Intake
A referral almost never arrives as data; it arrives as a fax. What turning it into a record by hand actually costs, where the errors land, and what scan-to-fill changes and what it does not.
A referral for home health almost never arrives as data. It arrives as a fax, or a PDF of a fax, or a photo of a printout someone left on a desk: a physician's order, a face sheet, sometimes a discharge summary stapled to both. Somebody has to turn that page into a record before a single visit can be scheduled, and that step is one of the most expensive things an agency does that nobody puts a number on.
Read the whole page, because the field you need is rarely where the form puts it. Find the name, the address, the diagnosis, the ordering physician, the visit frequency, the start-of-care date, the insurance. Type each one into the intake system by hand. Notice the fax cut off the right margin and call the referring office back. Do this again for the next fax in the pile.
None of that is clinical work. It is transcription, done by clinical staff, under time pressure, from a document that was never designed to be read by a computer or, often, by a person in a hurry.
| What happens | What it costs |
|---|---|
| A field is mistyped: a transposed digit in the phone number, a street misread from a bad scan | A clinician drives to the wrong address, or a call that should connect does not. |
| The frequency is read wrong: "2x/wk x 4wks" skimmed as weekly | A visit count that does not match the plan of care, caught late or not at all. |
| The fax sits in a queue until someone has a free half hour | A start-of-care date slips, which in home health can affect reimbursement timing. |
| Two referrals for patients with the same or similar name | The wrong chart gets opened. This is the failure that matters most, and it is exactly the one busy manual entry makes easiest. |
| An illegible field is guessed rather than left blank | A wrong value that looks confident is worse than a gap that visibly needs a phone call. |
A referral that takes fifteen minutes to key in is fifteen minutes before the clinician nearest that address even knows the patient exists. Multiply that by a caseload that turns over every week and the delay compounds: the day is planned around what is already entered, so a slow intake step is a quiet tax on how soon a new patient gets seen.
Each field carries a confidence score and is traceable back to where it appeared on the page, so a clinician can see at a glance which values to double-check and which were plainly stated.
If the model cannot find a start-of-care date on the referral, the field is left empty rather than filled with something plausible. A blank asks a question; a wrong-looking-right answer does not.
Scan-to-fill produces a draft, not a client record. The same create step the typed form uses is the only way a draft becomes a real patient, so a faxed referral gets exactly the review a hand-typed one would.
If a name on a referral is close to an existing patient's, that is flagged as a question for a person, never resolved silently in either direction. Acting on the wrong patient is the failure everything else in this list exists to prevent.
Scan-to-fill reads what is on the page faster and more consistently than a tired transcription pass; it does not replace the judgement a clinician or intake coordinator brings to an ambiguous order, a missing authorization, or a physician who needs a callback. A referral with a genuinely unclear frequency is still a phone call. The point is to spend that judgement on the two or three referrals a week that actually need it, instead of on all of them.
Obillions is in private testing on iPhone, with a public TestFlight next. If your agency is still keying referrals in by hand, see what the rest of the day looks like once scheduling knows drive time too, or try the Windshield Tax calculator for what the driving side alone is costing.
Photographing a referral and having the fields โ name, address, diagnosis, frequency, ordering physician โ fill in automatically, each with a confidence score and a link back to where it appeared on the page, rather than typing every field by hand.
Commonly ten to twenty minutes per referral once you count reading the whole page, keying every field, and calling back for anything illegible or missing. Agencies rarely measure this directly, which is part of why it stays unfixed.
It should be left blank, not guessed. A fabricated value that looks plausible is more dangerous than a visible gap, because a gap prompts a phone call and a wrong-looking-right answer does not.
That risk exists with any intake method, manual or automated. The safeguard is treating a close name match as a question for a person to answer, never as something resolved automatically in either direction.
The image is read on the device; only the extracted text is sent to a server, and real patient referrals are not sent to a model until an agency's data agreement covers it. We do not claim HIPAA compliance as a marketing line.
It is in private testing on iPhone. A public TestFlight is next, then the App Store. Follow the launch on X or write to hello@obillions.com.
A voice agent for a day of visits. You say what changed and the day re-plans itself: the order, the drive times, and who nearby could take the hour a cancellation just freed. Nothing changes until you confirm, and your voice and photos never leave your phone.
In private testing now. A public TestFlight comes next, then the App Store.