Medical record review: the complete guide

Medical record review is the process of reading a patient's or claimant's medical file closely enough to state, with a page citation for each statement, what treatment happened, in what order, and what the record does and does not support. It is one activity with several names. A paralegal calls it building the chronology. An adjuster calls it file review. A physician preparing for an evaluation calls it record review and bills it separately from the examination. A payer calls it clinical review. The reading is the same reading.

What separates it from simply reading is the obligation to produce something another professional can check. A review that cannot be traced back to the page it came from is an opinion about a file, not a review of one. That single requirement, a citation on every line, drives almost every decision in the rest of this manual: how the file has to be ordered before anyone reads it, why duplicates have to come out first, what a chronology column has to contain, and what a deposition will attack.

The 5 deliverables

Almost everything anyone asks for from a medical file is one of 5 things, or a combination. Knowing which one you owe before you start is the difference between 1 pass through the file and 3.

The gap analysis is the one people skip and the one that costs the most when it is skipped, because a missing record is invisible by definition. Everything else in the file announces itself. A provider who was referred to and never appears does not.

The same 900 pages get read by 6 or 7 different people across the life of a claim, each looking for something different and each producing something the next one relies on. The reason a chronology built for a plaintiff attorney rarely serves a defense evaluator is not quality. It is that they are answering different questions off the same paper.

2 consequences follow. First, a review is only reusable downstream if it is neutral: a chronology that only carries the facts helping one side has to be redone by the other, and a chronology that carries every fact serves both. Second, the person doing the reading is frequently not the person carrying the professional risk. The paralegal reads, the attorney signs. That gap is why citation discipline matters more here than in almost any other kind of document work, and chapter 15 is about what happens when it fails.

Every guide to record review assumes a clean PDF. Real productions are not clean, and the difference between the 2 is most of the work. Here is what turns up, in rough order of how often it ruins a schedule.

• Volume nobody sized in advance. A medical management company representative told RAND that a packet can run 10,000 to 15,000 pages. That is an outlier, but the working case is not small either: across 198,774 California cases with any paid record review, the median payment implies roughly 513 pages and the 95th percentile roughly 3,533.
• Duplicates, in bulk. The same discharge summary appears in the hospital production, in the referral packet the orthopedist received, and again in the primary care chart. Near-duplicates are worse than exact ones: same document, different scan, 1 page skewed, so a hash comparison misses it.
• Co-mingled patients. A page belonging to somebody else, sitting inside your production because it was adjacent in the source chart. It is a privacy incident and a credibility problem in the same page.
• Scans of faxes of scans. Optical character recognition on a third-generation fax is not the same problem as OCR on a born-digital PDF, and vendors who quote accuracy rarely say which one they measured.
• Handwriting. Therapy notes, nursing flow sheets, and the margin annotation that turns out to matter.
• Undated and unnumbered pages. A document with no date cannot be placed in a chronology, and a page with no Bates number cannot be cited.
• Template boilerplate. Modern electronic records carry review-of-systems blocks and copied-forward assessments that look like findings and are not. Reading them as new findings on each visit invents a treatment course that did not happen.
• Imaging that is not the report. Studies arrive as discs or as a link to a portal, and the file contains only the radiologist's narrative. See reading imaging as DICOM for what is in one and not the other.

Records that were removed before they reached you

There is one more thing in a production, and it is the thing you cannot see. In California, the 2021 medical-legal fee schedule started charging $3.00 a page past 200, which gives the party assembling the packet a direct financial reason to send fewer pages. RAND found that evaluators report this happening. One active evaluator described defense attorneys removing hundreds of pages that other physicians in the file had referenced, which is how the removal became visible at all.

The evaluator's objection was not about the money. It was that deciding which records are clinically relevant is a medical judgment, and a party to the dispute was making it. Whatever your role, the practical consequence is the same: treat the production as a sample, not a census, until the gap analysis says otherwise.

The workflow below is not a novel idea; it is the order that stops you reading the same page 3 times. The only part people get wrong is the sequence. Reading before ordering feels faster on page 1 and costs hours by page 400, because every contradiction found early has to be revisited once the rest of the file is in date order.

1. Collect. Gather the complete file from every provider named anywhere in the record, plus every prior production. Gaps found in week one are a phone call. Gaps found at deposition are a different kind of problem. Chapter 5.
2. Organize. Deduplicate, separate co-mingled patients, number the pages, sort by provider and date. Nothing is read yet. Chapter 6.
3. Read and extract. Pull every medically significant event, each one carrying the page it came from. Chapter 7.
4. Build the chronology. Put the extracted events in date order with the columns your reader actually needs. Chapter 8.
5. Analyze. Causation, treatment gaps, pre-existing conditions, contradictions between providers, work status over time. Chapter 9.
6. Report. Write the deliverable in the form the end reader needs, with citations they can check in seconds. Chapter 10.

2 rules govern the whole sequence, and both trade a small cost now for a large one later.

A file is complete when every provider named in it has produced records, every date range referenced has documents in it, and the billing matches the treatment. That is a test you run against the records themselves, which is why it can only be finished after a first pass, not before one.

The right of access, and the clock it starts

Under 45 CFR 164.524, a covered entity must act on an individual's request for access to their records no later than 30 days after receipt, with one extension of no more than 30 days. That is the outer bound on a patient-directed request, not a service level, and providers routinely use it. Plan the schedule against 60 days, not against a promise on a release form.

The same rule limits what a provider may charge an individual for the copy. Under 45 CFR 164.524(c)(4) the fee is confined to 4 categories:

• Labor for copying the information, whether in paper or electronic form.
• Supplies for creating the paper copy, or the portable media if the individual asked for electronic copies on media.
• Postage, where the individual asked for it to be mailed.
• Preparing an explanation or summary of the information, if the individual agreed to that.

Search and retrieval time is not on that list. A retrieval invoice that bills for locating the chart is billing for something the rule does not permit on an individual's own access request, and is worth reading closely.

What to request, beyond the obvious

That fourth item is the one that changes outcomes. The reliable way to find providers is not to ask the client to remember them; it is to read the file you already have and extract every name in it. That is what provider list extraction and missing-records detection are for, and it is the same job whether a person or a tool does it.

This is the step people treat as clerical and it decides the accuracy of everything after it. 4 things have to happen before a single clinical fact is extracted, and they have to happen in this order.

1. Deduplicate. Exact duplicates are easy. Near-duplicates, the same document rescanned at a different angle or printed from a different system, are the ones that matter, because each copy carries a different page number and a reader counts them as separate encounters. See deduplication.
2. Separate co-mingled patients. Pull out any page that belongs to another patient. Do not delete it, quarantine it, log it, and tell whoever produced it. See co-mingled record detection.
3. Number the pages. Nothing can be cited until it has a stable identifier. If the production came unnumbered, Bates stamp it before reading, not after, because a citation written against a temporary number has to be rewritten.
4. Sort. By provider, then by date inside each provider. Undated documents go to a holding set, not into the sequence; see undated document flagging.

Why duplicates are a substantive problem, not a tidiness one

A duplicate that survives into the chronology does 3 separate kinds of damage, and only the first one is cosmetic.

That last one is worth stating plainly. RAND's California data is the clearest evidence that this is settled practice rather than a preference: among all the groups they interviewed, including evaluators, attorneys, payers and administrators, no participant of any kind objected to removing duplicate records before they reach the evaluator. Duplicate removal is the one thing everyone in an adversarial system agrees on.

Co-mingling is a privacy event

Another patient's page inside your production is a disclosure of their information to you. Handle it as an incident, not as a filing error: quarantine the page, record what it was and where it came from, notify the producing party, and do not forward it onward. The same page, unnoticed, also becomes a credibility problem, because the first person to spot it at deposition will be opposing counsel.

Extraction is the first pass where anyone reads for meaning. The working definition of a medically significant event is narrow on purpose: something a clinician did, found, prescribed, or decided, on a specific date, that a later reader might rely on. If it fails that test it is context, not an event, and context belongs in the narrative rather than the chronology.

The citation rule

Every extracted line carries the page it came from, written at the moment of extraction. This is the single discipline that separates a review that survives review from one that does not, and it is the one most often relaxed under deadline. Restoring a citation afterwards costs several times what recording it costs, because it means finding the page again in a file you have stopped holding in your head.

2 conventions make citations useful rather than decorative. Cite the page, not the document, because a 40-page operative record with one relevant line is not a citation. And cite the Bates number where one exists, because it is the identifier every other party in the matter shares. See page-level citations for how this is produced automatically.

The chronology is the most common deliverable in this work, and most of what makes one good is decided by its columns. Too few and the reader has to go back to the file. Too many and no 2 rows are comparable.

Date order, with 2 exceptions

2 things break the sequence, and both need a stated convention rather than a judgment call made row by row.

For a worked artifact, see the sample summary. For the deliverable itself, chronology generation.

Analysis is where the chronology stops being a table and starts being an argument. 4 questions produce most of what anyone actually uses. None of them can be answered until the file is in order, which is why this is step 5 and not step 2.

Causation, as the record supports it

The record does not decide causation; a physician does, and in a contested matter a trier of fact does after that. What the record supplies is the evidence: the interval between the incident and the first complaint, whether the mechanism described to the first treater matches the mechanism in the pleading, whether the objective findings are consistent with that mechanism, and whether any treating provider has already committed to an opinion in writing. Collect those 4 and the causation section writes itself, in the voice of the file rather than the voice of the advocate.

Gaps in treatment

A gap is a period with no treatment where treatment would be expected. Every gap has an innocent explanation and a hostile one, and the file usually contains the material to distinguish them: a documented insurance lapse, a course of therapy completed, a period of incarceration, a move, an authorization denial. A gap analysis that lists the gaps without looking for the explanations hands the other side a free argument.

Pre-existing conditions

The question is almost never whether a prior condition existed. It is whether the current presentation differs from the documented baseline. That needs the prior records, which is why chapter 5 insists on requesting the prior period for the same body part even when nobody has asked for it. A pre-existing analysis without baseline records is an assertion.

Contradiction between providers

The same patient tells the emergency physician one mechanism and the orthopedist another. 1 provider records full range of motion and the next, 2 days later, records severe limitation. These are the highest-value findings in a file and they only appear when the chronology is complete, because a contradiction is by definition a comparison between 2 rows. Record both rows with both citations and let the reader decide what it means. See alteration detection for the narrower case where a document itself appears to have changed.

The deliverable is written for one reader, and that reader is usually carrying professional risk on it. Everything about the format follows from that.

4 rules that hold for all of them

Separate what the record says from what you conclude. A write-up that blends the 2 is unusable to a defense reader and dangerous to a plaintiff one, because the reader cannot tell which parts they are on the hook for. 3 more rules follow from the same idea.

Record review is bought in pages, hours, cases, seats and credits, and only one of those compares across vendors. A case is a different size at every firm. An hour is a different amount of work at every desk. A seat and a credit are whatever the seller defines them as. The page is the only unit with a fixed meaning, which is why it is the one to ask for.

So we went and looked at whether anyone publishes one. We read the published claims of 328 record review vendors, software and services together, and scored each on the same 8 facts a buyer needs, taken from that vendor's own site. The result is the pricing study, recomputed on every build of this site so it cannot drift from the underlying profiles.

A row marked "not published" in that study is a negative finding: we looked on the vendor's own site and it is not there. It is not a judgement about the vendor or about its pricing.

Disclosure is worst where the stakes are highest

The pattern is not random. The 2 categories closest to the medical-legal work itself publish nothing at all between them, covering 57 vendors, and the most transparent category in the study is the one furthest from it.

The published range, and why it is not a ranking

Across the 8 vendors that publish a per-page figure, the range runs from $0.008 to $0.35 a page, roughly fortyfold. That spread says less about margin than about deliverable. A page index, a chronology, a narrative summary and a physician-reviewed report are 4 different products quoted in the same unit. Read the deliverable before you read the rate.

4 questions that turn a quote into a comparison

There is exactly one jurisdiction in the United States that says, in law, what reading a page of a medical-legal file is worth. California's medical-legal fee schedule at 8 CCR 9795, effective 1 April 2021, sets it at $3.00 a page past the first 200 pages, and every number in this chapter comes from that schedule or from RAND's analysis of what the state actually paid under it.

2 structural details matter more than the headline numbers. The schedule's modifiers, which raise the fee for a psychiatric evaluation or an agreed medical evaluator, multiply the evaluation only, never the per-page record fee. And the first 200 pages are included, so the meter starts at page 201 and runs flat from there.

What that produces in practice

RAND, in RR-A4655-1 (May 2026), commissioned by the California Department of Industrial Relations, analysed 1,752,557 panel assignments from 2012 to 2024 across 5,034 evaluators. Within that, 198,774 cases carried at least one paid record-review line. The distribution across those cases:

Read the top of that distribution carefully. At the 99th percentile the record review alone pays $23,000, more than 11 times the $2,015 for the evaluation it supports. RAND found that the growth in California medical-legal spending since 2021 is driven principally by the per-page fee rather than by the evaluations, and that in an ML201 practice roughly 48% of revenue is record review. Nearly half of what a comprehensive-evaluation practice earns is paid for reading, not for examining.

Why a Californian number matters outside California

Almost nowhere else publishes one. Every other market prices this work privately, which is what chapter 11 measured. So $3.00 a page is not a national benchmark and should not be quoted as one; it is the only public data point in existence for what 1 page of medical-legal reading is worth when a regulator sets the number. Treat it as a reference point and as an illustration of what happens when the page becomes a billable unit: the volume goes up, the cost follows, and the parties start managing the page count instead of the case.

There are 3, and most organizations end up with 2 of them running at once. The choice is usually framed as cost, and it is really about where the judgment sits and who carries the risk.

The honest comparison

None of the 3 removes the human. In-house and outsourced put a person on every page. Software puts a person on the verification pass instead, which is a smaller job on a clean file and not a smaller job on a contested one. What changes is where the hours go, not whether there are hours.

The combination most firms land on is software for the organize-and-extract work, where the task is exhaustive and mechanical, and a human for the analysis and the signature, where it is neither. That split maps exactly onto the 6 steps: chapters 5 through 8 are where the machine earns its keep, and chapters 9 and 10 are where it does not.

The trade-offs are argued at more length in in-house versus outsourced and AI versus human review, and the money side in what record review costs.

6 failure modes account for most of the damage. 4 of them are about the file and 2 are about the reader, human or otherwise.

Why the records arrive late, and why it is structural

RAND asked California evaluators what most needed fixing about the medical-legal process. The answer was the timing of records. Under 8 CCR 35 the parties exchange proposed records 20 days before serving them on the evaluator, with 10 days to object, so the evaluator is not entitled to the records until 10 days after the evaluation. Under 8 CCR 38 the report is still due 30 days from that evaluation. The rules put the reading after the exam and then run the clock on both.

What that looks like in practice: records arriving a median of 9 days after the evaluation, and, in one evaluator's account to RAND, 2,000 pages landing 10 days after the exam with the report already 90% written. The report then either goes out incomplete or becomes a supplemental at $650.

What the 2 error rates actually say

The best evidence on this comes from clinical research rather than from litigation, because clinical research measures it. Garza and colleagues, writing in BMC Medical Research Methodology in 2022, studied medical record abstraction in the ACT NOW CE Study and reported an error rate of between 1% and 3% depending on whether you count all fields or only populated ones, under formal abstractor training and continuous quality control. They also found that each additional quality-control event reduced the error rate by 0.51 percentage points.

2 things follow. Trained humans working under a QC regime make errors at a low but non-zero rate, which is the honest baseline any AI claim should be measured against. And the paper's own framing is blunt: abstraction is a significant source of error in studies that rely on record data. On a 900-page file, 1% is 9 pages read wrong.

Research published in PLOS Digital Health in 2025 found that 42% of GPT-4-generated summaries of emergency department encounters contained at least one hallucinated detail when checked against the source record, absent human review.

That number is why every serious deployment of this technology is built around verification rather than around trust, and why citations are not a feature but the mechanism. A cited line can be checked in seconds. An uncited line has to be re-derived from the file, which means the review was not done.

A review becomes a problem in exactly 2 places: when the underlying records have to get into evidence, and when the person who signed the summary has to defend it under oath. The first is a rules question. The second is a methodology question.

Getting the records in

Medical records are ordinarily admitted under the business records exception, Federal Rule of Evidence 803(6), "Records of a Regularly Conducted Activity". The rule's conditions are that the record was made at or near the time by, or from information transmitted by, someone with knowledge; that it was kept in the course of a regularly conducted activity; that making it was a regular practice of that activity; that all this is shown by a custodian or other qualified witness or by a certification under Rule 902(11); and that the opponent does not show a lack of trustworthiness in the source, method or circumstances.

Rule 902(11) is the route that avoids putting a custodian on the stand. It makes a domestic business record self-authenticating where the proponent produces a certification of the custodian or another qualified person meeting 803(6)(A) through (C), and where, before the trial or hearing, the proponent gives the adverse party reasonable written notice of the intent to offer the record and makes the record and certification available for inspection, so the party has a fair opportunity to challenge them.

The practical consequence for record review is small but sharp: the certification page is part of the production. It is the page most likely to be dropped as administrative in chapter 6's organize step, and it is the page that decides whether anything else gets in. Keep it, index it, and note in the gap analysis where it is missing.

Defending the method

The questions come in a predictable order, and a review built the way this manual describes answers all 6 from its own artifacts.

The last 2 are the ones that do damage when they are improvised. Both are cheap to answer in advance and impossible to answer afterwards. See audit trail and chain of custody.

Medical records are the most sensitive category of document most firms handle, and record review means sending them somewhere. Whether that somewhere is an outsourced service or a software vendor, the same questions apply.

1. A signed business associate agreement, before any file moves. Not a security page, not a badge on a website, an executed BAA. If a vendor cannot produce one, the conversation is over.
2. Encryption in transit and at rest, stated plainly. What algorithm, where the keys live, who can decrypt.
3. Named subprocessors. Where the file physically goes, including any model provider, and in which jurisdiction. "Cloud" is not an answer.
4. A minimum-necessary posture. Only the people who need the file can reach it, and there is a record of who did.
5. An explicit training and reuse policy. Whether your records, or anything derived from them, are used to train, fine-tune or evaluate a model. This is the question most often left unanswered, and it is the one with the longest tail.
6. Deletion on request, with a stated retention period and a way to verify it.

Redaction is a separate decision

Redaction is not a privacy control on the vendor side, it is a decision about what the eventual reader gets. Removing identifiers before a file is reviewed can break the review, because a co-mingled page is caught by comparing patient identifiers. Redact on the way out, for the deliverable, rather than on the way in. See redaction.

Map the 6 steps against what each kind of worker is actually good at and the answer stops being controversial.

The line

Software drafts and cites. You review, you revise, you sign. That sentence is the whole boundary, and everything on the wrong side of it is a claim nobody should make: a signed medical opinion, a merit score on a case, an automated claims decision, a coverage denial without a human on it. None of those are record review. They are decisions that a record review informs.

The reason is not caution for its own sake. It is the 42% from chapter 14 on one side and the professional licence on the other. A draft that cites its pages costs seconds to check. A finished-looking output with no citations costs a full re-read, which means the software saved nothing and added a risk.

Demos are built on clean files. Your files are not clean. The only test worth running uses a closed matter you already know the answer to, which means you can grade the output instead of admiring it.

Pick the right file

Not the biggest one and not the tidiest. Pick a closed case of 400 to 900 pages that you personally worked, that contained at least one duplicate set, at least one handwritten document, and at least one thing you found late and wish you had found early. That last item is the actual test.

Run the same 9 checks on every vendor

• Every line cites a pagePick 10 rows at random and resolve all 10. 9 out of 10 is a fail.
• Duplicates were removed, and you were toldAsk for the count and the page numbers, then verify a few.
• The handwritten pages were readFind yours in the output. If they were skipped, was that disclosed?
• The thing you found late is in thereThe single highest-value check in the list.
• Nothing is asserted that the file does not sayRead the analysis section against the record, looking for invented specifics.
• The missing-records list is realDoes it name providers referenced in the file with nothing produced?
• The output is in a format you can work inOnly 30 of 328 vendors publish this. Sortable and pasteable, or a re-typing job.
• The price is computable before you buyA per-page rate you can multiply by your own page count, with the duplicate policy stated.

Score every vendor on the same file. The comparison is worthless otherwise, because file difficulty dominates every other variable in this work.

4 artifacts, copy them into your own process. Every one of them answers a question from chapter 15 before anybody asks it.

RECORD INTAKE LOG - one row per packet

Matter / claim number
Date packet received
Delivery route (portal, courier, email, disc, link)
Producing party
Pages received (count before anything is removed)
Pages after de-duplication
Duplicate pages removed (count)
Pages quarantined as another patient's (count)
Date range of records, earliest to latest
Providers and facilities represented
Bates range applied            Applied by whom
Certification / affidavit page present?   Y / N
Undated documents (count)
Handwritten documents (count)
Providers referenced in this packet with no records produced
Date missing-records request sent
Reviewed by                    Date completed

INCLUSION RULE - state before review begins

Scope of the review
  Bates range reviewed:
  Date range of care covered:
  Body parts / conditions in scope:

What was extracted as an event
  Every encounter, objective finding, diagnosis, treatment,
  medication change, work-status change, referral, and
  provider statement on causation, on a stated date.

What was not extracted, and why
  Exact and near-duplicate copies of a document already recorded
  Copied-forward assessments identical to a prior visit
  Template review-of-systems and past-history boilerplate
  Administrative pages (fax cover sheets, releases, confirmations)
  Pages quarantined as belonging to another patient

Undated documents
  Held in a separate appended set with the bounding evidence stated.
  No date was inferred to place a document in the sequence.

Prepared by:                      Date:
Software used, if any:            Verified by:

Re: [Matter] - records referenced in the production but not produced

Reviewing the records served on [date], Bates [range], the following
are referenced within the file itself and do not appear in it.

1. Providers named in the records with no records produced
   [Provider, specialty, and the page that names them]

2. Studies and reports referenced with no report produced
   [Study, date, and the page that references it]

3. Date ranges with no documents
   [Range, and the treatment on either side of it]

4. Billing without corresponding treatment records, or the reverse
   [Charge, date, and the page]

5. Certification or affidavit pages not included
   [Producing custodian]

Please produce these, or confirm in writing that they do not exist
or will not be produced. This list is taken from the file as served
and is not a request for anything outside it.

MEDICAL CHRONOLOGY

Matter:                          Claimant / patient:
Prepared for:                    Prepared by:
Date prepared:                   Version:

Records reviewed:  Bates [range], [N] pages as served,
                   [N] pages after de-duplication
Date range of care covered:
Providers represented:           [count]

Records referenced but not produced:  see Missing Records, p. [x]
Documents that could not be dated:    [count], appended
Pages quarantined (other patient):    [count], logged separately

Method: every row cites the page it came from. Statements in the
Findings column are what the record says. Anything that is an
inference by the preparer is marked [inference].

Software used in preparation:    [name, or none]
Output verified against source pages by:            Date:

The 9-point verification checklist

The same 9 checks from chapter 18, in a form you can tick.

Checked: 0 of 9.

• Every line cites a page10 rows at random, all 10 resolve
• Duplicates removed, counted and disclosedAsk for the page numbers
• Handwritten pages were read, or the skip was disclosedFind yours in the output
• The thing you found late is in thereThe highest-value single check
• Nothing is asserted that the file does not sayRead the analysis against the record
• The missing-records list names real providersReferenced in the file, not produced
• The output arrives in a format you can work inSortable and pasteable, not a re-typing job
• The price is computable before you buyA per-page rate and a stated duplicate policy

Questions readers ask

What is the purpose of medical record review?

To establish, from the records themselves, what happened to the patient and when, so a claim, defense, evaluation, or underwriting decision rests on documented facts. The output is usually a chronology, summary, or report that a decision-maker can verify against the source pages.

Who performs medical record review?

Paralegals and attorneys, claims adjusters, legal nurse consultants, physicians, and increasingly AI software with a human verifying the output. Who should do it depends on whether the deliverable needs clinical judgment and a signature or a cited draft to work from.

How long does medical record review take?

Manual review is commonly estimated in minutes per page, so a thousand-page file is days of work. AI software produces a cited draft chronology in minutes; the human verification pass on top depends on the stakes of the matter.

What is the difference between record review and a chronology?

Record review is the whole activity; the medical chronology is its most common deliverable, a date-ordered timeline of medically significant events. A review can also produce narrative summaries, gap analyses, and billing checks from the same reading of the file.

What does medical record review cost per page?

Only 8 of the 328 vendors we read publish a per-page rate at all, and among those 8 the published range runs from $0.008 to $0.35 a page. That fortyfold spread covers 4 different deliverables sold in the same unit, so compare the deliverable before the rate. The one published statutory figure is California's $3.00 a page past 200 under 8 CCR 9795, which prices a physician reading a medical-legal file, not a software pass.

How many pages is a typical file?

There is no national figure. The best public data is Californian: across 198,774 medical-legal cases with any paid record review, RAND found a median payment of $939, which at $3.00 a page past the first 200 works out to roughly 513 pages, and a 95th percentile of $10,000, roughly 3,533 pages. A medical management company told RAND that packets of 10,000 to 15,000 pages occur.

Are duplicate pages billable?

That depends entirely on the vendor, and it is worth asking before you sign anything, because duplicates are a large share of a litigation production. RAND found no participant of any kind, on any side, who objected to removing duplicate records before they reach the reviewer. Medrecords AI does not bill duplicate pages.

Can AI do medical record review on its own?

No, and no serious vendor should claim it. Research published in PLOS Digital Health in 2025 found that 42% of GPT-4-generated summaries of emergency department encounters contained at least one hallucinated detail when checked against the source record, absent human review. The working arrangement is that software drafts and cites, exhaustively and in minutes, and a qualified person reviews, revises and signs.

Bates number

A unique sequential identifier stamped on every page of a production, so any page can be cited unambiguously by every party in a matter.

Chronology

A date-ordered table of medically significant events drawn from a record, 1 row per event, each row citing its source page.

Co-mingled records

Pages belonging to a different patient inside a production. A privacy incident and a factual hazard in the same page.

Comprehensive medical-legal evaluation

In California, the full evaluation billed under ML201, which includes the first 200 pages of record review in its fee.

Gap analysis

The finding of what a record references but does not contain: named providers with no records, referrals with no follow-up, date ranges with nothing in them.

Medical record abstraction

The term used in clinical research for extracting structured data from a chart. It is the same activity as extraction in chapter 7, and it is the one with a published error rate.

ML-PRR

The California billing code for medical-legal record review, set at $3.00 per page past the first 200 by 8 CCR 9795.

Near-duplicate

The same document appearing twice in different scans or printouts, so a byte or hash comparison treats the 2 copies as different documents.

Production

The set of records served by one party on another, usually Bates numbered as a unit.

Self-authenticating record

A record admissible without a live custodian, under Federal Rule of Evidence 902(11), on a certification plus advance written notice to the adverse party.

Specials

The documented medical charges in a claim, tied back to the treatment that generated them.

Supplemental report

A further report answering questions raised after the original evaluation. In California it is billed under ML203 at $650.00, and it is frequently what late records produce.

This is a field manual, not a study. We surveyed nobody outside our own reading of vendor websites. Every number in it comes from a public source, named at the point it is used, and the sources are these 7.

What we did not do

• We did not survey attorneys, adjusters, nurses or physicians for this document. RAND interviewed evaluators and stakeholders, and where a view is attributed to them we cite RAND rather than restating it as our own finding.
• We did not publish a national average cost, a national average file size, or a national average minutes-per-page for manual review. No credible public figure exists for any of the 3, and chapter 12 says so rather than estimating one.
• We did not present California's $3.00 per page as a national benchmark. It is the only statutory figure of its kind in the United States, which is exactly why it is worth quoting and exactly why it does not generalize.
• We did not present any vendor's own marketing claim as an independent finding. Where a figure comes from a vendor's website, including our own, the line says so.
• We did not publish an accuracy benchmark for AI record review, ours or anyone's. There is no independent one, and an internally produced number is a marketing claim with a decimal point on it.
• We did not name or link any other record review platform, service or directory in this manual.

RAND, the California Department of Industrial Relations, the California Division of Workers Compensation, PLOS, BMC and the authors of the cited papers are not affiliated with Medrecords AI, have not reviewed this document, and do not endorse it. Their work is public and you should read it.