AI for medical records: what it can and can't do, task by task
Answer in brief: AI for medical records works best on reading tasks where every line it writes points back to a page you can open: summaries, chronologies, answering questions about a file, and flagging records that look missing. It is weakest where nobody checks the output against the source, and it never replaces your judgment, your signature or a BAA.
Pick by task, not by brand. The 7 tasks below use different tools, carry different risks and are priced in different ways. 3 of the 7 tasks, and most of the tool categories, are ones Medrecords AI, which publishes this site, doesn't sell.
Limits: Vendor facts come from each vendor's own website, read 2026-11-12, and we haven't tested every tool; prices and features change, so check the vendor before you buy.
What can AI do with medical records today?
As of November 2026, AI can read a record file of hundreds or thousands of pages and hand back a dated summary, a chronology, answers to your questions with page references, a list of visits that seem to be missing, draft codes, and plain-language explanations for patients. It does these by reading text that already exists. It does not examine the patient, decide causation, or know about care that was never written down. The useful question is not "does AI work" but "which task, checked how, by whom".
The 7 tasks compared (as of 2026-11-12)
Each task has its own users, failure modes and pricing. The table gives the short version; the task pages go deeper, and the tool directory lists 32 products with a source and a capture date for each.
| Task | Who uses it | Works well when | Fails when | Usual price model |
|---|---|---|---|---|
| Summarize | Clinicians before a visit; physician reviewers | Each line links to its source | Negatives flip, key findings are dropped | Per page, per seat, or bundled in the EHR |
| Chronology | IME, peer and expert reviewers; risk teams | Files are deduplicated first | Dates are guessed from print or fax dates | Per page or capped subscription |
| Missing records | Reviewers; risk and quality teams | Bills, referrals and orders are in the file | The file has no clue a visit happened | Usually bundled with chronology tools |
| Ask questions | Clinicians; reviewers | Answers cite the page or note | The answer has no citation or mixes patients | Per seat, per page, or EHR feature |
| Coding and abstraction | Coders, CDI, quality registries | Notes are complete and templated | Documentation is thin or ambiguous | Enterprise contracts, rarely published |
| Patient apps | Patients | Used to prepare questions | Treated as a diagnosis | Free or consumer subscription |
| Write the note | Treating clinicians | You read every line before signing | Plan and assessment sections go unread | Per clinician per month |
What the top search results leave out
On 2026-11-12 we pulled the Google top 10 for "ai for medical records" and 3 close searches ("ai medical records", "ai medical record review", "ai medical records summary") and read every result. For the exact phrase, Google returned 8 results: 2 hospital news stories about tools you can't buy (Penn Medicine's Chart Hero and Stanford's ChatEHR), 5 vendor home pages, and 1 academic scoping review.
- Every result covers 1 task at most. A scribe vendor sells scribing, a chronology vendor sells chronologies. None maps the tasks against each other, so a reader can't tell which category they need.
- None cites a published error rate. The Penn story says hallucination is a risk and human review stays critical, which is true and not enough to act on.
- None gives a test you can run. No result tells a physician how to try a tool on a file they already know.
- Prices appear only on vendor pages, and only for that vendor. No result compares per page, per seat and capped subscription pricing on the same file.
- Patient-facing record AI is missing. ChatGPT Health opened to US adults in July 2026, and athenahealth connected patient records to Claude on 2026-11-02. Your patients may arrive having read an AI summary of your notes.
This site fills those gaps: a task map, published error rates turned into checks, a scoring sheet with a pass mark, a cost calculator for the 3 pricing models, and a dated tool directory.
Where AI goes wrong with records, in numbers
The clearest published numbers come from a 2025 study in npj Digital Medicine by Asgari and colleagues. Clinicians checked 12,999 sentences in 450 AI-written clinical notes against the source consultation transcripts. Their results:
- 191 sentences (1.47%) contained a hallucination, something not supported by the source. 84 of those 191 were rated major, meaning they could change diagnosis or management if left uncorrected.
- 1,712 of 49,590 source sentences (3.45%) were left out of the note. 286 of those omissions were rated major.
- Of the hallucinations, 43% were fabrications, 30% were negations (the note said the opposite of the source), 17% were context errors and 10% were wrong causal links.
- Major hallucinations showed up most in the plan section.
The study measured notes drafted from visit transcripts, not summaries of long record files, but it is the clearest published breakdown of error types. 2 things follow. First, a low error rate on a long file still means errors: 1.47% of a 500-sentence summary is about 7 sentences. Second, the authors note that human-written notes also carry errors and omissions, so the fair comparison is AI plus your check against your current process, not AI against perfection. A 2024 Nature Medicine study rated by 10 physicians found that summaries from the best-adapted models were equivalent to or better than expert summaries in most cases. Both findings hold at once. The risks page turns them into a check routine.
Which task matches your job
The same phrase, "AI for medical records", covers very different jobs. Start from yours.
| You are | Your record problem | Task to look at first | Where to start |
|---|---|---|---|
| Primary care or specialist seeing new patients | Getting up to speed on a long chart before the visit | Summarize, ask questions | Ask whether your EHR already has it (Epic AI Insights, for example) before buying anything |
| Hospitalist or ED physician | Transfer packets of hundreds of pages | Summarize, ask questions | EHR-native tools for your own chart; outside packets need a tool that reads uploaded PDFs |
| IME, QME, peer or expert reviewer | 1,000 to 10,000 pages of outside records, a deadline and your signature | Chronology, missing records, ask questions | Medical-legal record tools priced per page; test on a closed file first |
| Hospital risk or quality team | Claim files, incident reviews, registry abstraction | Chronology, abstraction | Separate tools: record review for claims, abstraction platforms for registries |
| Clinic manager or practice owner | Documentation time, coding backlog, record requests | Write the note, coding, retrieval | Scribes are priced per clinician; coding tools are enterprise contracts |
| Physician whose patients use AI | Patients arriving with AI explanations of their results | Patient apps | Know what ChatGPT Health and Claude connectors can read, and ask patients what they used |
A rule of thumb: if the record lives inside your own EHR, look at what your EHR vendor ships first, because it already has the access and the agreement in place. If the record arrives from outside as a stack of PDFs, scans and faxes, you need a tool built for uploaded files, OCR and deduplication.
HIPAA, the BAA and what you can paste where
HHS guidance on cloud computing is plain: a service that creates, receives, maintains or transmits electronic PHI for a covered entity is a business associate, and it needs a BAA, even if it only stores encrypted data it can't read. For a physician that means 3 practical rules:
- Don't paste identifiable patient records into a consumer chatbot account that has no BAA with you or your organization.
- Ask every vendor for the BAA before the trial, not after. Some tools in our directory, such as BastionGPT, state a BAA on every plan, including the trial.
- Ask whether your data trains their models, and get the answer in the contract. A settings toggle can change; a contract term can't change without you.
Patient-facing tools are a different case. When a patient connects their own records to ChatGPT Health or Claude, the patient is choosing where their data goes. That is their right, and it is not the same as a clinician uploading a chart.
How to test a tool before you trust it
Vendor demos use files the vendor picked. Your test should use a file you picked: a closed case you know well, ideally 300 to 1,500 pages with handwriting, duplicates and at least 1 record you know is missing. Before uploading, write down 10 facts you would stake your signature on, including 2 negatives ("no loss of consciousness") and 1 date that is easy to get wrong. Then score the output against your list. Our scoring sheet does the arithmetic and gives a pass mark, and it includes a synthetic test page (no real patient) with planted traps if you can't use a real file in a trial.
3 results should end a trial early: a negative turned into a positive, a citation that points to a page that doesn't say what the summary claims, and a tool that won't sign a BAA before you upload.
What it costs, in 3 pricing models
Record AI is priced in 3 main ways, and the cheapest one depends on your volume:
- Per page: published rates run from $0.15 down to $0.05 a page for medical-legal tools (InPractice $0.15 to $0.07 prepaid; Medrecords AI 10 cents to 5 cents per deduplicated page; CaseMark $25 plus $0.01 a page, capped at $100).
- Per seat: a flat monthly fee per user, such as $99 per user for Abstractive Health's clinician app or $20 to $75 per user for BastionGPT.
- Capped subscription: a monthly fee for a page allowance plus overage, such as MediScan's $169 for 1,200 pages a month with 10 cents a page after that.
EHR-native features, coding platforms and abstraction platforms rarely publish prices. The costs page has a calculator that runs your file size and monthly volume through all 3 models.
Questions physicians ask about AI for medical records
Is AI for medical records accurate enough to rely on?
It is good enough to save reading time and not good enough to sign without checking. In a 2025 study, 1.47% of AI note sentences were hallucinated and 3.45% of source sentences were omitted. Use tools that cite a page for every line, and check the lines your opinion depends on.
Can I use ChatGPT or Claude to summarize a patient's records?
Not with identifiable records on a consumer account that has no BAA with you. HHS treats a cloud service that handles PHI for you as a business associate that needs a BAA. Use a tool that signs one, or your organization's approved, contracted AI service.
Which task should a busy physician start with?
The one that eats the most reading hours. For clinicians it is usually pre-visit chart summaries, often already in the EHR. For IME, peer and expert reviewers it is the chronology of an outside file, where per-page tools apply.
Does AI find records that are missing from a file?
Some tools flag likely gaps by matching bills, referrals and orders against the records present, or by flagging long breaks between visits. They can only flag what the file hints at; a visit with no bill, order or mention anywhere will not show up.
Who is responsible if an AI summary is wrong?
The physician who relies on it and signs. Tools like Medrecords AI state plainly that they don't diagnose or sign reports; a qualified reviewer checks the output and signs.
How much does AI record review cost?
Published medical-legal rates run from $0.15 to $0.05 a page, so a 1,000-page file costs about $50 to $150. Seat-priced tools run $20 to $99 per user a month. Many EHR, coding and abstraction tools don't publish a price.