points (81.6 minus 55.3). The survey included hypothetical scenarios, so these are not consent rates from real visits (JAMA Network Open, 2024).

An insurer, ProAssurance, advises verbal consent at every visit plus a signed form. It warns that raw audio may be demanded by the other side in a lawsuit, a step called discovery (ProAssurance, 2025).

At least two class actions over ambient AI recording are pending: Saucedo v. Sharp HealthCare (San Diego Superior Court, filed November 2025) and Washington v. Sutter Health (filed April 2026). A class action is a claim for a whole group of patients. They say an ambient AI tool recorded their visits without every party's consent. These are claims, not findings.

On October 4, 2026, the court record, called a docket, showed Sutter's requests to dismiss the case. A hearing was set for October 8, 2026, and there was no ruling (CourtListener, 2026). Ask for consent before every recorded visit, and plan for a patient who says no.

Who is liable when an AI-drafted note is wrong?

The signing clinician is first in line. In April 2024 the Federation of State Medical Boards adopted a policy that says:

the physician is ultimately responsible for the use of AI and should be held accountable for any harms that occur

Federation of State Medical Boards, policy on AI in clinical practice, April 2024.

That is policy guidance, not law. A legal analysis adds that a practice can be liable for how it chose and monitored the tool, and vendors for defects. Liability for fully automated notes is unsettled (JCO Oncology Practice, 2025). A Texas insurer advises signing only after review, with no automatic signatures, and an addendum to fix later errors (Texas Medical Liability Trust, 2025).

Manatt found that no state AI liability bill had passed in 2026 as of July 30, 2026 (Manatt Health, 2026), so ordinary malpractice law likely applies. Responsibility for the signed note stays with the clinician.

What This Means for Your Clinic: What Should You Do Now?

A small clinic can start carefully with a scribe under a signed agreement, a consent script and clinician review. It should treat any agent that writes to your chart as a pilot. These eight steps follow from the evidence above.

1. Start with notes, not agents. Why: notes have the best evidence, such as adopters spending 16.0 fewer documentation minutes per 8 scheduled hours (JAMA, 2023 to 2025). The best agent finished only 36.3% of simulated tasks (arXiv, 2026).
2. Ask your EHR vendor whether AI notes are included, an add-on or unpriced. Why: the solo-therapy add-on has a median of $35 a month (n 13, vendors with published prices in our calculator that charge an add-on; ClinikEHR is not in the pool because its AI notes are a capped allowance) (EHR cost calculator, 2026). Four enterprise scribes list no price (Microsoft, 2026).
3. Get a business associate agreement in writing before any patient detail goes in, and ask which plan it covers. Why: using a vendor that handles patient data without one is a HIPAA violation (HHS OCR, 2022), and coverage differs by plan (Claude's privacy center, 2026).
4. Write a consent script and a recording policy, and check your state's law with counsel. Why: in a patient survey, basic information drew 81.6% yes and fuller detail 55.3% (JAMA Network Open, 2024).
5. Read and sign every note yourself, with no automatic signatures. Why: 18% of reviewed notes had an omission (JMIR Medical Informatics, 2024), and medical boards put the responsibility on the physician (FSMB, 2024).
6. For any agent that touches your chart, ask whether it is read-only, which scopes it needs and who approves each change. Why: scopes separate read from write (HL7, 2024), and MCP only recommends human approval (Model Context Protocol, 2026).
7. Run a 30 to 60 day pilot that tracks minutes in notes, after-hours EHR time and errors found. Why: the largest study found no clear change in after-hours time (JAMA, 2023 to 2025). Large health systems told PHTI that typically 20% to 50% use a scribe offered to everyone (PHTI, 2025).
8. Keep contracts short and recheck the rules. Why: HTI-5 is still proposed, with its final rule at White House review (ASTP/ONC, 2025), and Colorado's main duties start January 1, 2027 (Colorado General Assembly, 2026).

What We Still Do Not Know

• Adoption in small practices. We found no AI adoption rate by practice size, and the CDC practice-size survey lists no AI question (CDC, 2024).
• Real-world agent results. The agent tests we found are simulations or early studies (npj Digital Medicine, 2026).
• Billing. Higher billing after scribes may mean more care or upcoding, and we cannot tell which (JAMA Network Open, 2023 to 2025).
• Federal guidance. We found no OCR HIPAA guidance on AI (OCR's January 2025 AI letter covers Section 1557 nondiscrimination, not HIPAA), and the FDA guidance we read does not mention generative AI (FDA, 2026).
• Rules in motion. No final HTI-5 rule is published (it is at White House review), Colorado's attorney general rules are proposed, and neither California class action had a ruling on the merits in the sources we read.
• Liability. No state AI liability bill had passed in 2026 as of July 30, 2026 (Manatt Health, 2026).
• Prices. Quote-only vendors do not publish prices, so our medians describe only vendors that do.

Frequently Asked Questions

What is agentic AI in healthcare, and how is it different from an AI scribe?

An AI agent is software built on a language model that plans steps and uses tools to act in real systems. A scribe only listens and drafts a note. The federal definition treats agentic AI and AI agents as the same thing. The line is blurring: Epic's AI Charting and Ambience's system draft the note and also queue up orders, and Ambience holds them for the clinician's signature.

Can a small clinic use the ChatGPT Epic integration?

Not by itself. The company behind ChatGPT announced it on September 1, 2026 for organization workspaces. It needs a business associate agreement and setup by an administrator on both the ChatGPT and Epic sides. It is not available to individual accounts, including the free clinician account. It is read-only, so it can show chart information but cannot write a note into the chart.

Do AI scribes really save time?

A little, and it depends on the product. In a UCLA randomized trial, one scribe cut time in each note by 9.5% and another showed no clear change. A Wisconsin trial found about 22 minutes less on notes a day. A five-system JAMA study found 16.0 fewer documentation minutes per 8 scheduled patient hours, with no clear change in after-hours time.

How many doctors use AI scribes, and what about small practices?

In early 2026, 81% of surveyed physicians said they use AI professionally, and 29% named voice-based documentation in Doximity's surveys. In 2026, 71% of health systems reported ambient scribing or documentation AI at some level of deployment. We found no small-practice rate, and the federal practice-size survey lists no AI question.

How much does an AI scribe cost?

Freed lists $39 (Starter) to $119 (Premier, billed monthly) a month per clinician; the page does not label the billing basis for the lower tiers. Heidi's Clinician plan is $110 to $150, and Upheal's AI Notes is $39. athenaAmbient costs nothing extra on athenaOne. Dragon Copilot, Abridge, Nabla and Suki list no price. In our calculator, the add-on inside an EHR has a median of $35 a month for a solo therapist (n 13). For a five-clinician group it is $188 (n 12).

Do I need patient consent to use an AI scribe?

Often yes, because of recording law, and a few states now add AI-scribe rules of their own. A journalists' guide says about 11 states primarily require every party's consent to record, including California, and four more require it for either in-person conversations or phone calls. Colorado requires written consent for AI recording of psychotherapy since August 12, 2026, Rhode Island requires written notice and consent for AI in therapy sessions, and Louisiana requires verbal disclosure before AI-transcribed recording from August 1, 2026. Insurers advise verbal consent at each visit and a signed form. Ask counsel about your state.

Do I need a business associate agreement before using an AI tool?

Yes, if the tool stores or processes patient information for you. HHS says using such a vendor without a signed agreement is a HIPAA violation. ChatGPT Business has no BAA, and ChatGPT for Clinicians lets an individual sign one. Claude's agreement covers Enterprise and the API after setup. Heidi, Freed and Upheal say they offer one; Upheal says every covered entity accepts its agreement at sign-up.

Who is liable if an AI-drafted note is wrong?

The signing clinician, first. The national body for state medical boards says the physician is responsible and accountable for harms. A legal analysis adds that practices can be liable for how they chose and monitored the tool, and vendors for defects. In 2026 through July 30, no state AI liability bill had passed. Texas requires review of all AI-created records when AI is used for diagnosis.

Conclusion

The state of agentic AI in EHRs in October 2026 is simple to state. AI scribes are real, and the best studies show modest, uneven gains. Time on notes fell by minutes, two studies found lower reported exhaustion, and the largest study found no clear change in after-hours EHR time. Agents that act are a mixed picture: a few, such as Epic AI Charting and healow Genie, are on sale, but the proof is thin. The best agent in a Stanford simulation finished 36.3% of tasks from start to finish, and the ChatGPT link to Epic is read-only.

Wrong notes and recording consent are the practical risks, and the clinician stays responsible for the signed note. Rules differ by state, and some are still changing. A small clinic can pilot a scribe under a signed business associate agreement. It should use a consent script, review every note, and treat any agent that writes to your chart as a pilot.

ClinikEHR is the EHR we build. Its public pricing page lists a Free plan at $0 for up to 50 clients and 2 staff, with no card needed. Essential is $99.90 a month and Team is $250 a month. Its AI Note Taker is a capped allowance: 300 audio minutes a month pooled across the clinic on Essential and 900 on Team. Compare prices in the EHR cost calculator.

Related Articles

• State of EHR 2026: Adoption, Costs, Rules and What's Next
• State of Small Clinics 2026: Costs, Staff, Rules and Software
• State of Behavioral Health Practice 2026: Demand, Pay and Rules
• Free Transparent EHR Cost Calculator 2027: Compare the Real Price of 50+ EHRs for Your Practice
• ClinikEHR AI clinical documentation features
• What Is MCP (Model Context Protocol) and Why It Matters for Healthcare AI Agents
• AI Agents and HIPAA: What Every Practice Should Know Before Automating
• Can You Use ChatGPT or AI for Therapy Notes? Ethics, HIPAA & the New Epic Integration
• AI vs Manual Clinical Notes: What Actually Works Better in 2026?
• AI Automation for Small Clinics: What to Automate First When You Don't Have a Big Staff
• Can You Connect an EHR's AI Agents to Claude or ChatGPT? Here's How

Statistics are as of the dates shown. Check the linked sources for updates. This is general information, not legal or medical advice.