Ask any operations manager where their team's hours go and the answer usually involves paper. Not literal paper anymore — PDFs attached to emails, scanned delivery dockets photographed on a phone, contracts returned as flattened scans, supplier invoices in forty different layouts. The documents are digital, but the work done with them is not. Someone opens the attachment, reads it, re-keys the fields into Xero or Airtable, checks it against a purchase order, chases a missing ABN, files it somewhere, and moves on to the next one. Multiply that by hundreds of documents a month and you have one of the largest, most measurable blocks of manual work in most service businesses.
That is why document processing is usually where we start with a new client. It is high volume, the inputs are consistent in structure even when layouts vary, the right answer is verifiable, and the time saved is easy to count. In one engagement with a wholesale distribution business, supplier invoice handling consumed roughly 22 hours a week across two people. After automation, the same volume took under four hours, almost all of it exception handling. That ratio — automating the routine 80 percent and routing the rest to a human — is typical, and it is the right target.
The anatomy of a document workflow
Every document process, regardless of industry, has the same five stages. Getting the automation right means being explicit about each one rather than treating "process the invoice" as a single step.
Intake. Documents arrive from somewhere: a shared inbox, a supplier portal, a scanner, a photo from a phone. The first automation win is usually just consolidation — one watched mailbox or upload folder that everything flows into, so nothing lives in a personal inbox. Make.com or n8n can watch a Gmail label or an Airtable form and pick up attachments the moment they arrive.
Extraction. The fields you care about get pulled out of the document: supplier name, invoice number, date, line items, totals, ABN. This is the stage people think of as "the automation," but it is only one of five, and choosing the extraction method matters more than choosing the tool around it.
Validation. Extracted values are checked against rules before anything is allowed downstream. Does the total equal the sum of the line items? Does the ABN match a supplier we know? Is the invoice number a duplicate of one already in Xero? Validation is what makes the system safe, and we cover it in detail below.
Routing. Based on the document type and the validation result, the document and its data go to the right place: a draft bill in Xero, a new record in the client file, a folder in SharePoint, or — critically — an exception queue for a human.
Archive. The original file is stored with a predictable name, linked to the record it created, so anyone can trace a Xero entry back to the source document in one click. This step is boring and frequently skipped, and it is the first thing you miss during an audit or a supplier dispute.
OCR versus AI extraction: when each is enough
Traditional OCR converts an image of text into characters. It has been reliable for decades on one condition: the layout is fixed. If every invoice from a supplier puts the invoice number in the same box, a template-based OCR setup will read it accurately and cheaply, forever. If you have three suppliers and they never change their format, template OCR is enough, and anything fancier is waste.
The problem is that most businesses do not have three suppliers with fixed layouts. They have three hundred, and layouts drift. This is where modern AI extraction earns its cost. Instead of reading a fixed region, a language-model-based extractor reads the document the way a person does: it finds "the invoice number" wherever it appears, understands that "Tax Invoice" and "Invoice" mean the same thing, and can pull line items from a table it has never seen before. It handles phone photos of crumpled delivery dockets that would defeat template OCR entirely.
Our rule of thumb in engagements: if your document sources are few and stable, use template OCR and spend nothing on AI. If sources are many or variable, go straight to AI extraction and skip the template phase — you will only end up rebuilding it. A hybrid is common too: cheap OCR for the two high-volume suppliers who send clean PDFs, AI extraction for the long tail. What we advise against is building fifty templates by hand to avoid an AI extraction cost that would have been lower than the maintenance of the templates.
Validation and the human review loop
Extraction gets the attention, but validation is where trust in the system is built. Before any extracted data is allowed to create a bill in Xero or a line in a client record, it passes through rules:
- Arithmetic checks: line items sum to the subtotal; subtotal plus tax equals the total.
- Referential checks: the supplier exists in your supplier table; the ABN matches; the bank details have not changed since last time — a change in bank details is one of the most common invoice fraud signals and should always route to a human.
- Duplicate checks: this invoice number from this supplier has not been seen before.
- Sanity checks: the date is not in the future; the amount is within the expected range for that supplier.
Anything that fails a rule does not get discarded and does not get forced through. It goes to an exception queue — in our builds, usually an Airtable view or a Softr interface — where a person sees the document next to the extracted fields, corrects what is wrong, and approves. That correction then feeds the archive and, where the tool supports it, improves future extraction.
On top of the exception queue, we recommend spot checks on the documents that passed. Review a random five percent of auto-approved items weekly for the first couple of months, then monthly. This is how you catch systematic errors — a supplier whose GST is being misread as part of the total on every invoice — that no single document would reveal.
Accuracy expectations: why 95 percent is fine
Clients often arrive expecting extraction to be either perfect or useless. Neither is true, and the expectation matters because it changes the design.
A well-built AI extraction pipeline on typical business documents lands around 90 to 98 percent field-level accuracy depending on document quality. Call it 95 percent. If your process treats that as a failure rate, you will be disappointed. If your process treats it as a routing statistic, it is excellent: 95 percent of documents flow through untouched, and 5 percent go to the exception queue where a human spends two minutes instead of the seven they would have spent keying the whole thing. The human is not a patch on a broken system — the human review loop is the design. The economics work at far lower accuracy than most people assume, because the alternative is 100 percent of documents handled manually.
What you should not accept is silent error. A system that is 95 percent accurate and tells you which 5 percent it is unsure about is production-ready. A system that is 98 percent accurate but confidently wrong on 2 percent, with no flag, will eventually put a wrong number into your accounts and shake everyone's confidence in the whole programme. Always prefer a tool and a prompt setup that reports uncertainty over one that maximises raw accuracy.
A worked example: supplier invoices into Xero
A concrete build we have delivered variations of many times:
- Suppliers send invoices to a dedicated address such as invoices@client-domain.com. Make.com watches the mailbox and saves each attachment to cloud storage.
- An AI extraction step reads the PDF and returns structured data: supplier name, ABN, invoice number, dates, line items, totals — plus a confidence flag for each field.
- Validation runs: supplier matched against the supplier table in Airtable, arithmetic checked, duplicate check against Xero's invoice history, bank details compared to the last known values.
- Clean results create a draft bill in Xero — draft, not approved — with the source PDF attached. The Airtable log is updated with status "auto-processed."
- Failures land in an exception view in Softr, showing the document beside the extracted fields. The bookkeeper works that queue daily, corrects, and approves; the correction pushes the draft bill to Xero and marks the record "human-reviewed."
- Every Friday, a reconciliation report compares the count and value of bills created in Xero that week against documents received. Any gap means something got lost between stages, and the report names it.
Total human time in steady state: the exception queue, typically ten to fifteen documents a week at a few minutes each, plus a half-hour reconciliation glance. The original manual process at that client was a full-time role's worth of keying across the team.
Common pitfalls
The failures we see in document automation are remarkably consistent:
Unreadable scans. Phone photos taken at an angle, fax-era scans, documents photographed on a wooden table with shadows. AI extraction copes with more than OCR, but there is a floor. Fix it at the source where you can — ask key suppliers to email PDFs rather than post paper — and build a "cannot read this" outcome that routes to a human instead of guessing.
Layout drift. A supplier redesigns their invoice and a template-based system silently starts putting the purchase order number in the invoice number field. This is the strongest argument for AI extraction on variable sources, and for spot checks regardless: drift is exactly the kind of error that passes every arithmetic rule because it is consistently wrong.
Missing fields. An invoice with no ABN, no due date, no line items. Decide in advance what the system does — for most clients, a missing required field means exception queue, not a blank in Xero.
Duplicate processing. The supplier resends the invoice because the first one was "lost" — which often means it went to someone's personal inbox. Consolidated intake plus a duplicate check on invoice number closes this almost completely.
No archive link. Six months later, someone disputes a payment and nobody can find the source document without digging through email. Store the file, name it predictably, link it to the record, every time.
Where to start
Pick the single highest-volume document type in your business — for most service and wholesale businesses that is supplier invoices — and count what it costs today: documents per month, minutes per document, who touches it, and how often errors surface downstream. That baseline tells you the size of the prize and gives you the numbers to measure the automation against later. Then build the five stages in order, and resist the urge to automate the second document type until the first has run cleanly for a month.
If you want an outside read on which document process in your operation would pay back fastest, book a 30-minute audit — we will map the workflow, estimate the hours it is costing you, and tell you honestly whether extraction automation is the right first move or whether something else should go first.