How to summarize medical records: by hand, by template and with AI
Answer in brief: to summarize medical records, start from the question the summary must answer, number every page, remove duplicates, pull each encounter into a dated grid with its page number, then write the summary for its reader and check every line your opinion rests on against the source page.
You can do that by hand, with a template, or with AI software that drafts the grid and summary with page citations for you to verify. Which one fits depends on file size, how often you do it, and whether the summary goes into a report you sign.
Limits: We have not timed the listed tools on a shared test file: time figures come from the cited studies and vendor pages, and the worked example is synthetic.
How do you summarize medical records?
Write down the question first: the referral questions for an IME, the treatment under review for a peer review, the reason for a referral or transfer. Then work in this order: paginate, deduplicate, sort by date, extract each encounter with its page number, write the summary in the template for that purpose, flag gaps and conflicts, and verify. Reading is the slow part. At the paces vendors publish, 40 to 100 pages an hour, a 500-page file is 5 to 12.5 hours of reading before you write a word.
3 methods compared: by hand, by template, with AI
Most guides on this topic are written for paralegals building a demand letter. This one is for the physician or nurse reviewer who reads the file and signs what comes out of it. The 3 methods differ in where your hours go.
| Method | What you do | Where the time goes | Main failure | Fits |
|---|---|---|---|---|
| By hand | Read every page, take notes, write a narrative | Reading and note-taking, roughly in proportion to page count | Fatigue: pages skimmed late in a long file, no page numbers to check later | Files under about 100 pages; cases where you need to read every word anyway |
| Template | Read every page, fill fixed fields (date, provider, finding, page), write the summary from the grid | Same reading, faster writing, easier checking | Fields you didn't think to include; normal findings left out that later matter | Repeat work of 1 type (IME, peer review) where the reader expects the same layout |
| AI-assisted | Upload, review a drafted grid and summary with page citations, verify, edit | Verification: opening the cited pages behind each line you rely on | Omissions most often, then wrong details and rare invented ones (see the error studies) | Files of hundreds or thousands of pages; frequent work; when you need a page cite on every line |
How long it takes: time-per-page benchmarks
No regulator or professional body publishes a standard reading pace for medical records. What exists is reading research, 1 timed abstraction study and vendor estimates. Here they are side by side, each with its source, read 2026-11-12.
| Figure | What it measures | Kind of source | Source, date |
|---|---|---|---|
| 238 words a minute | Average adult silent reading rate, English non-fiction (190 studies, 18,573 readers) | Peer-reviewed meta-analysis | Brysbaert, J Mem Lang 2019, 2026-11-12 |
| 60 seconds per report; 21.5 minutes per patient | Median time for trained abstractors to pull fields from imaging reports by hand | Peer-reviewed timed study | BMC Med Inform Decis Mak 2022, 2026-11-12 |
| 5 min 22 s per encounter | Chart review share (33%) of physician EHR time per outpatient encounter | Peer-reviewed study, as reported | Overhage and McCallie, Ann Intern Med 2020, via Healio, 2026-11-12 |
| About 100 pages an hour | Experienced clinical reviewer, standard records; complex records 2 to 3 times slower | Review service's estimate | medicalrecordsreview.com, 2026-11-12 |
| 40 to 60 pages an hour | Manual clinical record review | Software vendor's estimate | medchronosai.com, 2026-11-12 |
| 1 to 3 hours per 100 pages | Reading every word once, no extraction | Software vendor's arithmetic | botminds.ai, 2026-11-12 |
| 4 to 8 hours per chronology | Building 1 chronology by hand | Software vendor's estimate | assemblysoftware.com, 2026-11-12 |
What that means for a file. At 40 to 100 pages an hour, 40 pages take 24 to 60 minutes, 500 pages take 5 to 12.5 hours, and 2,000 pages take 20 to 50 hours. That is reading only. Extraction and writing come on top, and the 2022 study shows extraction alone runs about a minute per report even for a handful of fields.
The 3 vendor figures come from companies that sell software or services against manual review, so treat them as an upper bound on speed, not a target. Time your own pace on 1 closed file with the calculator.
The 8 steps, in short
Every method runs the same 8 steps. What changes is who or what does steps 4 to 6. The full method, with what to record at each step, is on the step-by-step page.
- Write the question. List the referral questions, the treatment under review, or the reason for referral. The question decides what goes in.
- Number every page. Apply Bates numbers or a page index before you read, so every fact can carry a page.
- Remove duplicates. The same ED note or MRI report often arrives 2 or 3 times from different senders. Mark copies; read 1.
- Sort and index. Order by date of service and list each document: date, provider, type, pages.
- Extract. For each encounter: date, provider, complaint, findings (pertinent negatives included), diagnosis, plan, work status, page.
- Write for the reader. Use the template for the purpose. Lead with what answers the question.
- Flag what is missing or in conflict. Referenced tests with no report, gaps in care, laterality or date conflicts, prior history of the same body part.
- Verify and sign. Check every line an opinion rests on against its page. Note which records you read in full.
Pick the template by purpose
A summary for an IME and a summary for a transfer share facts but not structure. Pick the template by who reads it and what they must decide. Copyable versions of all 4 are on the templates page.
| Purpose | Reader | The question it answers | Must include |
|---|---|---|---|
| IME or QME record review | Carrier, attorney, judge | Each referral question: diagnosis, causation, treatment, MMI, work capacity | Records list with page ranges, prior history of the same body part, every work status change, missing records |
| Peer review (records only) | Payer, plan, review organization | Was the treatment under review reasonable and necessary? | The treatment request, the documented indication, prior treatment tried and its result, the guideline cited |
| Referral to a specialist | The consultant | What do you want the consultant to answer? | The question, relevant history, results already done (so they aren't repeated), current meds and allergies |
| Transfer of care | Receiving team | What does the next team need to keep the patient safe today? | Active problems, hospital course, current meds with last doses, pending results, goals of care |
What goes wrong: error types in hand and AI summaries
Hand summaries fail in predictable ways: pages skimmed when you are tired, a scanned page that looked blank, notes without page numbers that can't be checked later. Nobody has published rates for these in medical-legal summaries, so we don't quote any.
AI summaries have been measured. 2 published studies, 1 of hospital course summaries and 1 of clinical notes, agree on the order of risk:
- Omissions come first. Physicians who gave feedback on 100 AI hospital course summaries reported omissions in 25, inaccuracies in 20 and invented content in 2. They rated 88 of the 100 as having no potential for harm (JAMA Network Open 2026).
- Invented content is rarer but often serious. In 12,999 LLM-written clinical note sentences, 1.47% contained a hallucination, and 44% of those were rated major: they could change diagnosis or management if left uncorrected. 3.45% of source sentences were omitted (npj Digital Medicine 2025).
- Negations are a common way to be wrong. In the same npj study, 30% of hallucinated sentences were negations: the note said the opposite of the source. "Denies radicular pain" for "reports radicular pain" is the medical-legal version.
So an AI-assisted summary needs a different check from a hand summary. You verify what is there line by line against its page, and you hunt separately for what is missing. The quality checks page turns that into 12 checks with a score.
Which method to use, by situation
| Situation | Use | Why |
|---|---|---|
| 1 file under about 100 pages, once a month | By hand with the template | Under 1 to 2.5 hours of reading at the published paces; software setup and a BAA may cost more time than they save |
| IME or QME files of 500 to 5,000 pages, several a month | AI-assisted, verified against the template | 20 to 50 hours of reading for a 2,000-page file by hand; page cites make your verification and any deposition faster |
| Peer review against a guideline | Template; AI to find the indication and prior treatment | The answer turns on a few documents, which you must read in full |
| Transfer or referral today | Template from the EHR, by the treating team | Speed and current meds matter more than history; this is clinical documentation, not record review |
| You need someone else to read it | A clinician-written service | Days, not hours, and $0.25 to $0.50 a page at published rates (see tools); you still verify what you rely on |
Rules that change how you summarize
- California medical-legal reports: only the signing physician may review and summarize prior records, except that if someone else excerpts them, the physician must review the excerpts and make whatever further inquiry is needed. A report that fails the rule is inadmissible and unpaid (Labor Code 4628). An AI draft is an excerpt you review, not a section you paste.
- California workers' comp reports should list all information received or relied on (8 CCR 10682(b)(4)). Keep the records index from step 4; it becomes that list.
- Hospital transfers: the hospital must send "all necessary medical information" about the current course of illness and treatment, post-discharge goals of care and treatment preferences (42 CFR 482.43(b)).
- AI and HIPAA: uploading records to any software without a signed Business Associate Agreement is a HIPAA problem. Consumer chat plans don't include one. OpenAI and Anthropic offer BAAs for specific business products only, with conditions (see tools).
- Billing pages: California's fee schedule pays $3.00 a page for record review beyond the 200 pages included in a full initial evaluation (8 CCR 9795, ML-PRR). The manual vs AI page works out what that pays per hour of reading.
Worked example: 1 synthetic file, 3 summaries
The worked example takes a synthetic 40-page workers' comp low back file (no real patient) and summarizes it 3 ways. It shows the page index, the hand narrative, the template grid and an AI-style draft with 3 planted errors: a missing prior history, a laterality flip and an EMG that was ordered but never appears in the file. Each method catches a different subset.
Questions about summarizing medical records
How long does it take to summarize medical records?
Reading alone runs about 40 to 100 pages an hour at the paces review vendors publish, so a 500-page file is 5 to 12.5 hours before writing. Complex records can take 2 to 3 times longer per page. AI software returns a cited draft in minutes to about an hour, and your time moves to verification.
What should a medical record summary include?
What the reader needs to answer their question, each with a date and a page number: the records reviewed, relevant prior history, each encounter's findings and diagnosis, treatment and response, work or functional status, and what is missing or in conflict. The template differs by purpose: IME, peer review, referral or transfer.
Can I use ChatGPT to summarize medical records?
Not with a consumer account: there is no Business Associate Agreement, so uploading a patient's records is a HIPAA problem. Even with a BAA-covered business product, a chat answer usually has no page citations and can drop parts of a long file without saying so. Purpose-built tools cite each line to a page.
Is an AI summary of medical records reliable?
Partly. In published studies, omissions were the most common error (25 of 100 AI hospital course summaries) and invented content was rare (2 of 100), but some errors were rated major. Treat an AI summary as a cited draft: verify every line you rely on and look separately for what is missing.
What is the difference between a summary and a chronology?
A chronology lists every encounter in date order with its page. A summary selects and explains what matters for 1 question, usually in fewer words. Most physician reviewers build the chronology first, then write the summary from it.
Do I have to read every page if I use AI?
For a signed medical-legal report, you are responsible for what it says. California requires the physician to review any excerpts made by someone else. Read in full the records your opinions rest on, verify cited lines, and note in the report which records you reviewed.