AI medical records summaries for personal injury cases: what to automate and what to check

Where chronology time goes

Ask a personal injury case manager where the week goes and the answer is usually records. Records and bills arrive from every provider who treated the client, in different formats, often hundreds or thousands of pages for one person. Someone reads every page and builds the chronology: date, provider, complaint, diagnosis, treatment, and the billed and paid amounts. The same person looks for gaps in treatment and earlier injuries, chases missing records and itemized bills, totals the medical specials and hands the file to the attorney for the demand.

On a contingency fee, every day between treatment complete and demand sent is a day the fee sits still. Case managers carry most of this work. Time recovered here creates capacity, because a contingency fee does not grow with the hours spent on a chronology. It improves margin when the firm lowers its costs or uses that capacity for more cases.

Texas gives the same records a second job. Rule 194.2(b)(10) of the Texas Rules of Civil Procedure requires initial disclosures in an injury suit to include "all medical records and bills that are reasonably related to the injuries or damages asserted", or an authorization to obtain them. Under Rule 194.2(a), those disclosures are due within 30 days after the first answer or general appearance, and a party served or joined later has 30 days from being served or joined, unless an agreement or court order sets a different time. A well-indexed record set serves the demand before suit and the disclosure after it. This guide is general information, not legal advice, and the responsible lawyer must check the rules and orders that govern each matter.

What the agent should write and what it should never write

A workable rule: let the agent do anything a careful new case manager could do from the records alone, and keep anything that needs judgment about liability or value with the attorney. The agent reports what the records say and where they say it. It never writes a medical opinion, and it never decides what a case is worth. In practice the line falls roughly as the table shows.

Every entry needs a page cite

A chronology without page cites has to be checked by reading the records again, which defeats the purpose. With a cite on every entry, the reviewer clicks from the line to the page and confirms it without hunting. Opinion 512 asks for exactly this kind of independent verification. It says a lawyer may not rely on AI output without an appropriate degree of review, and the lawyer remains fully responsible for the work whatever level of review they choose.

Scans and handwriting need their own rule. Typed records and clean scans read well, while handwritten clinic notes and poor faxes often come through garbled. So the agent gives each entry a confidence score and marks anything below your threshold for a person to check against the page image. Opinion 512 cites a Stanford study that found leading legal research tools hallucinated between 17% and 33% of the time. That figure comes from research tools rather than chronologies, yet it shows why a cite on every line matters: it makes each answer checkable.

Write the review steps down. The case manager checks every flagged entry first, then works through the unflagged entries and the specials totals against the bills. The responsible attorney sets how deep that review goes, based on the validation results and the case, and the review always looks for material omissions, including visits the agent did not flag. Each correction a reviewer makes is logged, so the next accuracy test can count it.

Test on closed files before live cases

Opinion 512 also gives a practical way to set the level of review. Its example is a lawyer summarizing a large set of contracts, who need not check every summary by hand if the tool was first tested for accuracy on a smaller subset. A firm can apply the same idea to chronologies, with the responsible attorney deciding how much review the results justify. Take ten closed files whose chronologies your team already finished and trusted, and run the agent on copies inside the firm's systems. Closed files carry the same confidentiality duties as live ones. Then compare the two versions line by line.

Count four things: visits the agent missed, wrong dates or amounts, prior complaints it failed to flag, and missing records it failed to notice. Agree a threshold for each before you look at the results, so the test cannot be bent to fit. Keep sampling after go-live, because record sets change as new providers and new formats come in.

Protecting medical records

Before medical records go into any AI workflow, confirm the firm's authority to use and disclose them, for example under the client's authorization. The firm's own duty of confidentiality applies on top. Opinion 512 asks lawyers to evaluate the risk that client information is disclosed outside the firm, and to read the terms of use of any tool before it touches client data. Texas Opinion 705 recommends reviewing and, where needed, renegotiating a tool's terms of service.

For a build, the agent runs in the firm's Microsoft 365 or cloud tenant, built to HIPAA standards. Before go-live we confirm the model provider's terms bar training on your data. Only the case team can see a client's records through the agent, and every page it read and every line it wrote is logged. Records are kept or deleted under the firm's retention policy. A consumer chat account often fails these tests on its terms alone, so case managers should not paste records into one.

What PI software already does

Several products already target this job. Filevine lists MedChron among its features, and EvenUp sells medical chronologies and demand drafting to injury firms. EvenUp's home page says its product can "Cut 9+ hours of manual work per case", which is a vendor claim we have not checked. If a product fits your template and your review steps, buy it, and we will say so during the sprint.

A custom agent makes sense when the firm's own way of working is the point: your demand template, your case management system, the flags your senior attorneys care about, and a review queue that matches how your case managers already work. It also makes sense when chronologies are one step in a longer workflow that includes records requests or the disclosures in a filed case.

Worked example: what the hours look like

This is an illustrative example with assumed numbers, not a client result. Take a 25-person personal injury firm where 5 cases a week reach treatment complete. Today each case takes about 300 minutes for the chronology and 180 minutes for the demand draft, 480 minutes in all. Five cases at 480 minutes is 2,400 minutes, or 40 staff hours a week.

Assume 4 of the 5 are routine: one injury event, under about 1,000 pages, no prior claims. With the agent, the case manager spends about 90 minutes checking the chronology against its page cites and the attorney about 90 minutes editing the demand sections, so 180 minutes per case. Four routine cases take 720 minutes. A fifth case with surgery and prior injuries still takes the full 480. The total falls to 1,200 minutes, or 20 hours, which gives back 20 staff hours a week.

Hours are only half the measure. If cases wait an assumed 10 business days for a demand today, a target of 3 days is the number to track. Surgery cases and very large record sets lower the routine share, so count a real month of closed files before trusting any estimate, including this one.

Starting small with chronologies

Pull last month's closed files and tag each one routine or complex. Pick ten routine files and run an agent on copies inside the firm's systems. Then compare its chronologies with the ones your team built. Ask your most careful case manager to mark every miss.

That test is roughly what our five-day AI Workflow Sprint does, for $2,500. You get a working prototype on redacted or synthetic samples and a fixed quote for the build, and the fee comes off that build. Full refund if we do not find 10 staff hours a week.

Common questions