How AI Document Processing Actually Works in a Medical Practice

A GP sends you a three-page referral as a scanned PDF. It is slightly crooked, there are handwritten notes in the margin, and the patient’s date of birth is buried in paragraph four. Software can now read most of that document and pre-fill your intake form.

The useful question is not whether it works. It is which parts it gets right, how you find out when it hasn’t, and what happens to the document along the way. Here is the pipeline, stage by stage. It is one piece of the broader digital health shift reshaping Australian specialist practices.

Stage 2: OCR, and why it decides everything after it

Optical character recognition turns a picture of a page into characters. A referral printed straight out of a GP’s practice management system (Best Practice, Medical Director, Genie) arrives as clean typed text and reads reliably. A page that has been printed, signed, photographed on a phone in poor light and faxed twice arrives as an approximation.

Margin annotations, stamps and pages scanned at an angle are the usual culprits. If most of your referrals turn up as poor-quality scans, that is a problem with how they reach you, not with the AI. That is worth reading alongside what happens after a GP hits send.

Stage 3: matching text to fields

Once the page is text, the system has a wall of words and needs to work out which part is which field. Medicare numbers follow a fixed format. A date of birth sits near a label like “DOB”. The referring doctor is in the letterhead or the sign-off. This part is largely structural: pattern recognition against the layouts that referral correspondence tends to follow.

Stage 4: interpretation, which is not pattern matching

A letter might read: “I would appreciate your opinion on this 67-year-old gentleman with progressive right knee pain, ?meniscal tear.” Nothing there is labelled. The system has to recognise a presenting complaint, treat the question mark as marking a provisional rather than confirmed diagnosis, and place the referral in the right specialty.

Urgency is the same problem. “Routine review at your convenience” and “urgent — red flag symptoms, please see within two weeks” both need to become a priority your triage can act on. The Australian Digital Health Agency continues to work on standardised clinical document formats, but referrals still arrive in every shape imaginable, so interpretation has to be flexible.

Stage 5: confidence scoring, the part that makes it safe

This is what separates useful AI from dangerous AI. Every extracted field carries a score representing how certain the system is about that specific value. A clearly printed Medicare number in a standard position scores well. A surname half-hidden by a fold in the scan does not.

Fields above the threshold are pre-filled. Fields below it are flagged for a human, shown beside the original source text so the reviewer can check the claim against the page rather than trusting it. That is how SimpleRef’s AI Intake behaves: it does not pretend to be infallible. It processes what it is confident about and explicitly asks for help on the rest.

The review queue is the actual product

Extraction on its own is a parlour trick. What makes it useful is the queue it feeds: incoming referrals listed with their extracted data already in place, confident fields pre-filled, uncertain fields highlighted, missing fields marked.

The reviewer’s job changes from type everything from scratch to check what was found and fix what was wrong. Crucially, the system tells them where to look. The worst case, a handwritten letter through a bad fax, still needs manual entry. It is now the exception you handle rather than the default you endure. From there the referral belongs to your triage board, which is a separate discipline: see the path from referral letter to booked appointment.

What happens to the document

Any AI processing of clinical correspondence has to sit within the Privacy Act 1988 and the Australian Privacy Principles: processed securely, visible only to authorised staff, and with a straight answer available about where it goes.

Ours is this. Patient records are hosted in Australia. AI Intake is included in every SimpleRef plan and is part of how the product works. When someone uploads a document to it, that document is sent to Google’s AI service for extraction, which is not hosted in Australia. Nothing is sent to Google unless a document is uploaded. We would rather state that plainly than have a practice discover it later; the full picture, including which other suppliers sit where, is on our security page.

The bottom line

Faster intake is worth having, but it only fixes the first hour of a referral’s life. A systematic review in the Journal of General Internal Medicine found only 35% of specialty referrals result in a documented visit. Most of that gap opens up long after the data entry is done, in the quiet weeks when nobody is watching the queue. Extraction feeds the queue after a referral lands; it does not replace it.

If you want to see the mechanics, the feature list covers AI Intake and the board it feeds, pricing is published, and you can ask us a direct question about your document mix.

Then do the only evaluation that counts: take twenty typical referral letters off your own desk, run them through, and see what gets caught and what gets flagged.