This site is published by Medrecords AI, which makes 1 of the products it covers. We say which one wherever it appears. How we handle that.

Best AI for Expert Witnesses

A defensible AI workflow for a physician expert: worked example

By Ahmed Jemaa, Co-Founder and CEO of Medrecords AI. Updated .

Answer in brief: let AI do the sorting, keep your hands on the checking and the opinion, and write down each step as you go. The 8 steps below take a synthetic 1,912-page standard-of-care file from intake to signed report, with every AI step logged and every relied-on fact checked against its page.

Limits: We read every rule, ruling and product page on the date shown and did not test any product with patient records. Court summaries here are not legal advice; how your court treats AI use depends on the jurisdiction and the judge, so ask retaining counsel.

The synthetic case

Synthetic, no real patient. Every name, date, page number and finding on this page is invented to show the method.

Retaining counsel asks you, an emergency physician, whether the emergency department met the standard of care for a 58-year-old man seen for abdominal pain on 2024-03-14, discharged, and admitted 31 hours later with a perforated appendix. Counsel sends 1,912 pages from 6 sources: the emergency department, the hospital admission, the primary care practice, the surgeon, the imaging center and the pharmacy. Your report is due in 5 weeks.

8 steps from intake to signed report

8-step AI workflow for a physician expert, split into AI steps and steps the expert does Row 1: 1 intake and tell counsel, 2 deduplicate and index with a tool under a BAA, 3 page-cited chronology, 4 check relied-on lines against pages. Row 2: 5 chase missing records through counsel, 6 literature search then read in full, 7 write the opinion yourself, 8 finish the AI-use log and disclosure. Steps 2, 3 and the search in 6 are AI-assisted; the rest are the expert's. 1. Intake count pages, tell counsel 2. Dedupe, index record tool under a BAA 3. Chronology every entry cites its page 4. Check relied-on lines against pages 5. Gaps missing records via counsel 6. Literature search, then read in full 7. Opinion written by you, page by page 8. Log, disclose AI-use log and report paragraph Green: AI does the work, you check it. Amber: AI finds, you read. White: you.
Our workflow. AI touches 3 of 8 steps, and none of them is the opinion.

Each step on the synthetic file

  1. Intake. You count pages per source against counsel's transmittal letter: 1,912 in all. You email counsel naming the record tool and the literature tool you plan to use, and confirm that the record tool's plan is under a BAA. Counsel confirms the protective order allows it. Log row 1.
  2. Deduplicate and index. The record tool finds that the hospital, the surgeon and the primary care practice each sent copies of the same CT report, discharge summary and lab panels. In this synthetic file, 1,388 pages are unique. 2 pages are unreadable scans; you ask counsel for clean copies. Log row 2.
  3. Chronology. The tool produces 214 dated entries, each citing a page. You mark the 31 entries your opinion will rest on: triage vitals, the ED physician's exam, the lab results and their result times, the CT order, the discharge instructions, and the return visit.
  4. Check. You open the cited page for all 31 marked entries and for 20 others picked at random. You find 2 problems: a nursing note time entered as 14:10 that reads 16:10 on the page, and an entry that merged 2 separate nursing notes into 1. You fix both from the source pages and note them in the log. Neither is a reason to drop the tool; both are reasons you check.
  5. Gaps. The ED physician's note refers to a "repeat abdominal exam" with no matching note in the file. That could matter to the standard of care. You don't treat the absence as a finding. You ask counsel to request the full ED encounter and the EHR audit trail for the visit.
  6. Literature. You search for guidance on evaluating undifferentiated abdominal pain that was in force on 2024-03-14, using search terms only, with no patient data. The tool returns 12 papers; you read 5 in full and cite 4. 1 software summary misdescribed a study's population, which you would have repeated if you had cited the summary.
  7. Opinion. You write the standard-of-care analysis yourself. Each step of the reasoning cites record pages and named sources. You attach the chronology as an exhibit only after checking every entry in it, because an exhibit under Rule 26(a)(2)(B)(iii) is something you vouch for.
  8. Log and disclose. You finish the AI-use log and add the disclosure paragraph from our disclosure template. The report says what AI did, what you checked, and that the opinion is yours.

What this file would have looked like without the checks

Had you relied on the chronology unchecked, your report would have put a nursing reassessment 2 hours early, at a point that bears on how long the patient waited. Opposing counsel's expert, reading the same page, would find it. The question at deposition writes itself: "Doctor, you say 14:10. What does page 1,204 say?" (Synthetic page number.) A single wrong time doesn't sink an opinion. Having no answer for how it got there can sink your credibility.

That is the reason for the 31 checks. They are the facts the opinion rests on, and checking them is your job under Rule 702(b) whatever produced the chronology.

Verification time calculator

Budget the checking before you quote the case. Every default below is illustrative; put in your own numbers.

With the illustrative defaults the total is about 7.5 hours. That time is the defensible part of an AI-assisted review, and it is billable expert time, not overhead. If the number looks too high to quote, the fix is to rely on fewer chronology lines, not to skip the check.

Rules this workflow follows