A demand workflow can only reason from what it can reliably read. A clean, text-searchable treatment record may move through extraction and chronology review smoothly; a third-generation fax, a tilted emergency-room scan, or a handwritten therapy note can turn the same workflow into a false-confidence problem.
The right operational question is not whether an AI tool can process the file. It is whether the file is reliable enough for automation to continue without a human resolving the uncertainty first. Plaintiff PI firms need a defined stop-and-review rule before unreadable source material becomes a polished but incomplete demand draft.
Why poor source quality is a demand-drafting risk, not just an OCR problem
Optical character recognition converts pixels into machine-readable text. When the source is crisp, OCR can make large record sets searchable and easier to organize. When the source is degraded, the output may still look orderly while omitting, merging, or misreading facts that matter to the demand theory.
Common failure points include faint fax copies, black page borders, skewed scans, double-sided pages with bleed-through, cropped margins, handwritten provider notes, checkboxes, stamps over text, unusual abbreviations, and records assembled from several facilities at different resolutions. A human reviewer sees the visual warning. Downstream software may receive only the extracted text and therefore lose the signal that the source was uncertain.
That matters because small extraction errors can change the legal meaning of a record. A missing “no” before a symptom, an unread treatment-gap explanation, a misread dosage, or a date assigned to the wrong encounter can distort causation analysis. The draft may remain grammatical while its factual foundation is incomplete.
The danger is greatest when the output is persuasive. Staff can spot obvious nonsense. They are less likely to question a coherent paragraph that silently excludes two unreadable pages. The workflow therefore needs to preserve the distinction between “not present in the record” and “present but not reliably readable.” Those are not the same evidentiary condition.
Where automated processing should stop
A good escalation policy does not require every imperfect page to be manually transcribed. It identifies defects that could materially affect liability, causation, treatment history, damages, or the credibility of the package.
The text and the image disagree
Human review should be mandatory when the OCR output appears complete but the page image contains handwritten additions, marginal notes, arrows, strikeouts, or checked fields that are missing from the text layer. The same rule applies when a scanned form uses columns or tables and the extracted text collapses unrelated fields into one sequence.
Comparison should focus on high-value sections: assessment, history, restrictions, referrals, diagnostic impressions, treatment plans, missed-appointment explanations, and discharge instructions. If the image carries meaning that did not survive extraction, automation should pause for a corrected transcription or an attorney-approved notation of the limitation.
Dates or document boundaries are unstable
A chronology depends on reliable encounter dates and reliable document grouping. Escalate when dates are partly cut off, handwritten, overwritten, inconsistent between header and body, or likely to be confused with print or scan dates. The workflow should also stop when one file contains interleaved records from different providers or when a separator page has been dropped.
This is especially important in treatment-gap analysis. An apparent six-week gap may be a true break in care, a missing record, or an encounter whose date was not captured. The draft should not choose among those explanations on its own.
The unreadable material touches a disputed issue
Not every blurred administrative page deserves the same attention. A partially unreadable cover sheet may be low risk. A partially unreadable note addressing prior symptoms, mechanism of injury, work restrictions, referral compliance, or future care is different. If the page could affect a point an adjuster is likely to test, it belongs in a human-review queue.
For a California claim, this also connects to evidence preservation and attorney supervision. Under California Rule of Professional Conduct 1.1, competence includes the learning and skill reasonably necessary for the matter. Using automation does not transfer the attorney’s responsibility to understand the source evidence that supports the firm’s position. Rule 1.6 confidentiality duties likewise remain relevant when medical records are routed through any outside technology provider.
Confidence signals conflict with practical reality
A numerical confidence score is useful only if the firm knows what it measures. Page-level OCR confidence may not reveal that the one unreadable phrase is the dispositive phrase. A high average can also conceal a few critical low-quality pages in a large production.
Firms should treat confidence as a triage signal, not a legal conclusion. If the extracted chronology conflicts with an intake note, billing sequence, diagnostic report, or attorney understanding of the file, the inconsistency should trigger source review even when the software reports a strong overall score.
A practical human-escalation protocol for PI firms
The most useful policy is short enough for intake staff, case managers, and attorneys to apply consistently. Build the following checkpoints into the records-to-demand workflow:
- Preserve the original file. Keep the source image and its original filename or production identifier. Never let cleaned OCR text replace the record from which it was derived.
- Run a page-quality screen before extraction. Flag blank pages, low contrast, skew, clipping, handwriting, table-heavy forms, mixed orientations, and pages with no usable text layer.
- Separate unreadable from absent. Use an explicit status such as “source requires review.” Do not translate uncertainty into a definitive statement that the record contains no reference to the issue.
- Route material pages to a person. Assign the review to someone who understands the file and the demand theory. A technical operator can identify poor OCR; legal staff must decide whether the uncertainty matters.
- Resolve the defect at the source when possible. Request a cleaner production, rescan the page, correct orientation, or obtain the native PDF. Manual transcription should record that it was transcribed from a difficult source and should be checked against the image.
- Document the resolution. Record which pages were reviewed, what was corrected, and who approved the correction. This creates a usable audit trail when the demand is revised later.
- Re-run downstream work after correction. A repaired chronology should feed the demand draft again. Patching one paragraph manually while leaving the stale extraction elsewhere invites inconsistency.
Firms should also sample apparently clean pages. If a batch contains repeated OCR defects, expand the review instead of assuming the problem is isolated. For a two-hundred-page record set, a sensible workflow may prioritize every page tied to diagnosis, causation, restrictions, referrals, and future treatment, then sample the remaining administrative material. The exact scope should follow the case, not a universal percentage.
This protocol complements a structured intake process. Our related guide on structured medical record intake explains why file organization and document boundaries have to be correct before automation can add much value.
How Legal Power AI fits
Legal Power AI is designed to help plaintiff firms move from organized case materials toward attorney-reviewed demand work, while keeping the lawyer in control of factual and strategic decisions. The Chronology Builder supports the records-review stage, but no AI-generated chronology or demand should be sent without checking it against the underlying file. Work-product protections, confidentiality obligations, and the attorney’s duty to verify accuracy still apply to AI-assisted drafts.
Make “stop” an intentional workflow state
The mature use of legal AI is not uninterrupted automation. It is knowing where automation is dependable, where uncertainty must remain visible, and where a person has to take over. Low-quality scans and handwritten notes are predictable conditions, so firms should not handle them through ad hoc judgment after a draft has already been produced.
A defined escalation rule protects more than accuracy. It keeps the chronology, demand theory, and supporting records aligned; gives staff a repeatable way to handle difficult productions; and prevents polished language from outrunning the evidence. The best workflow is not the one that processes every page without interruption. It is the one that stops before uncertain source material becomes an unsupported factual statement.
See a more disciplined records-to-demand workflow.
Legal Power AI helps plaintiff PI teams turn organized case materials into attorney-reviewed demand work without treating automation as a substitute for legal judgment.