Can AI summarize medical records for a personal injury case?
A capable model reads a 214-page production in about a minute. What to ask it for, what comes back wrong, and which parts of a record review are still yours.
Yes, and the part worth arguing about is not whether it can read them. A capable model will take a 214-page production and hand back a dated chronology faster than you can find the index. The question is what you asked it to produce, and how much of that output you are willing to put your name on without checking.
This page is about the work itself: what to ask for, what comes back wrong, and which parts of a record review are still yours. If you are asking whether you are allowed to upload the records in the first place, that is a different question and it has its own guide.
Three jobs it does well, one it does badly.
The chronology
Date of service, provider, complaint, findings, treatment, next appointment. This is mechanical extraction across hundreds of pages and it is the single biggest time sink in a personal injury file. A model does it in a minute and does it consistently, which a tired paralegal at 6pm does not.
The billing picture
Charges, adjustments, write-offs and what was actually paid, pulled out of ledgers that arrive in four different formats from four providers. Ask for it as a table with a page reference on every line, because you are going to check the total before it goes anywhere near a demand.
The gaps and the prior mentions
Where treatment stopped for eleven weeks. Where a 2019 note mentions the same shoulder. These are the two things the adjuster will find, so you want them first, and a model reading every page will surface them where a skim will not.
What it does badly is causation. Whether the gap in treatment is explainable, whether the prior shoulder complaint is the same injury, whether the treating orthopedist survives contact with the defense expert: none of that is extraction, and a model asked for it will produce a confident paragraph that reads exactly like the paragraphs that are correct. Use it to find the facts. Draw the inference yourself.
Rivera v. Hollis · records
214 pages readAsk for page references, not prose.
The difference between a summary you can use and a summary you have to redo is whether every line traces back to a page in about four seconds. Make that the first instruction, before anything about tone or length. A chronology with citations is a working document. A fluent narrative without them is something you end up reading the whole record to verify, which is the job you were trying to avoid.
- One line per date of service. Provider, complaint, objective findings, treatment, page. Not a paragraph about the course of care.
- Flag rather than conclude. Ask it to mark gaps over thirty days and every mention of the injured body part before the incident date, and to stop there.
- Separate the bills from the records. Two outputs. Mixing them produces a total nobody can audit.
Six steps from production to usable summary.
Check the scan before you check the model
Most bad output is bad input. A fax of a fax at 150 dpi, or handwritten intake notes photographed at an angle, comes back thin no matter what reads it. If a page is unreadable to you at full zoom, assume it is unreadable to the model and pull a clean copy from the provider.
Deduplicate the production
The same ER visit routinely appears three times across three productions, and a chronology that lists it three times looks like three visits. Ask for duplicates to be collapsed with the page ranges kept, so you can still point to each copy when the carrier claims you missed one.
Ask for the chronology first, alone
Do not ask for a chronology, a bill summary, a gap analysis and a narrative in one request. Each output has a different failure mode, and bundling them means checking four things at once. Run them separately and check each as it lands.
Spot-check ten lines against the pages
Pick the first, the last, the two largest dollar figures and six at random. Transposition is the error to hunt for: a correct date attached to the wrong provider, or a charge read from the adjacent column. If ten are clean the rest usually are. If two are wrong, re-run rather than repair.
Reconcile the billing total by hand
Once. Against the provider ledgers. This number goes into a demand and then into a settlement statement, and it is the one figure in the file that a client, a lienholder and the bar can all check independently.
Do the thinking part yourself
Causation, apportionment, whether the gap is defensible, what to concede and what to fight. You now have the facts of a 214-page record in front of you an hour after it arrived instead of a week. That hour was the point.
Read the whole record the day it lands.
Upload the production to the matter it belongs to and get back a dated chronology, a billing table and the gaps, each line pointing at a page. The agent is already inside the file, so there is no context to paste, and the models it runs on keep nothing. Included in the $49 plan.
Common questions about AI record review.
How long does 200 pages actually take?
Can it read handwritten notes?
Do I still need a nurse reviewer?
Will it invent a treatment date that is not there?
Can it write the demand letter from the same records?
Is it safe to upload the records at all?
Read the record the day it arrives.
The agent reads what you upload against the matter it belongs to, and keeps nothing. $49 per user per month.