General AI vs IME-specific tools
Short version: general assistants are good with language and weak with big files you must cite. IME record tools are the reverse. Use a general assistant on a BAA plan for short documents and literature, and an IME tool for the record file.
Limits: We read each product's help pages, terms and pricing on 2026-11-12 and did not test how any product handles PHI. Plans, prices and laws change, and nothing here is legal advice for your state.
The 4 tests, product type by product type
An IME report gets read by a claims examiner, an opposing attorney and sometimes a judge. Each of them can ask where a fact came from. That is why we score every product on the same 4 tests: whether its maker signs a BAA, whether it cites pages of your file, whether it copes with 1,000+ pages, and whether it holds your report format. Facts below were read from each maker's help pages on 2026-11-12.
| Product, plan | BAA | Cites pages of your file | Large file limits stated | Format control |
|---|---|---|---|---|
| ChatGPT Free, Plus, Pro | No | Not documented | 512 MB and 2 million tokens per document | Prompt only |
| ChatGPT for Clinicians | Yes, per clinician, in Settings | Cites journals, not your pages | Same as ChatGPT | Prompt only; starter prompts |
| Claude Free, Pro, Max | No | Not documented | PDFs up to 1,000 pages, 20 files per chat | Prompt only |
| Claude Enterprise, HIPAA-ready | Yes, 20 seat minimum | Not documented | Same upload caps | Prompt or projects |
| Gemini, personal account | No | Not documented | 100 MB per file, 10 files per prompt | Prompt only |
| Gemini in Google Workspace | Yes, Workspace BAA | Not documented | Not stated for Workspace; 1 million tokens on AI Pro and Ultra | Prompt, inside Docs |
| Microsoft 365 Copilot | Yes, Microsoft BAA via the DPA | Cites the work files it grounds on | Not stated on the page read | Prompt, inside Word |
| IME record tools (15 listed, ours included) | 7 of 15 say they sign one | Yes, the core feature | Built for it; 3,000 pages per case at InjuryMgmt | Templates or learned formats |
Sources and the rest of the products are in the dated tools table. "Not documented" means the help pages we read do not describe page-level citations to an uploaded record. You can ask any assistant for page numbers in a prompt. You then have to check every number, because nothing in these products ties the answer to a page.
Why a big IME file breaks a chatbot
Upload limits and reading limits are different numbers. A file can upload and still not be read end to end. Google says it plainly in its Gemini help: with a 1 million token context window, Gemini can understand up to about 1,500 pages of text, and past that window, answers can miss connections or details spread across the file. That describes how an IME file fails in a chat tool. In the synthetic file below, the detail that matters is a 2017 back complaint on page 1,412, not the headline diagnosis.
The published caps, as of 2026-11-12:
- ChatGPT: 512 MB per document and 2 million tokens per text or document file; up to 80 files every 3 hours.
- Claude chat: 500 MB per file, up to 20 files per chat, PDFs limited to 1,000 pages. Project files are 30 MB each, and their total must fit the context window.
- Gemini app: up to 10 files per prompt, 100 MB per file, and a 1 million token context window on Google AI Pro and Ultra (192,000 with Deep Think).
OpenAI's rule of thumb is that 1 token is about 4 characters, and 100 tokens about 75 words. Scanned records with tables, stamps and repeated headers are messier than prose, so measure a sample of your own pages.
File-fit calculator
Enter your page count. The default of 667 tokens a page comes from Google's own example (1 million tokens for about 1,500 pages). Change it if you have measured your files.
The calculator runs in your browser and sends nothing anywhere. It checks published caps only. It cannot tell you whether a model attends to every page inside its window.
The synthetic file, worked through
Synthetic example, no real patient. Claimant "J.D.", age 47, warehouse lift injury on 2023-03-14, lumbar strain with a later L5-S1 disc finding. The referral packet is 1,240 pages: 14 providers, 2 prior IMEs, physical therapy notes, pharmacy logs, and the same MRI report faxed 3 times (pages 212, 641 and 1,077).
What matters in this example is the second column's last line. The record tool did not make the call on apportionment. It showed you a page you would otherwise have missed, and you opened it. That is the whole value of a page citation: it turns "trust the AI" into "check page 1,412".
The chat tool's failure is not a hallucination in the strict sense. Every sentence it wrote may be true. The problem is an omission, and omissions are harder to catch because nothing on the screen looks wrong. Our failure modes table lists the 6 patterns to check for.
When a general assistant is the better pick
IME tools are narrow. A general assistant on the right plan beats them at work that is about language or medicine, not the file:
- Explaining an unfamiliar procedure note or abbreviation, with the record excerpt kept out.
- Searching literature on a causation question. OpenEvidence and ChatGPT for Clinicians cite published sources.
- Rewriting your own boilerplate: your methods paragraph, your fee letter, your scheduling emails.
- Proofreading a finished report for internal contradictions, on a plan with a signed BAA.
- Drafting a checklist of records to request for a given injury type.
If those tasks are most of your AI use, ChatGPT for Clinicians is free for verified US physicians and comes with an individual BAA. An IME company already on Microsoft 365 or Google Workspace can use Copilot or Gemini under the BAA it already has, for the features on each maker's covered list.
When an IME tool is the only sensible pick
- The file is over a few hundred pages, or arrives as several merged packets.
- You will cite records in the report and expect to defend each cite.
- Duplicates are common: the same imaging report from 3 providers, re-faxed progress notes.
- You want the output in your report layout every time, not the layout of the day.
Pick 2 tools from the tools table, run the same closed file through both, and count missed events and wrong pages. Speed is the least useful number in that test.