By Sagar Shankaran, Founder of CallSphere
Faxed records, handwritten margin notes and phone photos still get keyed in by hand at litigation firms. What that error rate costs, and what 2026 readers fix.
Key takeaways
Who in your firm knows exactly what came back from the records custodian at Mid-State Orthopedic last Thursday? Not roughly. Exactly — which date range, how many pages, whether the certification was signed, and whether the production actually covers the eleven months your client says he was treating there.
In most litigation firms the honest answer is: the records clerk knew for about ninety seconds while she was typing it in, and after that it lives in whatever she typed. Which is why the ortho notes stopping in March gets discovered at mediation, in front of the mediator, by opposing counsel.
Walk the stack. Four hundred and twelve pages faxed from a hospital release-of-information department, arriving as a PDF where every third page is skewed nine degrees and the bottom line is cut off. A cover sheet with a handwritten note in the margin: "partial — 2019 archived, request separately." An affidavit of custodian of records, signed and notarized, that has to be kept because it is what makes the records admissible under the business records exception. A "no records found" affidavit from a chiropractor your client swore he saw twice. A UB-04 itemized statement where the charges you need are in the right-hand column of a form designed for a billing clerk, not a lawyer. A disc of imaging that nobody will open. Two pages your client photographed on his phone — a discharge instruction sheet, taken at an angle, with a thumb in the corner.
The records clerk opens each one, reads far enough to know what it is, and types four or five fields into Filevine, Litify, Assembly Neos, or whatever your firm runs: provider name, date range produced, page count, whether a certification is attached, and any note. Then she names the file and moves on. She does maybe 240 of these a month.
What changed in 2026 is that a model can read the artifact as it arrived — the crooked fax, the handwritten margin note, the phone photo — instead of requiring a person to look at it and retype what it says.
Optical character recognition has existed forever, and every firm has tried it. It reads clean typed text at 300 dpi. It falls apart on the things litigation actually receives: a fax of a fax, a stamp overlapping a date, a physician's handwriting, a form where the meaning depends on which box has a check mark in it, a photograph taken in a parking lot.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for law firm in your browser — 60 seconds, no signup.
The 2026 readers do not transcribe character by character. They look at the whole page the way a person does — the layout, the letterhead, the stamp, the handwriting in the margin, the check box — and answer questions about it. "Which provider sent this?" "What date range does it cover?" "Is a custodian certification included and is it signed?" "Does anything on this page say the production is partial?" That last one is the question the fax cover sheet answers in ballpoint pen, and it is the one that has been getting missed.
flowchart TD
A["Fax from the records custodian"] --> D["Reader opens the artifact as it arrived"]
B["Certified mail with handwritten cover sheet"] --> D
C["Client's phone photo of discharge paper"] --> D
D --> E["Pulls provider, date range, page count, certification"]
E --> F{"Does the range cover the treatment dates on file?"}
F -->|Yes| G["Files to the matter and closes the records task"]
F -->|No| H["Opens a follow-up subpoena task for the records clerk"]
Everything that arrives — the fax server, the scanner in the mail room, the shared inbox the copy service sends to, the photos clients text in — lands in one place. Each item gets read once and turned into the same five fields the clerk was typing, plus two she was not: the earliest and latest treatment dates actually appearing inside the records, and a plain-English note of anything on any page that suggests the production is incomplete.
Then the useful part. Those dates get compared against what the file already says. Your client's intake sheet says he treated at Mid-State Orthopedic from February through December. The production covers February 3 to March 19. That mismatch opens a task — "supplemental request to Mid-State, 3/20 forward" — on the records clerk's list on Thursday afternoon, not at mediation in November.
The same comparison catches the other classic: a provider referenced inside somebody else's chart who is not on your request list at all. The ER note says "patient referred to Dr. Aponte for pain management." Is there a Dr. Aponte in your provider list? No. Now there is a task.
The clerk still opens the certifications herself and still eyeballs anything the reader marked as low confidence. Her day changes from typing to deciding.
Manual data entry into a case management system is not error-free and never has been. Published error rates for hand-keyed fields cluster around 1% per field, and it is worse when the source is a bad fax read at speed at 4:40 on a Thursday. Take a firm handling 240 record responses a month with five keyed fields each.
| Assumption | Keyed by hand | Read from the document |
|---|---|---|
| Record responses logged per month | 240 | 240 |
| Fields per response | 5 | 5 |
| Error rate per field (illustrative) | 1.0% | 0.3% after clerk review |
| Wrong fields per month | 12 | 3.6 |
| Partial productions never flagged (illustrative, 8% of responses) | 19 per month | near zero |
| Clerk minutes per response | 7 | 2.5 |
| Clerk hours per month | 28 | 10 |
Eighteen hours a month of a $26-an-hour clerk is $468 — real, but not why you would do this. The reason is the 19 partial productions a month that currently go unflagged. Most are harmless. A few are not. One of them is the file where you find out during the deposition that eight months of treatment was never produced, and now you are asking for a continuance, paying for a second deposition, or arguing about a discovery cutoff you cannot move. Price one of those at the cost of a continuance and a redo — for most firms, several thousand dollars and a credibility hit with the judge — and the case makes itself on a single avoided incident per year.
Certifications and affidavits. Whether a custodian affidavit is properly signed, notarized, and in the form your state requires is an admissibility question. Have the reader tell you a certification appears to be present; have a person confirm it is the right one. This is the page that gets you a records objection sustained at trial.
Still reading? Stop comparing — try CallSphere live.
See the law firm AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
Anything the reader marks as unsure. Handwriting reading has improved a great deal and is still not reliable on a provider's cramped addendum or a hand-marked date. If it cannot read the date range cleanly, it should say so and route the page, not guess. Insist on that behavior before you turn it on.
Protected health information handling. These are medical records obtained under HIPAA authorizations and subpoenas. Whatever you use must be a business account with training on your data turned off and a retention setting you chose, and your file-handling policy should say so in writing. If your malpractice carrier asks how client medical records are processed, you want a one-page answer ready.
The judgment call about what to do next. The system can tell you the ortho records stop on March 19. Whether to send a supplemental request, subpoena the archive at a copy cost of $340, or let it go because the treatment after March was unrelated — that is a lawyer's call about damages, not a clerk's and not a machine's.
Usually, and it is worth testing with your ten worst faxes rather than your ten cleanest. Multi-generation faxes of handwritten notes are the hard case. The realistic outcome is that 85 to 95% of a typical stack is read cleanly and the remainder is routed to a person, which is still a very different Thursday than the one you have now.
It can pull the totals and the line items, and that is genuinely useful for building a specials table. Do not let those figures go into a demand letter without a person tying the total back to the face of the bill. Damages numbers are the last thing you want to be approximately right about, and adjusters check them.
Imaging discs are a different problem and this does not solve them — you still need a radiologist or a treating physician to say what the films show. Portal downloads are typically clean PDFs and are the easiest input of all. Have someone map which of your top twenty providers send which format; that list is worth an hour of somebody's time.
Firms that try usually regret it. The records clerk who has been doing this for six years knows that a certain hospital always sends billing separately from records, and that a particular copy service quietly drops radiology. That knowledge is what makes the follow-up tasks correct. Use the time for chasing the gaps, which is the part of the job that has always been shortchanged.
Pick one active case with a discovery cutoff inside ninety days. Take every record response in that file — all of them, including the ugly faxes and the "no records" affidavits — and have them read and compared against your client's treatment history. You are looking for one thing: a gap nobody had noticed. If you find one, you have your answer about whether to do this across the docket, and you have saved a case in the process.
Records requests generate phone calls — release-of-information departments calling back about an authorization date, clients calling to ask whether the firm got their bills, copy services confirming an address. CallSphere builds AI voice and chat agents that answer the firm's line and web chat around the clock, capture who is calling and about which matter, and put the call-back on the right person's calendar. It does not read your medical records — it keeps the records clerk on the stack instead of on the phone.

Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Block billing, clerical time and wrong task codes cost defense firms 6% of every invoice. Reading all 4,200 entries first now costs a few dollars a month.
The 8:52 p.m. accident call does not leave a message. What an instant-answer intake line changes about a plaintiff firm's signed-case rate, with the arithmetic.
Why the medical chronology is the last serial job in a plaintiff firm, and what happens to your demand calendar when four agents split the chart by provider.
ERISA reimbursement, Medicare conditional payments and made-whole: how grounding keeps a confidently wrong lien answer out of your disbursement statement.
AI that answers only from a firm's approved memos and the client's file, showing the page — so a wrong travel answer is caught before it costs a decade.
A 1,200-page record in one question: timeline, inconsistencies, exhibit mapping. What it saves on a flat fee, and what still needs an attorney's judgement.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI