AI for each QME task: what it does, where it fails
Short version: AI earns its place on the parts of a QME report that are about finding things: pages, dates, priors, new records, page counts. It fails, or breaks the rules, on the parts that are about judging them: the history, the discussion, causation, apportionment and the ratings. Each section below names the rule that draws that line.
Limits: We read each product's own pages and the DWC sources on 2026-11-12 and did not upload records to any product. Prices, plans and rules change, and nothing here is legal advice; ask your own counsel about Labor Code 4628.
The 7 tasks at a glance
| Task | AI does well | AI fails at | Rule |
|---|---|---|---|
| Record review and chronology | Sorting, dating, page-citing, removing exact duplicates | Handwriting, faxed scans, events split across providers | LC 4628(a)(2), (c) |
| Declaration of records reviewed | Building the document list with dates and page ranges | Knowing which files arrived without a declaration | 8 CCR 41(c)(2), 9793(n) |
| History and exam dictation | Turning your dictation into text | Taking the history for you | LC 4628(a)(1) |
| Drafting from your notes | Layout, record sections, your dictated words in your template | Writing the discussion or conclusions | 8 CCR 41(c)(7) |
| Apportionment and impairment | Finding priors and prior awards with pages; table math | The how-and-why reasoning; picking the method | 8 CCR 10682(b)(12), (14) |
| Supplemental reports and depo prep | Showing what is new; listing conflicts with pages | Telling you what opposing counsel will ask | 8 CCR 9795 ML-203, ML-204 |
| Billing | Assembling codes, modifiers and time statements | Making AI a billable cost | 8 CCR 9795(b) |
Record review and chronology
- What AI does well
- Reads a 2,000-page file from both parties, sorts it by date and provider, removes exact duplicate pages, and builds a chronology with a page on every line. Good tools also flag priors, gaps in care and records that mention a study you never received. MediScan shows source page ranges on each entry, IME Coach opens the source page from each chronology line, and OctopusLM cites a page on every line.
- What fails
- Handwritten progress notes and faxed scans, where OCR misreads a date or a dose. Events split across providers, such as an MRI ordered by 1 doctor and read by another. Near-duplicates, such as the same report faxed with a different cover sheet, which only exact-match deduplication will miss. And the quiet failure: an event the tool never extracted, which no citation can reveal.
- What the QME must still do
- Labor Code 4628(c) says that when someone other than the physician excerpts the records, the physician reviews the excerpts and the entire outline and makes the further inquiries needed. Read the pages your opinions rest on. The DWC checklist adds that a physician should not rely on a summary from another source without checking it against the records summarized.
Tools that state page-linked output: MediScan, IME Coach, InjuryMgmt, OctopusLM, InPractice, ChartInsight, Workflow Labs, Dodonai, Docura, easyIME, Wisedocs, Medrecords AI (ours). Sources on the tool directory.
Declaration of records reviewed
- What AI does well
- Produces the list 8 CCR 41(c)(2) asks for: every document with its date, author, type and page range. It can compare that list against the records log the claims administrator attaches under 8 CCR 35(c), and show what is on the log but missing from the upload.
- What fails
- AI cannot tell which files came with the provider's declaration. Under 8 CCR 9793(n), documents sent without the declaration and its page-count attestation do not count as received, for billing or for your duties on records. If you upload everything in 1 batch, the tool treats it all as part of the file. Condensed deposition pages are another trap, covered on the home page.
- What the QME must still do
- Sort declared from undeclared submissions before upload, state the dates records arrived, and sign the verification of pages reviewed under penalty of perjury. The DWC checklist also asks you to say which records were missing and how that limits your opinions.
History and exam dictation
- What AI does well
- Transcribes your dictated history and exam findings into clean text, fast. IME Coach drafts the exam section from notes you type or dictate and states a BAA. ZyDoc offers job-type templates and letterheads, easyIME has its easyDictate tool, and Long Health EvalPath lists an AI scribe for the exam; none of those 3 states a BAA on the pages we read.
- What fails
- Any setup where software does the interviewing. Labor Code 4628(a)(1) reserves taking a complete history for the physician who signs. An intake questionnaire the worker fills in is fine as a source, but a chatbot that conducts the history is not you taking it. Transcription can also drop a negative ("denies numbness" becomes "numbness"), and 1 lost word reverses a finding, so read back every finding.
- What the QME must still do
- Take the history yourself, read the transcript against your memory of the exam, and state face-to-face time, which the DWC checklist lists under 8 CCR 49 to 49.9. Tell the worker before you record anything.
Drafting from your notes
- What AI does well
- Puts your own dictated or typed findings into your template: header, exam date and street address under 8 CCR 35.5(b), list of records, record summary sections you have checked, and the required declarations in the right place. IME Coach learns the layout from 2 or 3 of your past reports. Docura has QME, AME, P&S and supplemental templates.
- What fails
- 8 CCR 41(c)(7) says the evaluator writes every part of the report that discusses medical issues, the research used and the conclusions. An AI that generates the discussion from the records, even well, writes the part the rule reserves for you. The output also tends to sound certain where you would hedge, and to answer questions the parties never asked.
- What the QME must still do
- Write the discussion and conclusions in your own words. Use AI to turn your dictation into prose only if the content is yours, and note in the report which tool you used and for what. 8 CCR 35.5(c)(1) also asks you to address each question in the parties' cover letters; a checklist of those questions is a good AI job.
Apportionment and impairment sections
- What AI does well
- Finds what apportionment needs: prior injuries, prior awards, pre-existing degenerative findings, non-industrial conditions and the imaging that shows them, each with a page. InjuryMgmt states it calculates AMA Guides Fifth Edition ratings from findings you review and fills the PR-4 impairment fields; easyIME says its instructions feed impairment rating sections.
- What fails
- The DWC checklist asks for a non-conclusory discussion that meets the Escobedo "how and why" standard, with approximate percentages and the reasons behind them. A general assistant asked for apportionment will offer a percentage; the reasons are the part the checklist cares about, and they have to be yours. Table and grade choices on borderline findings, and whether an Almaraz Guzman deviation from the standard rating is justified, which the checklist asks you to explain, are judgment calls a tool cannot make for you.
- What the QME must still do
- Decide causation of the disability, the apportionment percentages and the rating method, and explain why. 8 CCR 10682(b) lists apportionment and the reasons for the opinion among the report's elements. Use AI only to make sure no prior injury page escaped you.
Supplemental reports and depo prep
- What AI does well
- Shows which records are new since your last report. That matters for billing: under 8 CCR 9795, ML-203 ($650) is not allowed for a supplemental that reviews information you already had. A record tool that answers questions with the page named also builds a deposition prep sheet: every statement about mechanism, prior complaints and work status, with conflicts side by side.
- What fails
- Version control. If the new records go into the same case as the old ones, the tool may blend them, and your supplemental will cite pages the parties say you already reviewed. AI also cannot predict cross-examination; it can only list the inconsistencies already in the record.
- What the QME must still do
- Keep each batch of records as its own upload, answer only the questions in the request, and prepare your own testimony. ML-204 pays $455 an hour for testimony time, with a 2-hour deposition minimum, whatever tools you used to prepare.
Billing
- What AI does well
- Assembles the pieces the DWC checklist expects in the report: the ML code, any modifiers with the required explanation (such as -93 for an interpreter), face-to-face time and the page verification. Practice software with AI help, such as Simplexam, generates, submits and tracks bills; Simplexam says its AI does not write reports.
- What fails
- Treating AI as a billable cost. 8 CCR 9795(b) says each ML fee includes record review, report preparation including typing and transcription, and overhead, and Labor Code 4628(d) limits charges to direct professional charges plus reasonable costs of tests and clerical expense. AI page counts can also differ from the declared count, which the DWC checklist says you must reconcile, not override.
- What the QME must still do
- Choose the code and modifiers, sign the page verification, and absorb the AI cost as overhead. The record review page count has its own rule set in 8 CCR 9795.
Common questions about AI by QME task
Which QME task benefits most from AI?
Record review and chronology, because it is the slowest part of a QME file and the one where page citations let you check the output quickly. It is also the task Labor Code 4628 watches most closely, so treat the AI output as excerpts you verify.
Can AI write the apportionment section of a QME report?
It should not. 8 CCR 41(c)(7) has the evaluator write all discussion of medical issues and conclusions, and apportionment needs a how-and-why explanation. AI can find every prior injury page so your reasoning starts from a complete record.
Is AI dictation allowed for the history in a QME report?
Transcribing your own dictation is clerical work. Labor Code 4628(a)(1) keeps taking the complete history with the physician, so the AI should record what you ask and hear, not run the interview.