AI medical record tasks: what works and what fails, 1 by 1
Answer in brief: AI handles 4 reading tasks well enough to save real time (summaries, chronologies, missing-record flags and questions answered with page links), 2 production tasks under close audit (coding and abstraction, and writing the visit note), and 1 patient task you should know about even if you never buy it. Each section below says what works, what fails, what it costs and what you still check.
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.
1. Summarize a record file
What it is: the AI reads the chart or an uploaded file and returns a short account: active problems, key events, results, medications, and what changed recently. There are 2 different products under this name. EHR-native summaries read your own chart before a visit. Medical-legal summaries read an outside file of PDFs, scans and faxes for a review or a claim.
- Tool categories
- EHR features (Epic AI Insights), in-house hospital tools (Penn's Chart Hero, Stanford's ChatEHR), clinician apps (Abstractive Health), and medical-legal tools (MediScan, InPractice, DigitalOwl, Wisedocs, Medrecords AI). See the directory.
- What works
- Summaries with a link from each line back to the source. Epic says AI Insights produce a chart summary "with links back to the original documentation", and EpicShare reports that clinicians at Riverside Health spent 23% less time reading notes after rollout (EpicShare, 2026-10-19). Penn built Chart Hero so every statement ties to chart data with 1-click access to the source.
- What fails
- Negatives that flip ("denies chest pain" becomes "chest pain"), findings that get dropped because they sit in a scanned attachment the tool didn't read, and old problems presented as current. In the Asgari study, 30% of hallucinated sentences were negation errors.
- What it costs
- Bundled in the EHR license, $99 per user a month for Abstractive Health's clinician app, or $0.05 to $0.15 a page for medical-legal tools.
- What you still check
- Every negative your opinion depends on, the current medication list against the latest note, and whether scanned outside records were read at all.
2. Build a chronology
What it is: 1 row per encounter in date order, with the provider, the facility, what happened and the page it came from. It is the backbone of an IME, a peer review, an expert opinion or a hospital claim file. Physicians read chronologies more than they build them, so the question is whether you can trust each row.
- Tool categories
- Medical-legal record tools: InPractice, Dodon, CaseMark, Superinsight, MediScan, Wisedocs, Medrecords AI and others. Most are sold to law firms and insurers as well as physicians.
- What works
- Files that are deduplicated before the chronology is built. Outside productions often contain the same visit 2 or 3 times, from the provider, the insurer and the attorney. A tool that removes exact duplicates first gives you fewer rows to check and doesn't bill you twice.
- What fails
- Dates. A fax header date, a print date and the date of service can sit on the same page, and the wrong one moves an event out of order. Undated handwritten notes are the next weak spot.
- What it costs
- Per page or capped subscription. Published rates run from CaseMark's $25 plus $0.01 a page (capped at $100) to InPractice's $0.15 a page for a 1,000-page pack.
- What you still check
- The rows your opinion rests on: open the cited page and confirm the date of service and the finding. Spot-check 5 random rows as well.
3. Find missing records and gaps in care
What it is: the AI compares what the file says should exist with what it contains. A bill with a date of service but no visit note, a referral to a specialist with no specialist note, an MRI ordered with no report: each is a sign that a record is missing. A separate check flags long breaks between visits.
- Tool categories
- Features inside chronology tools. Filevine's MedChron, for example, flags dates where a bill exists but the treatment record doesn't, and treatment gaps of more than 30 days (Filevine help center). MOS ReviewGenX and Medrecords AI offer missing-record identification too.
- What works
- Files with bills, orders and referral letters, because those are the clues the AI matches against. It saves the slow part of a review: noticing an absence.
- What fails
- 2 things. A visit with no bill, order or mention anywhere is invisible to any tool. And a gap in the file is not the same as a gap in care: the patient may have been seen, and the record just wasn't produced.
- What it costs
- Usually included with the chronology tool at no extra charge.
- What you still check
- Whether each flag is a missing record (request it) or a real break in care (relevant to causation or compliance). Don't state a gap in care in a report until the record request comes back.
4. Ask questions of a file
What it is: a chat box over the chart or the uploaded file. "Has she had a colonoscopy?" "When was the metformin stopped?" "Any prior lumbar imaging?" The answer comes back in plain English, ideally with the page or note it came from.
- Tool categories
- In-house hospital tools (ChatEHR at Stanford, piloted with 33 clinicians in June 2025; Chart Hero at Penn, 30 clinicians testing as of January 2026), general AI workspaces that sign a BAA (BastionGPT, Hathr AI), and Q&A features inside medical-legal tools (MediScan's MediChat, Medrecords AI's record Q&A).
- What works
- Narrow factual questions with a citation you can click. Penn's team tied every statement to chart data for that reason, and Stanford listed source citations as a feature in development when ChatEHR launched (Stanford Medicine, 2025-06-05).
- What fails
- Questions that need the absence of something ("Was she ever on anticoagulants?"), where a confident "no" can mean the tool didn't find it. Also general AI tools that answer from medical knowledge instead of from the file, with no citation to tell them apart.
- What it costs
- $20 to $75 per user a month for BAA-backed general AI (BastionGPT), from $47 a month (Hathr AI), or included in per-page medical-legal tools.
- What you still check
- Open the citation every time. Ask "no" questions 2 ways ("any anticoagulant", then "warfarin, apixaban or rivaroxaban") and compare.
5. Coding and abstraction
What it is: turning notes into structured data. Autonomous coding tools read the note and assign ICD-10 and CPT codes, sending unclear charts to human coders. Abstraction tools fill registry fields, such as the American College of Cardiology's NCDR registries, from the chart.
- Tool categories
- Autonomous coding (Fathom, CodaMetrix, CombineHealth), registry abstraction (Carta Healthcare's Atlas, Health Elements' Carbon) and general document extraction (Affinda).
- What works
- High-volume, templated documentation. Emergency department coding is where several vendors publish their case studies. The ACC says the Carta Healthcare tool for NCDR aims to reduce abstraction time and cost by 20% (ACC).
- What fails
- Thin or ambiguous documentation. The tool codes what the note supports, so a vague note gets a vague code or a human review. Vendors publish their own automation and quality figures; we haven't tested them.
- What it costs
- Enterprise contracts. None of the coding or abstraction vendors we checked publishes a price.
- What you still check
- Claims go out under your NPI, so ask for the audit sample, the rule for sending charts to a human, and who answers a payer denial. For physicians, the best defense is the note itself.
6. Patient-facing record apps
What it is: patients connecting their own records to an AI assistant to have results and notes explained. This one isn't a purchase for most physicians, but it changes visits. A Pew Research Center survey from August 2026 found that 34% of US adults have used AI assistants for health purposes, as reported by Fierce Healthcare.
- Tool categories
- ChatGPT Health (US users 18 and older, records connected through b.well, OpenAI help center), Claude with athenahealth's athenaPatient MCP connector (US Pro and Max users, launched 2026-11-02), Apple Health Records, and record services such as PicnicHealth.
- What works
- Explaining lab terms and helping patients write down questions. OpenAI states Health is "not intended for diagnosis or treatment".
- What fails
- Patients may read an AI summary that leaves out context only you have, or that reads an old problem as active.
- What it costs
- Free to a consumer subscription for the patient.
- What you still check
- Ask patients what they read and where. If they bring an AI summary, check it against the chart the same way you would check any outside summary.
7. Write the note (ambient scribes)
What it is: the AI listens to the visit and drafts the note. Strictly, this creates a record rather than reading one, but it is the most widely bought medical record AI, and it decides what every later summary has to work with.
- Tool categories
- Standalone scribes (Freed, Abridge), EHR scribes (Epic's Art, per Healthcare IT News) and scribe modes inside BAA-backed AI workspaces (BastionGPT).
- What works
- History and exam sections in routine visits.
- What fails
- The plan. In the Asgari study, the plan section held the largest share of major hallucinations, 21% of them.
- What it costs
- Per clinician per month, from $39 a month for Freed's Starter plan (40 notes a month).
- What you still check
- Every sentence before you sign, the plan section first. A wrong plan in today's note becomes a wrong fact in next year's AI summary.
How the 7 tasks connect
The tasks feed each other. A scribe writes the note, coding turns it into a claim, the note later becomes part of a record file, and a reviewer's AI summarizes that file years later. An error made at the start travels through every later step, and each step adds its own. That is why the check at each step matters more than the tool you pick: read the risks page for the error types, and use the scoring sheet before you commit to any tool.