Document AI that extracts, validates and knows when to ask a human
Extracting fields from a document is the easy part now. Knowing which extractions to trust, and routing the rest to somebody who can check them, is where the project succeeds or fails.
What is document AI?
Document AI is the automated reading of invoices, contracts, forms, statements and scans to extract structured data, classify the document and validate it against your systems. It suits Australian organisations processing high volumes of paperwork by hand, where accuracy matters enough that anything uncertain should reach a person instead of going straight through.
Get a fixed written quote- Typical timeline
- 7 to 14 weeks
- What drives cost
- The number of document types, how variable their layouts are, and how much of the output a person has to check before it is trusted.
- Best for
- High volumes of paperwork keyed in by hand today
- You own
- The pipeline, the extracted data, the audit trail and the accounts
- Built with
- OCR, extraction models, validation rules, exception queues
Your handover
Straight through processing is the number that matters
Accuracy per field sounds impressive and tells you almost nothing. The measure that changes your staffing is the proportion of documents that pass all the way through without a human touching them, which we call the straight through rate. A pipeline extracting ninety eight percent of fields correctly can still send every second invoice to review, because one uncertain field is enough to hold the whole document. That is the arithmetic to design around from the beginning.
- 01Assessment of your real documents with an achievable straight through rate
- 02Ingestion from email, scanner or portal
- 03Classification and extraction per document type
- 04Validation rules including GST, ABN and duplicate checks
- 05Confidence thresholds calibrated with you
- 06Exception review screen with document and fields side by side
- Integration into your accounting, ERP or line of business system
- Audit trail of every extraction, correction and approval
- Weekly reporting on straight through rate and error types
Set them high and everything is safe and slow
So the work is tuning the confidence thresholds against your tolerance for error. Set them high and everything is safe and slow. Set them low and mistakes reach your ledger. We calibrate against a labelled sample of your real documents, show you the trade off curve, and let you pick the point, because the right threshold for a two hundred dollar consumable is not the right threshold for a supplier contract. That choice belongs to the business, not to us and certainly not to a default setting.
Extraction is easy, validation is the work
A model can read a total off an invoice. It cannot tell you the total is wrong. Validation is where a document pipeline earns its cost: check that the line items sum to the subtotal, that the GST component is arithmetically consistent, that the ABN exists and matches the supplier on record, that the purchase order number is real and has budget remaining, that this invoice number has not already been paid. Each check is unglamorous and each one catches something that would otherwise have become a phone call in three weeks.
The checks also let the system be honest about uncertainty
The checks also let the system be honest about uncertainty. When extraction is confident and every validation passes, the document goes through. When extraction is confident but a validation fails, that is a business exception and it needs a person, which is a different queue from a document that was merely blurry. Distinguishing those two cases is one of the more valuable things we build, because it stops your finance team wading through scanning problems to find the genuine discrepancies. Where the validations depend on live data such as supplier records or budgets, they run against your ERP rather than a stale export.
How the engagement runs
How a document AI project runs
We begin with your documents, not with a platform choice. A sample of several hundred real items, including the bad ones people are embarrassed to send, tells us the achievable straight through rate before you commit to anything. Clients who send only their tidiest examples get an optimistic estimate and an unpleasant surprise, so we ask for the messy pile explicitly.
- 01Document sampleSeveral hundred real items including poor quality scans and edge cases
- 02Field specificationWhat must be captured, what is optional, and the correct value for each in the sample
- 03Pipeline buildIngestion from email, scanner or portal, classification, then extraction per document type
- 04Validation rulesArithmetic, ABN and supplier matching, purchase order lookups and duplicate detection
- 05Threshold calibrationConfidence levels set against the trade off curve you have chosen
- 06Exception queueA review screen showing the document beside the extracted fields, with the reason it was held
- 07Parallel runAutomated and manual processing side by side, differences reconciled
- 08Cutover and monitoringStraight through rate tracked weekly, with corrections fed back into the model
Two decisions on your side that keep the project moving
Rollout runs in parallel with the manual process at first. Both run, results are compared, and the differences are examined one by one. It costs a few weeks of duplicated effort and it is the only way to know the real error rate before you switch the manual path off. Nobody enjoys that phase and everybody is glad of it afterwards.
Choose the right level
What to expect from different document types
Vendors demonstrate on clean documents. Your reality includes a photograph of a delivery docket taken on a job site in poor light, folded twice, with a signature across the total. Expectations should be set per document type rather than in general, and we assess a real sample of yours before quoting rather than promising a headline accuracy figure that came from someone else's data.
Document type
01
Digital PDF invoices
Realistic outcome
High straight through rate
What limits it
Layout variation between suppliers, handled with a short learning period
02
Scanned invoices and statements
Realistic outcome
Good, with a review queue
What limits it
Scan quality, skew and stamps or handwriting over printed fields
03
Photographed dockets and receipts
Realistic outcome
Moderate, expect regular review
What limits it
Lighting, creases, thermal paper that has faded
04
Structured forms you control
Realistic outcome
Very high, and cheapest of all
What limits it
Only if the form is designed for capture rather than for print
05
Contracts and long agreements
Realistic outcome
Useful extraction, always reviewed
What limits it
Clauses are interpretive, and a missed obligation is expensive
06
Handwritten forms
Realistic outcome
Partial at best
What limits it
Handwriting recognition remains genuinely unreliable at scale
How we work this out during scoping
The table below reflects what we typically see. Where a type sits at the harder end, the answer is not usually a better model. It is changing the input: a form redesigned so the fields are machine friendly, a scanning standard for the field crew, or a supplier asked to send a structured file instead.
Privacy, retention and where documents are processed
Business documents are stuffed with personal information: names, addresses, bank details, sometimes health information or a tax file number that should never have been on the form in the first place. Sending those to a processing service overseas is a cross border disclosure under APP 8, and your organisation stays accountable for it. APP 11 obliges you to protect the information you hold, which in this context means encryption at rest, access limited to the people who need it, and audit logging of who viewed what.
More on privacy, retention and where documents are processed
We default to processing in Australian regions and keeping the document store and audit trail on Australian infrastructure, and where a client needs an overseas capability we document the flow so their privacy officer can assess it properly rather than discovering it later. Retention is set deliberately: records supporting tax positions generally need to be kept for five years under ATO requirements, while the intermediate artefacts a pipeline creates should be purged far sooner. Deleting what you no longer need is a security control, and it is the one most often forgotten.
When document AI is the wrong fit
If you process thirty invoices a month, the build and supervision will cost more than the keying. If your documents are unique every time, there is no pattern to learn and no throughput to gain. If a mistake is catastrophic and unrecoverable, keep a human on every item and use the pipeline only to prefill fields, which still saves time without transferring risk to a confidence score.
The rest of the answer
There is also a better answer hiding in plain sight for many finance teams. Electronic invoicing over the Peppol network, which the ATO has been encouraging for years, delivers structured invoice data directly into your accounting system with no reading required and no extraction errors possible. If a meaningful share of your suppliers can adopt it, ask them before building a pipeline to photograph what they could have sent as data. Where the paperwork is only one step in a longer process, the wider win is usually in workflow automation rather than the extraction itself, and for construction businesses chasing dockets and variations that is nearly always where the real time goes.
How we scope it
Four ways to scope your Document AI project
We do not publish package prices, because the same brief can be a short build or a long one. These are the shapes the work usually takes. Tell us which one sounds like you and you will get a fixed written quote that spells out exactly what it covers.
Proof of value
One use case, evaluated honestly before it goes near a customer
Fixed written quote, agreed before work starts
- Assessment of your real documents with an achievable straight through rate
- Ingestion from email, scanner or portal
- Classification and extraction per document type
Production build
In production, with a human approval step and an evaluation set
Fixed written quote, agreed before work starts
- Everything in Proof of value
- Validation rules including GST, ABN and duplicate checks
- Confidence thresholds calibrated with you
- Exception review screen with document and fields side by side
Embedded platform
Built into the product rather than bolted onto it
Fixed written quote, agreed before work starts
- Everything in Production build
- Integration into your accounting, ERP or line of business system
- Audit trail of every extraction, correction and approval
- Weekly reporting on straight through rate and error types
Document AI Model care
Monitoring, evaluation and retraining as the inputs drift
Rolling monthly, quoted in writing
- Evaluation set rerun as the model and the inputs change
- Cost and quality reported monthly, not assumed
- Prompt, tool and guardrail changes as the work shifts
- Rolling, cancel with 30 days notice
These are shapes, not menus. Most quotes end up somewhere between two of them, and we will say so when the honest answer is the smallest one. Describe the problem and we will tell you which it is.
Questions buyers usually ask
Frequently asked questions
How long does a document AI implementation take?
Typically 7 to 14 weeks, including a parallel run. Assessing your document sample and specifying the fields takes the first couple of weeks, the pipeline and validation rules the middle stretch, and the parallel run at least a fortnight. More document types means more time, since each type needs its own extraction logic, validation rules and testing.
What does document AI cost to run?
There is a build cost, quoted in writing after we assess your documents, and a per document processing cost paid to the platform or model provider. At high volume that running cost is the one to watch, so we report it per document from the pilot onward. Against manual keying it usually compares well, though only once the straight through rate is known rather than assumed.
Who owns the extracted data and the documents?
You do. Documents and extracted data sit in storage under your accounts, the pipeline code is in your repository, and the audit trail is yours to export. Your documents are not used to train a shared model. If you move to another provider, the extracted data is in your systems already and nothing has to be recovered from us.
What accuracy can you actually promise?
We do not promise a number before seeing your documents, because any figure quoted in advance came from somebody else's paperwork. After assessing a real sample we give a measured straight through rate for each document type and the error rate at your chosen confidence thresholds. Those numbers go in the scope. That is a slower answer than most vendors give and a considerably more reliable one.
Can it handle our supplier invoices arriving in dozens of formats?
Yes, that is the standard case. Modern extraction does not need a template per supplier, though a short learning period per new layout is normal and accuracy improves as corrections feed back. What causes genuine trouble is not format variety but poor scan quality and invoices that omit a purchase order reference, since neither can be recovered by better reading.
What happens to the documents that fail validation?
They go to an exception queue with the reason attached, showing the document alongside the extracted fields so a person can resolve it in seconds rather than reopening the original. Corrections are captured and used to improve extraction. We separate scanning problems from genuine business discrepancies, so your finance team is not hunting real issues among blurry images. Approval routing for those exceptions is usually automated as part of the same build.
Related services
Send us your messiest documents
A sample of real paperwork, including the poor scans, tells us more than any brief. We reply within one business day with an honest read on the achievable straight through rate and a fixed written quote.