AI in EHR: how AI improves electronic health records and how to add it

AI in an EHR takes on the work around the record: it drafts the visit note from the recorded conversation, summarizes a long chart before a visit, turns faxes and lab files into fields, sorts the inbox, and suggests billing codes. At Ronas IT, an AI feature that works with one EHR starts at $3,000, with project management included, and takes at least 1 week of build time once the EHR vendor has given API and test-environment access, which runs on the vendor's schedule. A clinician still reviews and signs the drafted note, and staff check what the model is unsure about.
There are three ways to add it: switch on the AI features your EHR vendor already offers, add an app that opens inside the EHR, or build your own AI service that works through the EHR's API. This guide is for clinic groups and health-tech companies that are choosing between them. For AI elsewhere in healthcare, such as medical imaging and remote patient monitoring, see our overview of AI in healthcare.
What AI in EHR means
AI in EHR is software that reads the patient record and does part of the work around it, inside the EHR or through the EHR's API. The table shows where it helps and what a person still checks in each area.
| Area | What AI does | What a person checks |
|---|---|---|
| Documentation | Drafts the visit note from the recorded conversation and summarizes the chart before a visit | The clinician edits the note and signs it |
| Data entry | Reads faxes, referral letters and lab files and fills the matching fields | Staff confirm the fields the model was unsure about |
| Inbox | Sorts patient messages and results by urgency and drafts replies | Staff send the reply or rewrite it |
| Coding and billing | Suggests billing codes from the note and drafts prior authorization requests | A coder approves the codes before the claim goes out |
| Clinical decision support | Scores the risk of events such as sepsis or readmission and sends alerts | The care team decides whether to act on the alert |
Why clinics add AI to the EHR
The clearest reason is time. In a 2017 study in the Annals of Family Medicine, 142 family medicine physicians spent 5.9 hours of an 11.4-hour workday in the EHR, counted per full-time physician, and 1.4 of those hours came after clinic hours. The researchers took the time from the EHR's event logs and checked the logs against direct observation.

The Permanente Medical Group started using ambient AI scribes, which listen to the visit and draft the note, in late 2023. According to the AMA's report on its follow-up analysis in NEJM Catalyst, the scribes were used in more than 2.5 million patient encounters in one year and saved physicians an estimated 15,791 hours of documentation time.
How AI improves the management of electronic health records
AI helps where staff retype, copy or search for information that is already in the record or in a document next to it. The six uses below differ in risk. The first five move and summarize information. The last one gives clinical advice, so it needs the most proof before anyone relies on it.
Visit notes from the conversation
An ambient scribe records the visit and drafts the note in the format the clinic uses. The clinician reads and corrects the draft before signing it. The note has to go into the right section of the record, so the scribe needs a working connection to your EHR as well as a good speech model.
Chart summaries before a visit
A patient with a long history can have years of notes, results and letters. AI can pull the active problems, current medicines, recent results and open questions into one short summary. Each line should link to the place in the chart it came from, so the clinician can check it in one click.
Documents turned into fields
Faxes, scanned referral letters and lab reports still arrive as files. AI can read a file, match it to the right patient and fill the fields in the record. Anything it is unsure about, such as a second patient with the same name, goes to a person instead of into the chart.
Our closest public case shows similar data work, without AI. On a web platform for analyzing lab test results, results came in from the laboratory's database without manual upload, the platform structured them into health indicators, and admins combined the indicators into reports. A user's doctor or health coach could open the results and leave comments on the reports.
Inbox sorting and draft replies
Patient messages, refill requests and incoming results share one inbox. AI can sort them by urgency, route them to the right person and draft a reply. Staff send the reply themselves, so the clinic stays responsible for every word the patient reads.
Coding and prior authorization
From a signed note, AI can suggest billing codes and draft a prior authorization request with the supporting facts from the chart. The codes go to a coder for approval, and staff send the request.
Risk prediction and alerts
Prediction models estimate how likely a patient is to develop sepsis, be readmitted or get worse, and send alerts. In a 2021 study in JAMA Internal Medicine, researchers at Michigan Medicine checked the Epic Sepsis Model, a proprietary model then in use at hundreds of US hospitals, on 38,455 hospitalizations. It did not identify 67% of the patients who had sepsis, while at the study's alert threshold it would have sent alerts for 18% of all hospitalized patients. Before a prediction model sends alerts in your clinic, test it on your own patients and count both the cases it misses and the alerts nobody needed.
Three ways to integrate AI into an EHR
The three ways are the AI features of your EHR vendor, an in-EHR app, and your own AI service on the EHR's API. Your own AI service gives you the most control, and it fits a health-tech company whose product is the AI feature.

1. AI features from your EHR vendor
Large EHR vendors now build AI into their own products. Epic says it has moved to embed AI throughout its software, with separate AI tools for clinicians, for patients, and for billing and other revenue cycle work. Oracle Health AI Agents draft documentation and suggest medical codes.
“If your EHR vendor already sells the feature and it fits how your doctors work, buy it. We would rather say so on the first call than build a copy you then pay to maintain.”
Roman Surikov, CEO at Ronas IT
2. An app that opens inside the EHR
SMART App Launch is an HL7 standard built on FHIR, a format for exchanging health data through an API. It lets an app open from a user's session inside the EHR and receive its context, such as the patient who is selected on screen. The app works with the permissions the user gives it, and the clinician does not switch to another system. It suits a tool that one company sells to clinics on different EHRs.
3. Your own AI service on the EHR's API
A backend service reads records through the EHR's FHIR API, runs the model and writes the result back, with no user starting it. The Backend Services part of SMART covers this case, such as a nightly job that summarizes new lab results. In the US, health IT certified under the ONC Health IT Certification Program to criterion § 170.315(g)(10) has to offer a standard API for one patient's data and for groups of patients, based on HL7 FHIR Release 4.0.1. What you may write back, and how you get access, depends on the vendor and your contract with it.
How the three options compare
| Option | Fits when | What you control | Main limit |
|---|---|---|---|
| EHR vendor's AI features | Your EHR already sells the feature you need | The settings the vendor offers | Works only in that EHR and follows the vendor's plans |
| In-EHR app | You sell one tool to clinics that use different EHRs | Your app's logic and screens | Data and write access differ from one EHR to the next |
| Your own AI service | The AI feature is your product, or the job spans several systems | The model, the prompts, the data flow and where it runs | You own the testing, the hosting and the agreements with every vendor |
Another route is to replace the EHR with a new system built around AI. We advise against it when the goal is a few AI features: it means moving every record and training every user again.
How to build an AI feature on top of an EHR
Here is how we would run this kind of project. It can open with an optional analysis phase of one week or more, which produces a written scope, budget and timeline. For the AI part itself, such as model choice and prompts, our generative AI integration guide goes deeper.
- Pick one workflow and one number. The number can be minutes per visit note, the share of faxes that staff type in by hand, or EHR time after clinic hours.
- Map the data. List the FHIR resources the feature reads, what it writes back, and whether the EHR vendor gives you a test environment.
- Sign the agreements. Every vendor that will see patient data signs a business associate agreement before real data flows. That usually includes the model provider and the hosting provider. The development partner needs one only if it receives patient data: it can build against the EHR vendor's test environment and synthetic records instead. In that case, your staff run the real-record test set from the next step on each build we deliver to your environment.
- Build a test set from real records. Only your staff and the vendors under these agreements handle the records. Write down the output you expect for each case, and rerun the set on every build that changes the model or the prompt.
- Build in sprints. A sprint usually runs for one to three weeks, and at its end you get a result to look at, such as a new version on the test server. On our side, our engineers write and test the code with AI agents against the vendor's test environment and synthetic records, while the real-record set stays with your staff. An engineer reviews and accepts each change before it reaches that server.
- Pilot with one team, then add more teams. Compare the number from the first step before and after the pilot, and add every wrong output to the test set.
“Demos run on clean charts. Real charts have notes copied from the last visit, old medicine lists and scanned pages turned sideways, and that is where a model fails first. Put those records in the test set before anyone writes the first prompt. Otherwise you find these failures only in the pilot.”
Alexander Storozhevsky, Lead developer at Ronas IT
Risks and rules to plan for
The main risks are wrong output that reaches the chart, patient data shared without the right agreement, and clinical advice software that counts as a regulated medical device.
Risks in the model's output
- Wrong or invented text. A model can write things nobody said. Keep the clinician's review before signing, link summary lines to their source, and check each release on the test set.
- Alert fatigue. A model that alerts too often makes the alerts that matter harder to notice. The authors of the sepsis study above called this a large burden of alert fatigue. Set thresholds on your own data and track how many alerts lead to action.
- Patients the model has not seen. A model tested on one patient group can work worse on another. Test it on the patients your clinics actually serve.
- A model the provider retires. Providers switch off older model versions on dates they set. Anthropic gives at least 60 days' notice before it retires a publicly released model, and OpenAI at least six months for its generally available models. A new model can word notes and summaries differently, so rerun the real-record test set before it reaches the chart, and budget for that move in support.
US rules to check
- Business associate agreements. Under 45 CFR 160.103, a vendor that, on behalf of a covered entity, creates, receives, maintains or transmits protected health information for a function or activity regulated by the HIPAA rules, such as data analysis or billing, is a business associate, and so is its subcontractor that does the same. The vendor's promise to safeguard the data has to be in a written agreement. AWS, for example, lists Amazon Bedrock among its HIPAA eligible services and tells covered entities and business associates not to use any of them with protected health information without first signing an AWS business associate agreement.
- Recording consent. An ambient scribe records the visit. Ask your lawyer what consent your state requires before you record.
- Information about the vendor's predictive models. Since 2025, EHRs certified to the decision support criterion, § 170.315(b)(11), have had to give identified users access to source attributes of predictive models, such as how they were validated and tested for fairness. The HTI-5 final rule removes those requirements from December 28, 2026, so if your clinics rely on a vendor's risk scores, ask for that information in the contract.
- Medical device rules. Software that suggests a diagnosis or a treatment can be a medical device. The FDA's guidance on clinical decision support software explains which decision support functions fall outside the device definition. Our guide to medical device software covers the device side.
Whether a feature falls under HIPAA, FDA or state rules is a question for your lawyer. Our overview of US healthcare laws and regulations lists the main ones, and the flow maps and architecture review from the analysis phase give your lawyer the facts to work from.
What AI in EHR costs
For one AI feature on one EHR, our price starts at $3,000 for the AI integration, and the connection to the EHR's API is part of that work. An optional analysis phase before the build costs $1,000 to $2,000. Here are our starting prices for each part, and for options you may need instead or in addition.
| What you need | When it applies | Starting price | Timeline |
|---|---|---|---|
| Analysis phase: flow maps of the workflow, a feasibility check, an architecture review and a written scope | Optional, before the build, when the scope is not written yet or you want to check feasibility first; without it, the analysis happens during the build | $1,000 to $2,000 | At least 1 week |
| AI integration: the AI feature in your product or workflow, with the connection to the EHR or another system it reads from or writes to | For each AI feature | From $3,000 | At least 1 week |
| Machine learning: a custom model trained on your data | Only for predictions, such as a risk score | From $15,000 | At least 6 weeks |
| Healthcare app: a new product built around the record | When you build the whole product, not one feature; the price covers one platform | From $30,000 | At least 10 weeks |
| Support for the whole app the feature runs in: monitoring, fixes and updates | Optional, after the feature goes live; from $1,000 for one app, and from $2,000 per month for a whole system with several client apps | From $1,000 per month | Ongoing |
Project management is part of every build price above, and the design of any new screens the feature needs is included in the AI integration price. Model usage and hosting are paid to the providers, and production runs in your own cloud account. Our prices do not include legal advice; that comes from your lawyer.
What to do next
- Look at your EHR vendor's AI features. If one fits, start there.
- Pick one workflow. Write down the number you want to change.
- Ask your EHR vendor about the API. Find out how you get access, whether there is a test environment, and what you may write back.
- List every vendor that will see patient data. Include your development partner only if it will receive real records, and confirm each one will sign a business associate agreement.
- Get a rough estimate. It is free and works from these answers. If the scope is not written yet, you can add the optional analysis phase to turn them into a scope with a budget and a timeline, or leave the analysis to the build.
First published 2023-04-17 by Dmitry Lauretsky, Former COO at Ronas IT. Rewritten by Evgeny Leonov.
Frequently Asked Questions (FAQs)
How is AI being used in modern EHR software?
For the work around the record. AI writes a first draft of the visit note from a recording of the visit, summarizes a long chart before the patient arrives, pulls data from faxes and lab files, triages the inbox, and proposes billing codes. Some systems also score the risk of events such as sepsis and send alerts. A clinician signs the drafted note, staff check what the model is unsure about, and a coder approves the codes before the claim goes out.
How does AI improve efficiency in EHR systems?
It takes over retyping, copying and searching. The Permanente Medical Group, a large US medical group, gave its physicians ambient AI scribes in late 2023. In a follow-up analysis in NEJM Catalyst, reported by the AMA in 2025, the scribes saved an estimated 15,791 hours of documentation time in one year.
How does an AI scribe connect to an EHR?
The note has to go into the right place in the record, so a scribe works through the EHR vendor's own feature, as an app launched from within the EHR, or over the EHR's API. Before you buy one, ask which EHRs it supports and what it writes back to each.
How do you integrate AI into an existing EHR?
There are three ways. You can switch on the AI features your EHR vendor sells, add an in-EHR app through the SMART App Launch standard, or build your own service that reads and writes through the EHR's FHIR API. The third gives you the most control and the most work: testing, hosting, and agreements with every vendor that sees patient data.
How do you implement AI notes in EHR workflows?
Start with one team and one number, such as minutes per visit note. Make sure the scribe can write the note to the right section of the record, build a test set from real records, and keep the clinician's review before signing. Ask your lawyer what consent your state requires before you record visits. Compare the number before and after the pilot, then add more teams.
Is AI in an EHR HIPAA compliant?
HIPAA compliance belongs to the whole setup, not to the AI. Under HIPAA, a vendor that handles protected health information for a covered entity, such as a clinic, is a business associate, and so is its subcontractor, which can include the model provider. Each one needs a written business associate agreement before patient data reaches it. Your lawyer confirms which vendors and data flows are allowed.
What is the difference between AI in an EHR and an AI-native EHR?
AI in an EHR adds features to the system a clinic already runs. An AI-native EHR is a new system designed around AI from the start, so adopting one means moving every record and training every user again. If the goal is a few AI features, we recommend adding them to the current EHR.
Does AI in an EHR need FDA clearance?
It depends on what the feature does. A feature that proposes a diagnosis or a treatment may count as a medical device. The FDA's January 2026 guidance on this kind of software lists the decision support functions that are not devices. A regulatory lawyer or consultant makes the call for your feature.
How much does it cost to add AI to an EHR?
With us, one AI feature connected to one EHR starts at $3,000, and that price covers the connection to the EHR's API, the design of any new screens, and project management. An optional analysis phase ($1,000 to $2,000) can come before the build. A new healthcare product built around the patient record starts at $30,000 for one platform, such as a web app for clinicians.






