Guide · Documents · Melbourne, October 2026
Somewhere in your business, right now, someone is holding a docket up to the window to work out whether that’s a 6 or a coffee stain. You’ve been told OCR will fix it. You’ve also been told AI will fix it. This guide lays out document AI vs OCR in plain English: what each one actually does, where OCR is all you need, where it quietly falls over, and what both cost per page in 2026. There’s a docket to test them on. Dave signed it.
Fixed price A$1,950 + GST. We look at your real documents, not a demo PDF.
What OCR gives you
WEIGHBR1DGE DOCKET No. 1O417 DATE 07/1O/26 CUST J. Citizen Constructi0ns REGO XYZ-l23 MATERIAL Mixed C&D waste GROSS 24.86 t 24.6B TARE 11.2O t NET l3.66 t SIGNED Dove
Every character, in order, with no idea which one matters. Two gross weights, no hint which is real. And Dave is now a bird.
What document AI gives you
Fields with names, a rule that does the subtraction, and a person who decides. Dave stays Dave.
Short answer: The difference in document AI vs OCR is reading versus understanding. OCR (optical character recognition) turns an image of a page into text. Document AI usually uses OCR as its first step (some newer tools read the page image directly with a vision model), then works out what kind of document it is, pulls out named fields like the ABN, total or net weight, checks them, and sends doubtful values to a person. Use OCR when you only need searchable text. Use document AI when someone currently retypes values from the page into another system.
Side by side
If you only read one thing on this page, read this table. It’s the whole argument, minus the jokes.
| OCR | Document AI | |
|---|---|---|
| What it does | Turns pixels into characters | Turns a document into named, checked fields |
| What you get | A text layer or a text file, in reading order (mostly) | Structured data: “Supplier ABN: 12 345 678 901”, with where on the page it came from |
| Knows what a document is? | No. An invoice and a lunch menu are both text | Yes. It classifies each document, and each page of a bundle |
| Layout changes | Template-based setups break when a supplier redesigns their invoice | Finds the total wherever this supplier has decided to put it |
| Handwriting | Clean print: very good. Handwriting: patchy | Better, with a confidence score per value, and still not perfect |
| Checks the values? | No | Yes, when paired with rules: sums, dates, ABN format, matches to your records |
| Typical cost per page | Fractions of a cent to free | A few cents, plus the cost of setting up the workflow |
| Best for | Making scans searchable, archives, copy and paste | Anything someone currently retypes into Xero, a register, a CRM or a spreadsheet |
Per-page prices are covered in detail, with dates and sources, in the costs section below.
The quiet hero
Optical character recognition looks at an image of text and works out which letters and numbers are in it. That’s the job. It’s an old, mature technology: Tesseract, the best-known free OCR engine, was built at HP Labs between 1985 and 1994, released as open source in 2005 and developed by Google from 2006 to 2017. Today OCR software is built into your phone, your scanner, Adobe Acrobat and most document management systems, usually for free.
And for a lot of jobs, it’s perfect. If you want a decade of scanned contracts to be searchable, OCR. If you need to copy a paragraph out of a PDF, OCR. If your bookkeeper wants to find every letter that mentions “variation 14”, OCR. No AI required, no project, no invoice from us.
What OCR doesn’t do is understand anything. It doesn’t know that “12 345 678 901” is an ABN, that “24.68” crossed out next to “24.86” means the operator corrected the weight, or that page 37 of a permit file is the insurance certificate. To OCR, it’s all just characters, laid out roughly left to right. [Adjusts reading glasses.] Very confidently laid out, mind you.
Open one of your PDFs and press Ctrl + F (or Cmd + F on a Mac). Search for a word you can see on the page. If it finds it, the PDF already has a text layer and doesn’t need OCR at all. Most invoices emailed straight from accounting software pass this test. Most scans, phone photos and “print, sign, scan” documents fail it.
Plenty of “we need OCR” conversations could end right here, which is great news for everyone’s budget and terrible news for our pipeline.
What changes
Document AI (also called intelligent document processing, or IDP) isn’t a rival to OCR. It’s usually OCR plus four jobs that a person used to do in their head. In the order they happen:
The image becomes text, with the position of every word kept. That position matters later.
If skipped: there’s nothing to work with. Even Microsoft documents its Read OCR model as the engine underneath its prebuilt invoice and layout models.
Invoice, delivery docket, credit note, insurance certificate, care plan. A 60-page scanned bundle gets split into the documents inside it.
If skipped: the credit note gets processed as an invoice and you pay a supplier for the privilege of being refunded.
Supplier, ABN, date, GST, total, line items, net weight, lot number. Wherever they sit on this particular layout, with a confidence score for each one.
If skipped: someone still retypes them, just from a nicer-looking text file.
Ordinary rules, not AI: does gross minus tare equal net? Is the ABN 11 digits? Does the PO number exist? Is this docket number already in the register?
If skipped: mistakes arrive in your system faster and in larger numbers. Congratulations, you’ve automated the error.
Anything with low confidence or a failed check goes to a named person, with the source image beside the value. Everything else flows through.
If skipped: nobody notices the wrong net weight until the EPA, the auditor or the client does.
Notice that only two of those five steps are AI. That’s on purpose. Rules are cheaper, faster and completely predictable, so we use them for anything a rule can do: the maths, the formats, the lookups. AI earns its place in the two steps that need judgement about messy input, and a person keeps the decisions. If you want the long version of that idea, we wrote it up as human in the loop, defined.
This is the core of what we build in our document intelligence work: AI extraction plus rules-based validation, with every value linked back to the page it came from. That last part matters more than it sounds. When an auditor asks where a number came from, “the AI said so” is not an answer anyone enjoys giving.
The five documents
These are the documents we see most in the industries we work with: waste and EPA compliance, construction, building surveying, aged care and accounting. Each one is easy for a person to read and surprisingly hard for OCR alone.
Waste and EPA
Construction and accounting
Construction
Building surveying
Aged care
The pattern
Handwriting deserves its own moment, because it’s where most OCR software projects go to die. Printed text comes in a few thousand fonts. Handwriting comes in one font per person, and some people’s font changes after lunch. Even in lab conditions it isn’t solved: Microsoft’s TrOCR model set a state-of-the-art result on a standard handwriting benchmark by getting about 2.9% of characters wrong. That benchmark is neat writing on clean paper. Your site diary is neither.
Modern document AI handles handwriting far better than old OCR did, and it’s genuinely impressive on neat block capitals. But “better” isn’t “always right”, so we never let handwritten values flow straight into a register. They get a confidence score, and anything under the threshold gets a human review. Dave’s handwriting, for the record, is always under the threshold.
Accuracy
Every OCR software brochure has a big number on it, usually starting with 99. Here’s the catch: that’s almost always character accuracy. Your business doesn’t run on characters. It runs on fields: the total, the ABN, the BSB, the net weight.
Do the maths on a typical invoice page with about 1,000 characters. At 99% character accuracy, around 10 of them are wrong. If those 10 land in the terms and conditions, nobody cares. If one lands in the total or the bank account number, the whole invoice is wrong. Errors also compound: a 1% character error rate can mean a word error rate of 5% or worse. And older or damaged paper is much harder. When the National Library of Australia measured raw OCR on digitised newspapers for Trove, accuracy ranged from 71% to 98% depending on the page.
That’s why the number that matters is field-level accuracy on your documents, and why the step after data extraction, checking, matters more than the extraction itself.
The fix isn’t a more accurate model. It’s checks. A rule that confirms the bank details match the supplier record catches the error whether a person, OCR or AI made it. If you want to see what errors from manual keying actually cost, start with the hidden cost of manual data entry.
Costs
Short version: the software is cheap, and the person checking the output is not. Here’s what the big three cloud providers list per 1,000 pages for plain OCR and for document AI models.
| Provider | Plain OCR | Invoice or receipt model | Custom extraction |
|---|---|---|---|
| Google Document AI | US$1.50 (first 1,000 pages free) | About US$100 (US$0.10 per document of up to 10 pages) | US$30 |
| Azure AI Document Intelligence | US$1.50 (500 pages a month free) | US$10 | US$30 |
| AWS Textract | US$1.50 | US$10 | Not listed as a custom model; prebuilt forms model US$50 |
First volume tier, pay as you go. AWS Textract has no like-for-like custom extraction price, so the forms model is shown instead. Prices fall at higher volumes and change over time, so check the provider’s page before you budget. Sources: Google, Azure (East US, via Microsoft’s retail prices API), AWS.
A made-up but very familiar transfer station takes about 300 paper dockets a week. Someone keys each one into the waste register, which takes about 90 seconds with checking. Here’s what the three approaches cost in people time, at A$55 an hour (wages plus on-costs) over 46 working weeks.
| Approach | Assumption | Hours a week | Labour a year |
|---|---|---|---|
| Manual keying | 90 seconds a docket | 7.5 | A$18,975 |
| OCR, then a person copies the fields | 45 seconds a docket: less typing, same checking | 3.75 | A$9,488 |
| Document AI, rules and human review | 1 in 5 dockets flagged, 60 seconds each, plus 30 minutes of spot checks | 1.5 | A$3,795 |
Illustrative only. Hours a year = hours a week × 46. Labour excludes the cost of building and running the workflow; for what that costs, see our guide to AI consultant cost in Melbourne. Actual results depend on your documents, volumes and process.
At about 13,800 dockets a year, a custom extraction model at US$30 per 1,000 pages comes to roughly US$414 a year. Even the dearest option in the table, about US$1,380, is a small slice of the roughly A$15,000 labour gap between manual keying and document AI. The page price is not where the money goes. The real comparison is people time, and that’s where OCR alone disappoints: it removes the typing, but not the reading, matching and checking.
Our rule of thumb is the same one from our ROI of AI automation guide: a first document project should pay for itself within about a year from the hours it gives back. On these numbers, a modest build clears that bar. On 30 dockets a week, it might not, and we’d tell you to keep the clipboard.
7.5 hours a week
About 1.5 hours a week
Decision
You don’t need a strategy workshop for this one. You need to look at what the person doing the job today actually does with the page.
Three or four “yes, lots, it’s bad, too many” answers? That’s a document AI job. Mostly “no”? Buy decent OCR software, or use the one already in your scanner, and spend the money on something else. Sometimes the real fix isn’t documents at all but plain workflow automation: getting suppliers to email PDFs that pass the Ctrl+F test instead of posting paper.
Human review
Neither OCR nor document AI should be the last word on anything that costs money, affects a person or ends up in a compliance record. Here’s how we split the work on a typical document workflow.
| Step | Handled by | Why |
|---|---|---|
| Turning the image into text | OCR | Mature, cheap and good at print. |
| Working out what each document is | AI | Needs judgement about layout and content. |
| Data extraction of named fields | AI | Layouts vary; every value gets a confidence score. |
| Sums, formats, duplicates, lookups | Rules | Predictable checks should be predictable. |
| Low-confidence or failed values | Person | With the source image beside the value. |
| Approving payments and submissions | Person | Accountability stays with your team. |
| Anything that decides about a person | Person | For example, the severity of an aged care incident. |
None of these rules says you can’t use document AI. They do mean the workflow has to keep the source document, record who checked what, and know where the data goes. That’s design work, and it belongs at the start of a project, not after go-live.
Common mistakes
We’re not proud of it, but we find this part fun. Here’s what goes wrong most often.
Our take
We use both, often in the same workflow. OCR does the reading, AI does the classifying and data extraction, rules do the checking, and a named person on your team makes the calls that matter. Every value stays linked to the page it came from, so you can always answer “where did that number come from?”
We don’t sell a document AI product. We start with how the work actually moves through your business, then pick the cheapest thing that fixes it. Sometimes that’s document AI. Sometimes it’s OCR you already own. Sometimes it’s asking three suppliers to stop sending faxes. (Yes, still.)
See the full service on our document intelligence page, browse the automation use cases, or read how the same idea works for invoice processing in Xero.
Who wrote this
FAQ
No. OCR turns an image of text into text. Document AI usually starts with OCR, then classifies the document, extracts named fields such as the supplier, ABN, total or net weight, and gives each value a confidence score. Paired with rules and human review, it can check those values and send doubtful ones to a person. OCR reads. Document AI reads and organises.
Modern OCR engines use machine learning, so technically yes. Tesseract, the best-known free engine, has a neural network (LSTM) engine in its current version 5. But OCR only recognises characters. It doesn’t know what the document is or what any value means, which is the part people usually mean when they say document AI.
Partly. OCR software is built for print and does well on clean typed text. Handwriting is much harder, and even research models get a few percent of characters wrong on neat handwriting. Document AI reads handwriting better and scores its confidence in each value, so shaky values can go to a person instead of straight into your records.
Usually not. Open the PDF and press Ctrl+F (Cmd+F on a Mac), then search for a word you can see. If it finds it, the PDF already has a text layer. Most invoices emailed from accounting software pass this test. Scans, phone photos and print-sign-scan documents fail it and need OCR first.
At October 2026 list prices, plain OCR costs about US$1.50 per 1,000 pages on Google, Azure and AWS. Prebuilt invoice models cost about US$10 per 1,000 pages on Azure and AWS, and custom extraction about US$30. The bigger cost is setting up the workflow and the people time spent checking the output, not the page price.
It depends on your documents, so ask for field-level accuracy on a sample of your own, including the messy ones. A 99% character accuracy figure can still mean a wrong total or ABN on some invoices. The reliable answer is to pair extraction with rules that check the values and a person who reviews anything that fails a check or has low confidence.
It can be, with care. The OAIC advises against entering personal information, particularly sensitive information, into publicly available generative AI tools, and says a person should verify personal information obtained through AI. Before choosing a tool, check where documents are processed and stored, whether they are used for training, and who reviews the output.
A 30-minute fit call is free. The Pain Point Audit is A$1,950 + GST for a half-day walkthrough of your documents and processes, with a written report within 5 business days and a fixed price for the first build. The fee is credited in full against a first build of A$5,000 + GST or more signed within 60 days.
Bring us the tray. The Pain Point Audit is a half-day look at how documents really move through your business, with a written report in 5 business days: which documents cost the most time, what OCR can handle, where document AI earns its keep, what stays with your people and a fixed price for the first fix. Dave can keep his handwriting.
Fixed price A$1,950 + GST. Credited if you go ahead with a build.
