A legal team remains responsible for every date and provider name used in a demand or filing, even when AI performs the first review.
AI medical record review uses review software to convert a production into a sorted, summarized timeline that a medical chronology paralegal verifies against the source.
That software runs as a fixed sequence of stages, commonly called a processing pipeline: intake, OCR, entity and event extraction, deduplication, and chronology generation. This article explains each stage, its documented failure modes, and the outputs a paralegal must verify.
Intake and Document Handling: Formats, Volume, and Scanned Pages
Intake converts the production into a uniform sequence of page images with stable page or Bates identifiers. Image quality here caps the accuracy of every later stage, and the software can only process what the medical records retrieval step obtained, so the intake page count sets the baseline for the missing-record check. Page volume, scan resolution, and the proportion of handwritten pages determine how much manual handling a production will need.
Production size also decides how the tool must chunk the file, because no processor reads an unlimited number of pages in one pass. Google's Document AI limits cap Enterprise Document OCR at 15 pages per synchronous request and 500 per batch request, so a production of any real size is split and processed in segments before anything is assembled into a timeline.
The first quality gate at intake checks resolution. Google Cloud Document AI sets a 200 dpi minimum and recommends 300 dpi or higher, and its Enterprise Document OCR scores each page across eight defect dimensions, among them blurriness, faintness, glare, and undersized fonts. Pages scoring below the readability threshold get routed for manual review.
Handwritten pages get separate handling because handwriting recognition has narrower support. AWS Textract documentation supports handwriting recognition only in English, and Azure OCR documentation supports handwriting and cursive only for English.
Where OCR and Text Extraction Quality Degrades
OCR quality governs everything downstream, so the useful question is how far apart document types sit. That gap tells a reviewer which pages to pull. In the EACL Industry Track benchmark of 60 photographed obstetric ultrasound reports degraded by blur and shadow, Qwen-2.5 VL still reached a character error rate (CER) of 0.031 on printed text.
Handwriting sits an order of magnitude worse. In the UBOMI BUHLE handwriting benchmark, Gemini 3.1 Pro posted a word error rate of 0.50 and CER of 0.31 on real handwritten maternity forms. Handwritten pages therefore go to manual review regardless of the page confidence score.
The ConfBERT paper notes that detection methods struggle with real-word errors, misreadings that still form valid dictionary entries, so a garbled drug name can pass spell-check as another real word.
Pages likely to have failed extraction share visible signals:
- An image-only page that produced no text must be flagged as potentially unreadable rather than automatically treated as blank.
- A page confidence score below the tool's threshold flags a re-read; AWS Textract guidance suggests thresholds as low as 50% for archival handwritten notes and 90% or higher for high-stakes decisions.
- Fax-embedded pages, rotated scans, handwritten forms, and tables crossing a page break warrant manual review; Google Document AI documents that a split table may be returned as two tables.
Pages carrying any of these signals get flagged as they are processed and pulled for manual review later. That keeps the verification workload proportional to the production rather than uniform across it.
What Entity and Event Extraction Produces
Extraction turns OCR text into structured fields: provider, date of service, diagnosis, procedure, medication, and billing entry. Each fails in its own way, and three carry the most weight in a personal injury or workers' compensation file. The rates below come from related clinical benchmarks and may not transfer directly to legal medical-record productions.
The Consolidated Clinical Document Architecture (C-CDA) treatment section holds "prospective, unfulfilled, or incomplete orders and requests only"; an extractor that ignores section context records an ordered imaging study as a performed one.
Deduplication: Page-Count Reduction Versus Information Loss
Deduplication collapses pages appearing more than once across producing providers into one canonical copy, and the volume involved is substantial. A JAMA Network Open study of 104,456,653 notes found 50.1% of electronic health record text duplicated from earlier text about the same patient, 45.9% of it from a different author. Two failure modes turn that reduction into loss.
Suppressed Copies That Carried Provenance
The clinical content of a repeated consult report is redundant, but the location of each copy is not. That a primary care physician held a cardiologist's report on a given date is evidence of what the treater knew, so the deduplication log needs to record where each suppressed copy appeared, by page or Bates range.
Amended Notes Collapsed Into Originals
An American Health Information Management Association (AHIMA) practice brief advises that any amendment be approved by the entry's author and that previous versions remain accessible through revision history. A pass treating the amended version as a duplicate of the original erases both the amendment and the question of why the record changed.
Both failures are invisible in a before-and-after page count, so the reviewer asks for the list of collapsed pages and samples it against the production.
How the Chronology and Summary Are Generated
The chronology stage drafts a narrative for each encounter and sorts the entries by date of service, ordering same-day encounters by facility and time where the record supplies it. A usable entry carries date, provider, facility, encounter type, findings, treatment, and a citation to the source page. An AI medical chronology that returns entries without page citations is an unverifiable draft.
Omissions outnumber fabrications by a wide margin. The Draper et al. evaluation of seven commercial AI scribes found omissions accounted for 83.8% of all errors and 60% of high-severity ones. That figure describes an adjacent clinical summarization task and may not transfer directly to legal record review, but it sets the priority: check for what is missing before checking what is wrong.
Chunking is a documented path to sequence error. A 2025 JAMA Network Open study reported that feeding summaries of summaries into the model frequently produced chronological inaccuracies, a workaround its authors attribute to older context limits.
Verification Checkpoints for the Reviewing Paralegal
AI review moves the paralegal from first reader to verifier, which relocates the error risk that the manual review process guide assumes. Verification concentrates on the outputs where the documented failure modes land, and every check runs against the source page rather than the tool's confidence display. Verify dates and provider attribution first. Then determine whether each service was ordered or performed. Errors in these findings change what the chronology proves.
- Confirm every date against the date-of-service field on the cited page rather than relying on the print date or fax header. A dictated note's creation timestamp may differ from the date of service. Cross-check billing entries, whose CPT and ICD-10-CM codes form a second timeline.
- Check the provider against the signature or attestation block, not the letterhead; classify the name as treating, referring, ordering, or co-signing. Flag entries where a facility name stands in for a clinician.
- Pull the pages flagged for handwriting, fax distortion, rotation, or another extraction risk, then confirm each one produced a corresponding entry. Treat "[illegible]" markers, ellipses, and entries far shorter than the visible page as extraction failures requiring manual entry.
- Confirm each specialist visit and every procedure or therapy against a completed-service note. An order or referral does not establish completion, nor does a prescription. Label unresolved orders as "ordered, completion not documented in received records."
Each check produces a record of what was verified and against which page. That record is what makes the chronology defensible when an entry is challenged.
Why Source Traceability Is a Defensibility Requirement
A verified entry is one a reviewer can trace to a specific page in the certified production, and that trace is what an entry has to survive when opposing counsel challenges it. National Association of Legal Assistants (NALA) guidance advises that every substantive claim be traceable to specific text in a document the firm controls, and an AI entry without a checked source page fails that standard on its face.
Verification is not a one-time acceptance test. The NIST AI Risk Management Framework treats validity and reliability in deployed systems as something confirmed by ongoing testing and monitoring in production rather than established once at rollout, which means a new tool or an updated model resets what the firm has actually verified. A wrong date of service or wrong provider expands review of that stratum, and recording an order as treatment does the same.
HIPAA, Security, and Chain of Custody in AI-Assisted Review
Vendor review covers HIPAA status, contractual safeguards, and security controls. Firms commonly require a signed business associate agreement, role-based access control, and a written statement that client records are not used to train the vendor's models. A current SOC 2 report supports vendor-risk assessment, though HIPAA does not mandate one.
Chain of custody runs through the page and Bates identifiers assigned at intake, and the production itself rests on the custodian's certification that the copy matches the original. AHIMA guidance adds that whoever certifies or testifies to those records must understand the policies and processes keeping them accurate. The intake record, deduplication log, OCR confidence record, and page-citation map document the AI-processing stage, and together they let counsel or a records custodian give that account.
The same standard applies to the review tool's own access logs. AHIMA audit-trail guidance calls for capturing the user, application, workstation, document, event description, and timestamp, and NIST guidance holds that the source of copied information, who copied it, and what was edited must all remain traceable.
Verification Priorities in Personal Injury and Workers' Compensation Files
Which entries earn full source-page verification depends on what the file has to prove. Provider-attribution errors concentrate at care transitions in both practice areas, so those entries get a signature-block check even when page confidence is high.
Personal Injury: The Incident Date and First Examination
The incident date, the first post-incident examination, and each escalation of care carry the weight here. A wrong date on the first examination undermines the causation argument for every entry after it, which is why those entries get checked in full rather than sampled.
Workers' Compensation: The Pre-Injury Baseline
A workers' compensation defense will argue that prior treatment of the same body part makes the current condition unrelated to the work injury. The reviewer labels pre-incident records as baseline and confirms the chronology has not merged a prior visit into the post-incident sequence.
Both: Documented Gaps Versus Unproduced Records
A defense argument commonly treats an unexplained gap as evidence against continuity of the injury. An entry reading "last documented visit [date/Bates]; next documented visit [date/Bates]" with retrieval status attached is defensible; a gap that turns out to be an unproduced imaging-center file is not.
A Defensible AI Medical Record Review Requires Source Verification
An AI review pipeline fails in predictable places, which is what makes those failures checkable: low-resolution and handwritten pages, hedged diagnoses read as confirmed, amended notes collapsed into originals, and events dropped in summarization. A defensible chronology retains the page-citation map and deduplication log so a reviewer can trace any entry back to the production.
Tavrn's services fit this controlled workflow by supporting medical-record processing while legal teams retain responsibility for source-page verification, security review, and the final work product. Martay Law Office processes records exceeding 22,000 pages and grew case filings 65% year over year at that volume, and firms working at similar scale can automate medical chronologies without giving up the verification record.
To learn more, book a demo.
