We Built an AI Document Scanner That Replaced Adobe Scan

Paper refuses to die. We run a dental practice in Darwin, and despite everything being "digital" now, a surprising amount of our daily intake is still physical documents. Referral letters from GPs. Lab reports. Insurance pre-approvals. Medical history forms that patients fill out by hand. Compliance certificates. X-ray reports from external imaging centres.

Every single one of those documents needs to be scanned, read, understood, and filed in the right place. And until recently, that was someone's job — standing at the scanner, opening each PDF, figuring out what it was, pulling out the key information, navigating to the correct patient record, and saving it. Over and over, 15 to 20 times a day.

It's the kind of task that feels small until you add it up. So we did.

The old workflow: four minutes per document

Here's what the process looked like before we automated it:

1. Scan the document — place it on the scanner, wait for the PDF (~1 minute)
2. Open and read it — figure out what you're looking at (~1 minute)
3. Determine the document type — is it a referral? A lab report? An insurance form? (~30 seconds)
4. Extract key data manually — patient name, referring doctor, date, findings, relevant details (~1 minute)
5. Navigate to the correct patient or folder — find the right record in the system (~30 seconds)
6. File it — save, rename, attach to the record (~30 seconds)

That's roughly four minutes per document. At 15 to 20 documents a day, that's 60 to 80 minutes of staff time — every single day — spent on what is essentially data entry with extra steps.

Over a week, that's five to six hours. Over a year, it's over 250 hours. An entire month and a half of working days, spent scanning and filing paper.

What we tried first

We looked at the usual suspects:

• Adobe Acrobat Pro (includes Adobe Scan) — roughly ~$30/month. Good OCR, decent scanning from mobile, and it does produce searchable PDFs. But it doesn't classify anything. It doesn't extract data. It doesn't know what a referral letter is or which patient it belongs to. You still have to read, understand, and file every document yourself.
• ABBYY FineReader — roughly ~$20–25/month. Better OCR accuracy than Adobe in our experience, especially with mixed layouts. But the same fundamental problem: it gives you text from a scan. What you do with that text is still your problem.

Both tools solve the "turning paper into digital text" part. Neither solves the "understanding what the document is and doing something useful with it" part. The OCR was never the real bottleneck — the human processing afterwards was.

What the AI does now

We built a system that handles the entire pipeline. Here's the workflow:

1. Document placed on the scanner — this is the only manual step. Someone puts the paper on the scanner and hits the button.
2. OCR extracts the text — local OCR processes the scanned image into machine-readable text. No cloud service, no subscription.
3. AI classifies the document type — the text is sent to a local language model that determines what kind of document it is: referral letter, lab report, insurance pre-approval, compliance certificate, medical history form, X-ray report, or something else entirely.
4. AI extracts key data — depending on the document type, the AI pulls out the relevant fields. For a referral letter, that's the patient name, referring doctor, date, reason for referral, and any clinical notes. For a lab report, it's the patient, the test type, the results, and any flags. For insurance, it's the policy number, approval status, and covered items.
5. Auto-files in the correct location — the system matches the document to the right patient record and saves it in the appropriate folder, properly named and categorised.
6. Logs the action — every processed document is logged: what it was, what was extracted, where it was filed, and when. Full audit trail.

Time per document after placing it on the scanner: about 10 seconds. Fully automated. No one needs to read it, classify it, or file it.

Before and after

The numbers are hard to argue with:

• Time per document: dropped from ~4 minutes to ~10 seconds
• Daily time spent on document processing: dropped from 60–80 minutes to under 5 minutes (just placing documents on the scanner)
• Weekly time saved: 5+ hours of staff time returned to actual patient-facing work
• Monthly software cost: $0 — down from ~$30/month for Adobe Acrobat Pro
• Misfiled documents: effectively zero. The AI matches to patient records consistently. Humans, especially when rushing through a stack of 20 documents before lunch, make mistakes. The AI doesn't get tired or distracted.

At a loaded staff cost of around $40–45/hour, five hours a week of saved time is roughly $200–225 per week — over $10,000 per year. That's the real saving. The $30/month in cancelled software is almost a rounding error compared to the time.

We wrote about this pattern — where the software subscription is cheap but the hidden staff time cost is enormous — in our post on the real cost of AI tools. Document processing is one of the clearest examples.

What it handles well

The system reliably processes:

• Referral letters — from GPs, specialists, and other dentists. Extracts referring provider, patient details, reason for referral, and clinical notes.
• Lab reports — crown and bridge work, denture cases, orthodontic appliances. Pulls out case details, materials, and completion status.
• X-ray and imaging reports — from external imaging centres. Extracts findings, recommendations, and the referring clinician.
• Insurance pre-approvals — identifies the insurer, policy holder, approved items, and any exclusions or waiting periods.
• Medical history forms — patient-completed forms with conditions, medications, allergies, and contact details.
• Compliance certificates — equipment servicing records, radiation safety, infection control audits.

Each document type has its own extraction template, so the AI knows exactly what fields to look for. It's not guessing — it's following a structured process with AI handling the understanding part.

What it can't handle

We're always honest about limitations, because setting false expectations helps no one.

• Handwritten notes — OCR still struggles with handwriting, especially the famously illegible variety that doctors are known for. If a GP handwrites a referral letter (it happens more than you'd think), the system flags it for manual review rather than guessing badly.
• Badly damaged documents — crumpled, water-stained, or heavily faded documents don't OCR well. The AI can only work with the text it receives, and if the OCR output is garbled, the classification and extraction will be unreliable.
• Completely novel document types — if someone sends us a type of document we've never seen before, the AI may misclassify it on the first encounter. It flags low-confidence classifications for human review, so nothing gets silently misfiled.

The design principle is the same one we apply across all our AI systems: when the AI isn't confident, it escalates to a human. It never silently guesses. A misfiled document is worse than an unfiled one, so the system errs on the side of caution.

This isn't just a dental practice problem

Document processing is one of those universal pain points that every business with a physical paper trail deals with. We built this for our practice, but the pattern applies directly to:

• Law firms — court documents, contracts, affidavits, correspondence. A litigation practice might process dozens of documents a day, each needing to be filed against the correct matter.
• Accounting firms — invoices, receipts, bank statements, tax documents. Every client has a folder, and every document needs to end up in the right one.
• Real estate agencies — contracts of sale, building inspection reports, strata documents, valuation reports. Each property and each transaction generates a stack of paper.
• Trades businesses — compliance certificates, insurance documentation, quotes, purchase orders. The admin side of a trades business generates more paper than most people realise.
• Medical practices of all types — pathology results, specialist reports, imaging, insurance. Same problem we had, different speciality.

If your staff spends more than 30 minutes a day processing, classifying, and filing documents, there's a strong case for automating it. We talked about the broader opportunity in our post on what AI can actually do for small business — document processing is one of the highest-return areas we've found.

Why local matters

Our system runs entirely on local hardware. The OCR runs locally. The AI classification runs locally. No document content — which may include patient names, medical conditions, Medicare numbers, insurance details — ever leaves our network.

This isn't just a nice-to-have for a healthcare practice. When you're processing documents that contain protected health information, the simplest way to guarantee compliance is to never send that data anywhere. No third-party processing agreements. No wondering which country your patient's referral letter is being OCR'd in. It stays on our machine, gets processed, gets filed, and that's it.

The bottom line

Adobe Scan and ABBYY are fine tools for turning paper into PDFs. But turning paper into PDFs was never the hard part. The hard part was always what comes after: reading the document, understanding what it is, extracting the important information, and filing it where it belongs. That's the work that eats 60 to 80 minutes a day, and that's the work the AI now handles.

We went from a person standing at a scanner processing documents for over an hour a day to a system that does it in seconds. No monthly subscription. No cloud dependency. No misfiled documents turning up in the wrong patient's record three weeks later.

If your business processes paper — and most do, whether they like it or not — get in touch. We'll walk through your current document workflow and show you exactly where the time is going and how much of it we can give back. No obligation, just a practical conversation about paper, AI, and getting your staff back to the work that actually matters.