Every medical malpractice and personal injury file arrives the same way: a production of records that is too large to read twice. Somebody has to turn it into a document the attorney, the expert and the adjuster can work from. That document is a medical record summary, and in 2026 a growing share of them are generated by software before a human touches them. This article is about how to tell a summary you can rely on from one you cannot, regardless of who or what produced it. It is written for the attorney or paralegal who has to decide, on a Tuesday afternoon, whether the summary on the screen is good enough to build a demand, a deposition outline or an expert referral on.
The word "summary" hides two different documents. The first is a narrative: a few pages of prose that tell the story of the care in order. It reads well, it is easy to produce, and it is nearly useless in litigation, because nothing in it can be checked without re-reading the record it was supposed to replace. The second is a working document: a dated list of what happened, each entry tied to the page it came from, with the reviewer's observations kept separate from the facts. That is the one an expert will accept, an adjuster will engage with, and opposing counsel cannot dismiss with "where does it say that?"
Software makes the first kind cheaply and in volume. It can also make the second kind, but only if it was built to, and only if the person receiving it insists on it. The failure mode is not that the software lies, although it can. The failure mode is that a fluent, confident narrative gets treated as verified work product when it is a draft that nobody has checked. A summary that was never checked against the record is an assertion, and an assertion is worth exactly what it costs to produce.
A summary you can defend has five parts. If a document is missing any of them, it is not finished, whatever the cover page says.
The backbone is a date-ordered table of events, one row per event, across every provider, with the source document and the page number on every row. Not "hospital records"; the specific note and the specific Bates or PDF page. The medical chronology guide sets out the eight columns that survive deposition and a filled-in sample. The short version: date and time, provider, facility, event stated the way the record states it, source page, a short quote, significance, and a flag. The first six are facts. The last two are analysis and must be kept in their own columns so a reader can strip them out and still have a document that stands on its own.
The citation is the whole point. A row with a page number can be verified by anyone in ten seconds. A row without one has to be taken on faith, and nobody in litigation is paid to take things on faith.
A useful summary points at the places where the care departs from what the record itself says should have happened: a critical lab called to a nurse at 21:12 with no provider entry until 23:50; an imaging read recommending follow-up with no follow-up order; a discharge instruction that contradicts the consult note. These are not standard-of-care opinions. They are observations about the record, stated with the interval computed and the two pages cited. The expert decides whether a deviation is a breach. The summary's job is to make sure the expert sees it.
The most valuable line in many summaries is the one that says what is not there. A gap is an interval in which the record contains no entry where one would be expected: a four-hour stretch in an ICU flowsheet, a missing operative report the discharge summary refers to, an outside consult that was requested and never appears. A summary should list gaps as their own entries, cite the pages that bracket the gap, and separately inventory records that were referenced but not produced. Half of the follow-up subpoenas in a case come from this list.
Where wording matters, the summary should reproduce it, in quotation marks, with the abbreviations and errors as written. "Pt c/o CP x2d, will obtain labs" is evidence. A paraphrase of it is not. Quotes settle the "is that what it says" argument before it is raised, and they preserve the record's own language, which is often more damaging or more exculpatory than any paraphrase.
Anything that is the reviewer's judgment, whether the reviewer is a nurse consultant or a language model, belongs in a column or section marked as such. Mixed together, a single wrong inference contaminates every fact around it and hands opposing counsel a way to attack the whole document. Separated, the facts survive even when the analysis is challenged.
Some summaries should go back to whoever produced them without further reading.
You do not need to re-read the record to know whether a summary can be trusted. You need a sample, chosen so that a bad summary is likely to fail it.
If a vendor, whether a review service or a software product, is going to produce summaries you will attach your name to, these are the questions worth asking before the first file goes in. The answers matter more than the demo.
The obligations do not change because the first draft came from software. The American Bar Association's Standing Committee on Ethics and Professional Responsibility addressed the use of generative tools directly in Formal Opinion 512 (July 29, 2024). Its themes are the familiar ones: competence, which includes understanding what the tool can and cannot do; confidentiality of client information put into the tool; communication with the client where it matters; candor toward the tribunal; supervision of the people and systems doing the work; and reasonable fees that reflect the time actually spent. The opinion is worth reading in full before adopting any tool. For the purposes of this article, the practical consequence is simple: the lawyer who signs the demand or takes the deposition is responsible for every fact in the summary, which is why the summary has to be checkable.
The thin product page at medical record review software covers the category at a high level. This article is the longer answer to the question that page raises: what does "good" look like, and how do you know when you have it.
The Records Analyzer was built around the checklist above rather than around a narrative. It takes the production as uploaded, indexes every page, and returns a dated chronology in which each row links to the page it came from, so the citation check in the ten-minute test is a click rather than a search. Deviations and gaps are reported as their own entries with the bracketing pages cited. The analysis is kept apart from the extracted facts. Pages the system could not read are listed, not dropped. You can see the citation behaviour without an account on the 30-second citation demo, and the chronology guide shows what the finished table should look like when it is done by hand, which is the standard the software has to meet.
It does not decide whether the care met the standard. That is the expert's work, and a summary that does it for them is a summary an expert will refuse to sign. What the software does is make the expert's first read shorter and the attorney's spot check faster, and it does that by producing a document in which every line can be traced back to the page. That is the whole difference between a summary you can use and one you cannot.
A date-ordered chronology with a page citation on every row, the deviations in the care stated as facts with the interval computed, a list of gaps and records referenced but not produced, short exact quotes where wording matters, and a clear separation between extracted facts and the reviewer's analysis. A summary missing any of these is a draft, not work product.
Not for anything that will be tested. Without citations, no statement in the summary can be verified without re-reading the record, an expert cannot trace a fact to its source, and opposing counsel can dismiss it with a single question. Send it back and ask for the cited version.
Count the pages in the production against the rows in the chronology, open three citations from different parts of the document and confirm the event appears on the cited page, verify the one interval the case turns on against both bracketing pages, confirm there is a gap list and a coverage statement, and read the analysis with the facts covered to see whether every claim is supported by a row. That takes about ten minutes.
No. ABA Formal Opinion 512 (2024) applies the existing duties of competence, confidentiality, communication, candor, supervision and reasonable fees to generative tools. The lawyer who relies on the summary is responsible for its contents, which is the practical reason every fact in it needs to be checkable against the record.
Upload a production to the Records Analyzer and get back a dated chronology in which each entry links to the page it came from, with deviations and gaps listed separately. Free to start, no card.
Open the Records Analyzer →Sources. American Bar Association, Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024). The chronology column layout and sample are from the MedLegal AI medical chronology guide. No statistics about summary accuracy or time savings are cited in this article because no independent published figure was located; treat any such claim from any vendor, including us, as a claim to be tested on your own records.
MedLegal AI is software, not a law firm. This article is general information for attorneys, paralegals, legal nurse consultants and patients; it is not legal or medical advice and does not create an attorney-client relationship. Verify every statute, rule and figure against the current source before relying on it. Questions: [email protected]