By Sagar Shankaran, Founder of CallSphere
Home medical equipment intake keys 22 fields off a grainy fax. 2026 readers pull the face-to-face date, NPI and AHI straight off the scan and flag the rest.
Key takeaways
It is 4:50 on a Thursday in February, which in this trade means the fax server is still working. Fourteen pages arrive from a sleep lab: a Standard Written Order signed with a scrawl, six pages of chart notes, a home sleep test summary with the AHI inside a table, a copy of an insurance card shot at an angle, and a cover sheet with a line handwritten in the margin: "pt prefers nasal pillows, do not call before 6."
Your intake coordinator opens Brightree and starts typing. Name, date of birth, address, primary and secondary payer, policy number, ordering physician, that physician's NPI, the face-to-face date, the AHI, the code for E0601 and the mask family, the referral source. Somewhere between twenty and twenty-five fields, read off a grainy fax at the end of a long day.
Every home medical equipment company has this hour. The referral does not arrive as data. It arrives as a picture of a piece of paper that was itself a picture of a piece of paper, and everything downstream — the authorization request, the delivery ticket, the claim, the audit response two years from now — is built on whatever your coordinator managed to read at 4:50 p.m.
The rest of the paper behaves the same way. The delivery technician comes back at 6 p.m. with a pile of proof-of-delivery tickets, half signed by a daughter with her initials next to a scratched-out date. A caregiver texts a phone photo of a cracked joystick on a K0823 power chair so you can decide whether the service call is warranty or billable. A physician's office faxes a chart note with the face-to-face date circled in pen, because that was faster than reprinting it.
Here is the plain definition worth keeping: multimodal document reading means the computer looks at the actual fax, photo or scan the way a person does — the crooked insurance card, the circled date, the initials in the margin — and pulls the fields out of it, instead of waiting for someone to retype them.
Nobody gets denied for "bad typing." They get denied for a wrong NPI, a face-to-face date outside the window, a date of birth off by a digit against the payer's file, a transposed policy number, a quantity that does not match the order. Those are typing. They come back as a clearinghouse rejection if you are lucky and a documentation denial from your DME MAC if you are not, and a billing specialist spends half an hour reconstructing what the fax actually said.
The workaround everyone pretends is fine is double-entry with a spot check. Somebody keys it, somebody else eyeballs the high-dollar orders — the power mobility file, the enteral pump, anything going to a prior authorization vendor like eviCore or Carelon — and the small stuff ships on trust. That is why the same three fields keep appearing on your denial report every month.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for healthcare in your browser — 60 seconds, no signup.
It gets worse in exactly the season when you can least afford it. October through March, respiratory referrals climb, oxygen and nebulizer setups climb, and the intake queue that ran clean in July is forty packets deep by 3 p.m.
Document readers are not new. The 2024 version needed a clean, upright, machine-printed page and a template telling it where to look. Hand it a fax rotated two degrees, or a form where the physician's office wrote the visit date in the margin instead of the box, and it either skipped the field or returned garbage with total confidence.
By 2026 that changed in a way that matters for this trade specifically. The current models read the artifact itself — the photo, the fax, the scan, the handwriting — rather than requiring a clean digital original. One can look at page nine of a fourteen-page packet, understand that the thing circled in pen next to "OV" is the face-to-face date, and put it in the right field. It reads the packet the way your best intake coordinator reads it, and it does not get tired at 4:50.
flowchart TD
A["Sleep lab faxes 14-page packet, 4:50 p.m."] --> B["Reader opens every page as an image"]
B --> C["Pulls NPI, face-to-face date, AHI, codes, policy number"]
C --> D{"Every claim-critical field found and legible?"}
D -->|No| E["Order held with the exact page flagged"]
D -->|Yes| F["Draft patient and order posted in Brightree"]
E --> G["Intake coordinator fixes only the flagged field"]
G --> F
F --> H["Authorization specialist works the order, not the keyboard"]
The packet lands in the fax queue at 4:50. By 4:51 a draft patient and draft order sit in Brightree with the demographics, the ordering physician and NPI, the face-to-face date, the sleep study result, the product codes and the referral source populated — each field noting which page it came from, so your coordinator clicks straight to page nine instead of scrolling fourteen pages.
Two fields come back flagged: the secondary policy number is half-cut by the fax margin, and the signature date is ambiguous because the pen slipped. Your coordinator opens that short review list at 8:10 Tuesday morning, looks at two cropped images beside the fields, fixes both in under a minute, and releases the order to the authorization specialist.
The same thing happens to the delivery tickets. Monday night's pile gets read overnight: signature present, date present, patient name matching, the serial number written on the line for the concentrator. Three come back flagged — one unsigned, one with an unreadable serial, one dated 2/3 in a way that could be either. Those go to the delivery supervisor. The other twenty-nine are filed against the right orders, so when an audit letter arrives in fourteen months the proof of delivery is already attached to the claim instead of sitting in a banker's box.
Illustrative assumptions, not a claim about your company. Put your own numbers in the same shape.
| New referrals keyed per month | 300 |
| Fields keyed per referral | 22 |
| Fields keyed per month | 6,600 |
| Keying error rate (illustrative) | 1.0% |
| Wrong fields per month | 66 |
| Share landing on a claim-critical field | 1 in 4 |
| Denials or rejections caused by keying | about 17 per month |
| Rework at 35 minutes each, $26 per hour loaded | about $258 per month |
| Never recovered (1 in 5), average episode value $340 | about $1,156 per month |
| Keying time at 6 minutes per referral | 30 hours per month |
Call it roughly $1,400 a month in avoidable money and most of a part-time position in avoidable keystrokes, on a 300-referral operation. Note also the piece that never shows up in a spreadsheet: a denial found in week six delays first billing on that setup by two to three weeks, which is real cash timing on a rental where you already fronted the equipment.
Still reading? Stop comparing — try CallSphere live.
See the healthcare AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
This is where honesty matters more than enthusiasm. Reading a document well is not the same as deciding what it means, and DME has fields where a confident wrong answer is expensive.
Keep a human on the qualifying clinical value — the AHI, the arterial blood gas or oximetry result behind an oxygen setup, the wound measurements behind negative pressure. Not because the reader gets the number wrong, but because a person has to decide whether the number and the date together actually qualify the patient under the policy, and that is a judgement, not a transcription. Keep a human on the signature and date for anything requiring a written order prior to delivery, because a mis-read date there is the difference between a clean claim and a refund request. And keep a human on anything ambiguous — the point of the flag-and-review list is that the system may say "I am not sure."
One more limit worth saying out loud: reading a packet does not chase a missing chart note. If the physician's office never sent the face-to-face documentation, no reader can invent it, and somebody still has to call that office.
No. The reader sits in front of your system, not instead of it: the fax comes in, the fields get pulled, and a draft order is created in whatever you already run — Brightree, Bonafide, TeamDME!, NikoHealth, CareTend. If a vendor says the only way to read your faxes is to move your whole operation, that is a sales strategy, not a technical requirement.
Treat it like every other vendor that touches protected health information: you need a business associate agreement, you need to know where the documents are processed and stored, and you need to know how long they are kept. Ask for those three in writing before the trial. Any serious vendor answers all three without flinching.
Do not chase a percentage. Run it alongside work your coordinator is already doing for two weeks — same packets, both results — and compare field by field. You want two things: does it beat your current error rate on the claim-critical fields, and does it flag what it is unsure about instead of guessing. The second matters more.
Yes, and it is an easier win because the stakes are lower. A caregiver photographs a broken caster or a torn sling, the reader names the part and pulls the equipment record for that serial number, and your service coordinator gets a repair ticket with the part identified instead of a text thread.
Do not start with the whole fax queue. Pick your highest-volume referral source — usually one sleep lab or one hospital discharge department — because their packets arrive in the same format every time. Run those through reading for two weeks alongside normal intake, tally the fields the reader got right, wrong, and correctly flagged, then decide whether to widen it.
One last note on the other half of intake. Paper is only one way referrals arrive; the rest come by phone, and those calls do not wait for a callback. CallSphere builds AI voice and chat agents that answer the phone and web chat around the clock, take the caller's details, book the delivery or fitting window and hand a complete lead to your intake queue — the same goal as the fax side of that desk: nothing sits unread until 8 a.m.
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.
Carbon-copy scale tickets, margin notes and container photos still get keyed by hand. What 2026 document readers catch, and the unbilled extras they surface.
Broadcast reconciliation eats 26 hours a month at a mid-market agency. How 2026 document readers handle faxed affidavits, margin notes and makegoods.
Handwritten shorts on back-door delivery tickets leak thousands a year across a 14-store group. What 2026 document reading changes, with the arithmetic.
Heat numbers get retyped four times between the mill and your AS9102 Form 2. What 2026 document reading changes at an aerospace supplier's receiving dock.
Colo cross-connect orders still arrive as faxed scans with handwritten port pairs. Here is what 2026 document reading does to the provisioning desk and the MMR.
Twenty-nine bad numbers a month enter a brewery's records through hand keying. How 2026 document readers handle pencil, faxes and hop certificates of analysis.
© 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