Review an AI medical summary in two directions: verify its statements against the original sources, then inspect the sources for important information the summary omitted. Keep unsupported wording, unresolved contradictions and missing material visible. A citation or fluent paragraph is a starting point for review, not proof that the summary is complete or accurate.
Define the review before running the tool
Define the intended use, approved inputs and responsible reviewer before generating a summary. Confirm that the firm permits the particular tool and its data handling for this material. A first-pass organizational aid and a document relied on for a substantive decision may require different review depth; the responsible professional should set that depth.
ABA Formal Opinion 512 discusses understanding generative AI's limitations and using appropriate independent verification or review. It does not certify any particular medical-summary product or prescribe this checklist as a compliance standard. Consult the rules and professional guidance applicable to the firm when deciding how AI-assisted work may be used.
Check what the summary says
For each statement under review, open the referenced original and inspect the surrounding context. Check that the source belongs to the right person and encounter, supports the wording and preserves qualifications. A correct page number can still point to text that says a symptom was denied, a diagnosis was considered or a procedure was planned rather than completed.
- Does the linked source support this exact statement?
- Are encounter dates and document dates distinguished?
- Are planned, reported and completed events kept separate?
- Has uncertainty or a negative finding been dropped?
- Does nearby context change the meaning?
Separate extraction from inference. If a source lists a date, a summary can report that dated entry; it should not silently convert the entry into a broader clinical or causal conclusion. Preserve the source's uncertainty and attribute conflicting statements. The medical chronology guide explains why a chronology organizes available sources without replacing other verification work.
Check what the summary leaves out
Next, review the original material for omissions using the task's agreed questions. Checking only the claims already written cannot reveal a provider, event or qualification the summary never mentioned. Use a source inventory to see which files and periods were included, and flag unreadable or unsupported files rather than assuming they were successfully processed.
A public AI-and-paralegals discussion includes accounts of missing facts and the time needed to recheck outputs. Those accounts concern different tools and are not a benchmark. A public LinkedIn example from PI attorney Brooks Schuelke similarly presents AI-assisted document comparison as an imperfect starting point. These observations support testing the actual workflow, not declaring a universal success or failure rate.
Record corrections and unresolved questions
The downloadable Excel review worksheet includes a fictional example and blank review rows. Record omissions, conflicts and input-coverage issues as separate rows; retain the summary version, intended use and approval scope with the log. Use a correction log that preserves the original output, source location, reason for the change and reviewer decision. Distinguish unsupported claims from omitted facts, date mismatches, identity problems and unreadable evidence. If a question needs a provider response or professional interpretation, assign it separately rather than hiding it in a revised paragraph.
Swipe across the table to see every column. Keyboard users can focus the table and use the arrow keys.
| Check | Record | Possible action |
|---|---|---|
| Support | Output statement and original source location | Confirm, narrow or remove unsupported wording |
| Omission | Relevant source detail absent from the summary | Add supported detail or flag for the reviewer |
| Conflict | Both sources and their dates | Preserve the conflict and assign clarification |
| Input coverage | Files, periods and unreadable material | Resolve missing input before relying on completeness |
| Decision | Reviewer, version, scope and open questions | Approve the defined use or keep the work pending |
Fictional example: a summary says a patient completed six therapy visits, but the supplied packet contains five dated visit notes and a plan for another visit. The reviewer replaces the unsupported statement with what the packet supports and flags the remaining appointment status as unresolved. Missing confirmation does not become a missed visit or a completed visit.
Measure the whole task
Measure preparation, generation, source review, corrections and follow-up separately. Include failed runs and incomplete packets. A faster first draft is useful only in the context of the time and quality needed to reach the firm's accepted result. The hold-time cost worksheet offers the same whole-task principle for provider calls; use your own observations rather than invented savings.
For a pilot, choose an approved test set and have a qualified reviewer establish the relevant source facts before comparing outputs. Record the sample size, document types, unresolved items and review method. Do not extrapolate a few clean examples into an accuracy percentage for every case, handwriting style or record packet.
Make approval an explicit decision
Approval should identify the version, reviewer, scope and remaining limitations. The missing-provider audit can turn unresolved references into a separate worklist. VerifyMD keeps source review and human approval central to its supported document workflow; it does not guarantee summary accuracy. Book a workflow demo to evaluate a fictional source discrepancy and any provider follow-up, which requires setup and owner activation for live calls.
Explore provider follow-up and source review for your firm. Live calls require agreed setup and activation.
Book a demo