Jump to section
A file lands on your desk as a stack of scans: an affidavit fax that went through two machines, a client intake form filled out in ballpoint, a registry stamp that half-printed. Somewhere in there is the applicant's name, a limitation date, and a claim amount, and someone has to key it in. We sat down with a litigation office manager who spent a year turning that pile into structured data, on condition we keep her firm's name out of it.
Why Scanned Documents Still Dominate Legal Work
Q: We keep hearing everything is digital now. Why are you still drowning in scans?
Because the profession runs on paper the way it always has, it just photographs the paper first. A client shows up with a shoebox. Opposing counsel serves an affidavit as a flat PDF with no text layer. It looks like text on your screen, but to a computer it is a photo of text, a completely different thing.
Q: So the PDF format itself is the problem?
Not exactly. A PDF can carry a real text layer, and if it does you can copy and search it and life is good. The trouble is the scanned ones. When a document has been printed, signed, and scanned, all that survives is the image. That is where optical character recognition comes in. OCR looks at the picture and guesses which shapes are which letters. Emphasis on guesses.
What OCR Reliably Reads and What Trips It Up
Q: After thousands of documents, where does OCR actually earn its keep?
Clean typed text on a good scan is close to solved. A pleading printed black on white in a standard font comes through beautifully. Body paragraphs, statute citations, most headings. I trust that at a glance.
Q: And where does it fall apart?
Handwriting first. Anything filled in by hand is a coin toss. Then it is the things that hurt most in law. Numbers. It loves to read a 1 as a 7, an 8 as a 3, a lowercase l as the digit 1. That matters when the number is a claim amount or a file number, and dates get scrambled the same way.
Warn. The most dangerous OCR error we saw was a limitation date read one digit off, a 2023 that came through as 2028. Nobody notices a plausible wrong date the way they notice garbled words. Treat every extracted date as unverified until a human confirms it against the image.
Q: Anything else that consistently trips it?
Tables and columns. A bill of costs in columns can come out interleaved, so a figure from one row lands next to a label from another. Stamps overlapping printed text confuse it. Faxes are the worst. And accented characters in French documents or in a client's name can silently vanish, which is its own headache in a bilingual practice.
How You Verify Extracted Fields Before Trusting Them
Q: You said treat everything as unverified. What does verification actually look like on a busy day?
We built a habit around confidence and consequence. OCR tools give you a confidence score per field, and low confidence gets flagged automatically. But the tool can be very confident and very wrong, so we sort by consequence too. A field that decides a deadline, a party's legal name, a dollar figure, or a court file number gets a human eye whatever the score says.
Q: How do your staff check a field without re-reading the whole document?
Side by side. The extracted value on one side, the cropped image of where it came from on the other. You are not proofreading the document, you are confirming that this string matches that patch of image. A few seconds per field, and that is the whole game.
We stopped asking whether the OCR was right. We started asking which fields would hurt us if it was wrong. Litigation office manager
Turning Text Into Structured Matter Data
Q: Reading the text is one thing. How do you get it into fields you can actually use?
That is the part people underestimate. OCR gives you a wall of text. Turning that into applicant name, respondent name, style of cause, filing date, and claim amount is a separate step called extraction, and it is where the value lives. We standardized on a small set of canonical fields, then mapped everything into them, same names whether the data came from an intake form or a court document. Here is the tiering.
| Field type | OCR reliability | Verification rule |
|---|---|---|
| Typed body text | High | Spot check only |
| Party names and style of cause | Medium | Always confirm against image |
| Dates and limitation periods | Medium to low | Always confirm, flag if computed |
| Dollar amounts and file numbers | Low | Always confirm, double key if critical |
| Handwritten entries | Very low | Assume wrong, key by hand |
Once the fields are clean, they flow into the matter. Names populate the style of cause, the filing date drives the calendar, disbursement figures land where billing can see them. We use the Data Import and Custom Forms pieces of A1 CMS to catch the mapped fields, but the discipline matters more than the tool.
Tip. Keep the original scan attached to the matter, not just the extracted text. When someone questions a date six months later, you want the source image one click away.
The Mistakes You Made Early On
Q: What would you do differently if you started over?
Three things. First, I trusted confidence scores too much and consequence too little, and it took a near miss on a deadline to fix that. Second, I ran everything through the same pipeline, when handwritten intake forms need a different path than typed pleadings. Third, the big one, we did not save the source snippet at first, so verification meant reopening the whole document. Once we stored the source patch beside each value, checking got fast enough that people actually did it.
Q: Any advice on the confidentiality side?
Know where the images go. If it is a cloud service, that is client material leaving your walls, and the usual duties apply. Our notes on cloud storage and client confidentiality cover the questions worth asking, and the broader reality check on legal AI is a good read before you buy anything.
The honest takeaway is that OCR reads typed text accurately but struggles with numbers, handwriting, and tables, and it never signals when it is guessing. If you verify against the source image, scaled to the consequence of each field, it will process thousands of documents that would otherwise require manual keying. If you treat it as infallible, it will eventually record a wrong limitation date on a live file, and that is not a mistake you get to undo. The technology is ready. The discipline around it is the work.
