OCR Claim Files: Making a Carrier's File Searchable

To OCR claim files, keep the carrier's PDF untouched, page through a copy to mark where each letter, report and note begins, cut it into parts your tool accepts, and run recognition on each part. The result is a file you can search for a date, an adjuster or a policy term, with every hit traceable to a page.

This post is for anyone reading a carrier's file from the outside or the inside: policyholder counsel in a coverage dispute, coverage counsel advising the insurer, or a paralegal handed 900 pages and asked when the denial was decided. The running example is invented. Marguerite Ostrander's kitchen flooded, Lakeshore Casualty denied most of the claim, and her lawyer has received the claim file in discovery.

Denial letters are the pages to search before the rest of the log. Look for the first-party claim number, the policy limit, the deductible, and each exclusion the letter names. When a later memo mentions subrogation or an endorsement, open that page and read it beside the image before you quote it. The denial and the policy form should show the same words.

What is inside an insurance claim file?

An insurance claim file is the carrier's whole record of one claim: the notice of loss, the adjuster's running notes, letters in and out, the policy, investigation material, estimates and valuations, payment records, and internal reviews. Each of those sections has its own page design, so one file mixes forms, logs, letters, tables and photographs.

Regulation shapes the contents in some places. California's rule on file and record documentation, 10 CCR 2695.3, requires an insurer's claim files to hold all documents, notes and work papers, including copies of all correspondence, in enough detail that pertinent events and their dates can be reconstructed. Other states set their own record-keeping rules, so read the one that governs your carrier. The practical point for a reader is the same everywhere: the file is meant to be a dated account of what the carrier did.

The sections below are the ones you will usually meet, with what each asks of recognition software.

The claim notes are usually the single most useful section, because they narrate everything else. They get a post of their own on converting adjuster notes into a sortable table.

A file received by assigned defense counsel is a different thing from one produced in litigation. It arrives with an assignment, it is not redacted, and it comes alongside suit papers and a policy. The post on insurance defense OCR for an assigned file covers that package.

Why does a claim file arrive as one enormous PDF?

A claim file arrives as one enormous PDF because the carrier's claim system stores each item as a separate record and an export prints them all to a single file. The export turns every note, letter and photograph into page images laid end to end.

Several habits of these exports explain what you see when you open the file.

• Everything is an image. Even letters that began as typed text come out as pictures of pages. Nothing can be selected or searched until recognition is run.
• The order is the system's, not yours. Some exports run newest first. Others group by document category, so all letters precede all photographs. A reply can sit two hundred pages ahead of the letter it answers.
• Attachments are detached. An email prints on one page and its attachment appears as a separate item somewhere else, often with nothing to connect them except a file name in the email header.
• Duplicates multiply. A letter is stored when drafted, again when sent, and again each time it is forwarded. Email chains reprint every earlier message.
• Color pages inflate the size. A hundred pages of photographs can weigh more than the remaining eight hundred pages together.
• Stamps arrive last. In a produced file, Bates numbers and confidentiality legends are stamped onto every page after export, sometimes over the bottom line of text.

None of this is a defect to complain about. It is how the file was made, and knowing it tells you what to do first: establish the order, find the boundaries and identify the duplicates before you read for content.

One request is worth making early. Ask whether the carrier can produce the file with each document as its own PDF, or with a list of documents by page range. If the answer is no, you will build that list yourself, and the following section shows how.

Marking where each document starts inside a claim file

Mark each document's first page in a list before you cut anything: one row per document, with its first and last page in the carrier's file. That list tells you where the file can be divided safely, and it becomes the index you search and cite from for the remainder of the case.

Building the list is a page-turning job, and it goes faster with a rule for what counts as a new document.

1. Work on a copy. Save the carrier's PDF as received and never alter it. Page numbers in that file, or its Bates numbers, are the references everything else depends on.
2. Page through at thumbnail size. Boundaries show up as changes in appearance: a new letterhead, a switch from text to photographs, a form with boxes, a page that is mostly a table.
3. Log a row at each change. Record the first page, the last page, a short type and a date if one is visible.
4. Group rows into parts. Combine neighboring documents into parts that fit your tool's limits. Schema Law accepts files up to 20 MB, with up to 250 pages per document on the Solo plan and 1,000 on the Team plan, as listed on the pricing page.
5. Cut the copy at the row boundaries. Put the page range in each part's name.
6. Upload each part and export. Keep the searchable PDF of every part, and export tables to Excel where the section is mostly figures.

For the Ostrander file, the first rows of the list look like this.

Photographs get a part to themselves. They are heavy, they hold almost no text, and isolating them keeps the text-rich parts small. The general rules for cutting large files are set out in the post on batch OCR, under how a batch should be split before upload.

Linking attachments back to their emails

Once the parts are searchable, attachments can be reunited with their cover emails. Search for the attachment's file name as printed in the email header, or for a distinctive phrase from the attachment's first line. Add a "Sent with" column to the list and record the page of the email there. In the example, the plumber's report at LC-0172 turns out to have reached the adjuster by an email at LC-0407, eleven days after the date on the report. That gap is a fact, and only the list makes it visible.

Photographs, repair estimates and recorded statements in a claim file

Photographs, repair estimates and recorded statements each need different handling, because only part of each is text. Photographs yield captions at most. Estimates are tables that belong in a spreadsheet. A recorded statement appears in the file as a transcript, a summary or a bare reference to an audio file, and OCR reads only what is printed.

Photographs

A photo page from a claim system typically prints two or four images with a caption under each: a number, a date and time, the photographer's name and a line of description. Recognition can read the captions. It does not describe what a picture shows. Text that happens to appear inside a photograph, such as a serial plate on a water heater, is sometimes picked up and often garbled, so read those from the image.

Captions are still worth converting. A search for "supply line" across the photo part finds every image the adjuster labeled that way, which is quicker than scrolling through 104 pictures.

Repair estimates

An estimate printed from estimating software is a long table: room by room, line by line, with quantity, unit, unit price, tax and a total, then subtotals per room and a summary page with overhead, profit and depreciation. Export these pages to Excel and check the arithmetic before you rely on a figure.

• Add the line totals for one room and compare the sum with the printed room subtotal.
• Add the room subtotals and compare with the summary page.
• Look at any row where a description wrapped onto a second line, since wrapped text can push numbers into the wrong column.

The post on converting a scanned PDF to Excel explains how tables are found on a page and why columns sometimes merge. When two estimates exist, the carrier's and the contractor's, converting both to spreadsheets lets you line them up room by room.

Recorded statements

Look at what the file actually holds. A full transcript is typed, with speaker labels, and converts like any other transcript. A summary written by the adjuster is the adjuster's account of what was said, and should be labeled that way in your notes. A claim note saying "RS taken, uploaded" tells you audio exists somewhere outside the PDF. OCR cannot turn audio into text. If the recording matters, request the audio file itself.

Redacted reserves and withheld pages in a produced claim file

A produced claim file often has reserve figures blacked out and whole pages withheld as privileged. OCR reads only what is visible, so a redacted passage yields nothing or a few stray characters, and a withheld page appears as a slip sheet or a gap in the numbering.

Whether reserves and communications with counsel must be produced is contested, and the answer depends on the jurisdiction and the type of claim. This section covers what the redactions do to your converted text, not whether they were proper.

Black boxes. Recognition may skip a solid box, or it may read the box edge as a run of capital I's, underscores or block characters. Where a box cuts through a sentence, the words on either side join up in the text, so a search hit can read as a complete sentence when it is not. Open the image for anything you plan to quote.

Redaction labels. Many productions print a word such as "Redacted" or "Privileged" inside the box. Those labels are text, and they are useful. Search for them to list every redacted page in minutes, then note the page and the stated reason in a spreadsheet.

Slip sheets and gaps. A page reading "Document withheld" with a Bates number is a placeholder for something removed. Sometimes there is no placeholder, only a jump in the numbering. Converted Bates numbers let you check the sequence, though stamped numbers are themselves prone to misreads, so confirm any apparent gap against the page images.

Leftover text. If the carrier's PDF already has a text layer when you receive it, a box drawn over the image may not have removed the words beneath. That is a problem for the producing party and a delicate situation for the recipient, governed by your own rules on inadvertently disclosed material. The post on hidden text in a PDF explains the mechanism.

Compare your list of redactions and gaps with the privilege log. In federal court, Rule 26(b)(5)(A) requires a party that withholds otherwise discoverable information as privileged to expressly make the claim and to describe what was withheld in a manner that lets other parties assess it. State rules differ. A redaction on LC-0033 with no matching log entry is a question for a meet-and-confer letter.

Reserve history has one more hiding place. Even when the dollar figures are blacked out, claim notes often record that a reserve was changed and when. Searching the notes for "reserve" gives you the dates of each change, which may matter on their own.

Which searches find the coverage decision in a claim file?

Start with the words a carrier uses when it decides: "deny", "denial", "decline", "reservation of rights", "exclusion", "coverage opinion", "authority" and the title of the policy provision relied on. Then search the notes dated before the decision letter, since the letter only announces a decision that was made earlier.

Work from the letter backwards.

1. Find the decision letter. Search for "we must respectfully" or "unable to extend coverage" or simply "denied". Note its date and the provisions it quotes.
2. Search the provision. Take a distinctive phrase from the quoted exclusion, such as "continuous or repeated seepage", and search the whole file. The first appearance tells you when that theory entered the claim.
3. Search the vocabulary of approval. "Authority", "approved", "per supervisor", "committee" and "roundtable" find the points where someone other than the adjuster weighed in.
4. Search the people. List every name in the notes and search each one. A name that appears only twice may be the manager who approved the denial.
5. Search the gaps. Sort your list by date and look for weeks with no entries. Silence before a decision letter is worth a question at deposition.

Carriers abbreviate heavily, so a search for the full word can miss the entry that matters. Build a short list of the abbreviations this file uses as you read, and search both forms. Recognition errors have the same effect: "denial" read as "denia1" will not match. If a search for an obvious term finds nothing, try a neighboring word before you conclude the file is silent.

In the Ostrander file, "seepage" first appears in a note on 24 February, four days after the plumber's report is dated and a week before the report reached the adjuster by email. The decision letter is dated 2 April. Three searches and the document list produced that timeline, each point tied to a Bates page.

A converted claim file is a finding aid. Every quotation and every date that goes into a brief should be read from the page image, and the citation should be to the carrier's Bates number, never to a page of your export. Before you upload a file that is subject to a protective order, read the order; the guide to OCR for documents under a protective order covers what to look for. Schema Law's stated commitments on how uploaded documents are handled are a starting point for that comparison.