how a twelve-person clinic stopped keying referrals by hand
this is an invented example. no client of ours is described here. we have written it as a full project, including the part most vendors skip, because "ai can read your documents" is useless and a specific build you can argue with is not.
this one carries more risk than our other worked examples, so the privacy work comes first, in the project and on this page.
the situation
carling family health. twelve people. two physicians, a nurse practitioner, four admin, the rest clinical support.
referrals arrive three ways: faxes that land as PDFs in a shared inbox, emails with attachments in six different formats, and a provincial portal that exports something structured but only for one program.
priya handles them. she keys each one into the practice management system: patient details, referring provider, reason, urgency, any attached history.
ten minutes each, thirty a week, five hours. but it is never five uninterrupted hours. it happens in four-minute fragments between phone calls, which is exactly where transcription errors come from. twice last year a referral sat unprocessed for nine days because it arrived during a busy afternoon and got buried.
week zero: the privacy decision
before a single document is processed, one question gets answered in writing: what may be sent to a model, and what may not.
in ontario this is phipa. it is not a reason to avoid the work, it is a reason to do this first. what gets decided and written down:
- de-identification before anything leaves the clinic's systems. names, health card numbers, addresses and dates of birth are stripped locally and re-attached locally. the model sees "patient, 60s, referred for X".
- where processing happens, and whether data residency is a requirement for this clinic.
- what the provider retains, and for how long.
- who approved it, with a date.
this is a deliverable, not a disclaimer. it is roughly a week of work and the clinic keeps it regardless of what gets built afterwards. if a vendor skips this step, that tells you what you need to know about the vendor.
what was actually built
1. the de-identification layer (week two)
runs inside the clinic's own environment. strips identifiers, keeps a local mapping, re-attaches on the way back out. nothing identifying crosses the boundary.
2. extraction with calibrated uncertainty (weeks three and four)
each document is read and the fields extracted, referring provider, reason, urgency, history. every field carries a confidence score.
anything below the threshold does not get written. it goes into a queue.
3. the review queue (week four)
priya opens one screen each morning. it shows only the fields the system was unsure about, with the source document beside them. she confirms or corrects, and it writes through to the practice management system.
the queue is the product. a version of this without it would be faster, and we would not build it.
the 3% that matters more than the 97%
an extraction system at 97% accuracy sounds excellent. it is dangerous.
the danger is not the three per cent. it is not knowing which three per cent. a system that silently gets thirty documents a year wrong, scattered invisibly through the other nine hundred and seventy, is worse than one that is eighty per cent accurate and tells you every time it is unsure.
so the design requirement is not accuracy, it is calibrated uncertainty. The system knowing what it does not know. the threshold is set deliberately, reviewed after the first month, and tuned so the queue catches the errors without becoming so long that priya stops reading it carefully.
how long it took
| week | what happened |
|---|---|
| 0 | privacy assessment. what may leave the clinic, written down and approved. |
| 1 | document audit, 200 real referrals sorted by source and format. three formats accounted for 80%. |
| 2 | de-identification layer built and tested against the 200. |
| 3 | extraction built for the three common formats. |
| 4 | review queue built. priya tested it on live referrals in parallel with keying them by hand. |
| 5 | parallel running. every document done both ways and compared. |
| 6 | threshold tuned. it was too confident on urgency, which is the one field where being wrong matters most. we lowered it and the queue got longer. |
| 7 | live, hand-keying stopped. |
live in seven weeks, of which one was privacy and two were parallel running. we would not shorten either.
what changed
| before | after | |
|---|---|---|
| minutes per document | 10 | 3 |
| hours a week | 5 | 1.5 |
| longest a referral sat unprocessed | 9 days | same day |
| errors found | when someone noticed | at the queue, before writing |
≈ 3.5 hours a week back.
in a twelve-person clinic that is most of a day. realistically it goes into phones answered and appointments booked, not anything strategic, which is fine, because answering the phone is the constraint in most clinics.
the thing priya noticed was not the hours. it was not holding thirty half-finished documents in her head.
what it would cost
honest numbers for a build this size:
- the privacy assessment: $3,000–5,000, delivered as a document the clinic keeps
- the build: $12,000–18,000, weighted toward extraction and the queue
- running it: roughly $80–150 a month at this volume
- payback: 3.5 hours a week of admin time, plus the referrals that no longer sit for nine days
in ontario the planning portion may be eligible for a matched grant through the digital competence centre.
what we would not build
- clinical judgement. summarising a referral for triage is a different thing from reading it, and it is not what this is for.
- auto-processing of anything flagged. flagged means a person looks, every time.
- the decision to accept a referral.
- anything touching the portal's structured feed. it already arrives clean; leave it alone.
when this is not worth doing
if your referrals all arrive through one structured portal, the savings are much smaller and you should not do this. this pays where the inputs are messy and arriving in three or more formats.
if you have fewer than about fifteen documents a week, the queue-reading habit will not stick and the build will not pay.
if you want this worked out with your real numbers rather than carling's invented ones, that is what a readiness review is, and for a health organisation it starts with the privacy question, not the software.
other worked examples: funder reporting in a charity · quoting in a trades business