AI is now genuinely good at reading the routine paperwork that buries a back office: customer purchase orders that arrive as email PDFs, delivery dockets and proof-of-delivery photos, supplier invoices, timesheets on paper, and certificates of currency. It can pull out the fields that matter, match each document to the right order, job or supplier, and flag anything that doesn't line up. It still struggles with handwriting, poor scans and messy records behind the scenes, so a person should check before anything is confirmed, dispatched or paid. And the real value isn't the reading itself. It's that the information goes into your stock, job or accounts system once, instead of someone typing it in again.
The paperwork we're talking about
If you run operations or finance in wholesale, manufacturing, transport, construction or facility services, the pile probably looks something like this:
- Customer purchase orders arriving by email as PDFs, in every customer's own layout, sometimes as a photo of a printed form.
- Delivery dockets and proofs of delivery (PODs): signed paper dockets, driver photos, consignment notes with "2 cartons damaged" scrawled across the bottom.
- Supplier invoices for stock, freight, materials, subcontractors and hire equipment, often referencing a purchase order number, sometimes not.
- Timesheets on paper, in a photo from site, or in a spreadsheet a supervisor emails on Monday morning.
- Certificates of currency and other compliance documents from subcontractors, each with an expiry date someone is meant to track.
Each one carries information another part of the business needs: the order must become a sales order, the signed docket must trigger an invoice, the timesheet must reach payroll and the job cost. In many businesses, a person reads the document and types it in, sometimes twice.
What AI reading does well today
On typed, reasonably clean documents, today's document-reading tools handle three jobs well.
1. Pulling out the fields. Customer name, order number, delivery address, line items, quantities, unit prices, dates, ABN, GST, totals. On a supplier invoice, the supplier, invoice number, PO reference and lines. On a POD, the consignment number, who signed, when, and any notes about shortages or damage. On a certificate of currency, the insurer, policy type, cover amount and expiry date.
2. Matching it to the right record. The useful part is matching: this PDF is an order from that customer; this docket belongs to Tuesday's delivery run; this invoice is against a particular PO and job. A good setup looks up your own records rather than guessing.
3. Flagging what doesn't line up. This is where the time goes when people do it by hand. For example:
- The customer ordered 40 units at last month's price, but your price list changed.
- The POD says 18 cartons delivered against 20 dispatched.
- The supplier invoice charges for 12 hours of hire when the docket says 8.
- The timesheet shows a worker on two sites at once.
- A subcontractor's certificate of currency expires next week.
The AI does the first pass and shows a person the document and the matching record side by side. Clean matches go through with a quick review.
Where it struggles
Being honest about the limits saves a lot of disappointment.
- Handwriting. Neatly printed numbers are often fine. A driver's scrawl on a rain-soaked docket, or a supervisor's ticks and crossings-out on a timesheet, is still unreliable.
- Poor scans and photos. Crumpled paper, shadows, a docket photographed at an angle on the tray of a ute. Accuracy drops quickly.
- Unusual layouts. A customer who orders by free-text email ("same as last week but double the 20mm") or a supplier with a strange multi-page layout needs more setup and more checking.
- Messy master data. This is the big one, and it isn't an AI problem. Master data is your core lists: customers, suppliers, products, jobs, sites. If the same supplier exists three times under slightly different names, product codes on the customer's PO don't match your codes, or job numbers are written differently on site than in the office, the matching step fails. The AI reads the document perfectly and then can't work out where it belongs.
Tidying the master data and agreeing how documents are referenced often does more than any choice of tool.
Why a person still checks
AI reading makes mistakes confidently: a 7 read as a 1, or an invoice matched to the wrong one of two similar orders. That's fine when the AI prepares and a person approves. It's not fine if a misread order ships the wrong stock or a supplier gets paid twice.
So the AI reads, matches and flags; a person confirms the order before release, approves the bill before payment and signs off the timesheet before payroll. Their time goes on the exceptions, not the documents that match.
The real value: information entered once
Here's the pattern most businesses miss. The problem was never really "reading documents is slow". It's that the same information gets typed in more than once, and then stops before it reaches the place that needs it.
In a typical wholesale or transport week, a customer emails a PO and someone types it into the order system. The driver brings back a signed docket, and someone types the delivered quantities into the invoice. The supplier's bill for the same stock arrives, and someone else types that into the accounts and checks it against a printed PO. That's double data entry (the same information typed into two or more systems) at every step, and each re-key is a chance for a mistake.
When it's fixed, it looks different. On Monday morning, the weekend's emailed orders are already sitting in the stock system as draft sales orders, each one flagged if a price or product code doesn't match. Signed PODs from last week's runs have already turned into invoices, apart from three with shortages that are waiting for the account manager. Supplier bills are matched to the purchase order and the goods received, so the goods received not invoiced list is short and believable. Nobody has typed any of it in. People have checked it.
That's also why your figures get closer to real time: when dockets reach job costing the day they're signed, you can see what a job has cost this week, not after month-end.
Do you need anything beyond what you already have?
Be honest with yourself here, because it might be a setting rather than a project.
For supplier bills going into your accounts, a standard product may be enough. Most common accounting systems, Xero and MYOB among them, have bill capture built in or through well-known capture apps that read the supplier, date, amount and GST and post a draft bill. If typing supplier invoices into the accounts is your main pain, try those first.
The gap is documents that need to reach your operational systems. Bill capture apps are built to feed the accounts. They generally don't:
- turn a customer's emailed PO into a sales order in your stock or ERP system
- push signed dockets into job costing, or PODs into invoicing (see our guide to getting from proof of delivery to invoice)
- match a supplier bill against goods received in a warehouse system, not just the accounts
- send timesheet hours to both payroll and the job cost
- track certificate expiry dates against the subcontractors booked on next week's jobs
That's where AI reading has to be connected to your other systems, with your own matching rules and an approval step: less about the AI, more about where each document's information needs to end up and who checks it on the way. See our page on manual data entry.
How to start
- List the documents that arrive in a month, who handles them, and which system each gets typed into.
- Follow the most painful one all the way, such as customer POs or PODs, and note every place it gets keyed in.
- Fix the references: job or run numbers on every docket, PO numbers on every supplier bill, no duplicate suppliers or customers.
- Use what you have first, such as the bill capture in your accounting system.
- Pilot one flow with a person approving every result until you trust where it's right and where it isn't.
Our business process improvement guide has more on mapping a flow like this before you change anything.
When it's worth getting help
If your documents only need to reach the accounting system, a capture app and tidy supplier records will probably do.
It's worth getting help when documents have to land in operational systems, when the same paperwork is keyed into two or three places, or when matching fails because the underlying records don't agree. That's integration and workflow design, with AI doing the first pass. Our AI workflow automation and document workflow automation and integration services cover exactly this. If you'd like the whole picture mapped first, our $5,000 Process & Systems Audit traces how documents and information move through your business and produces a written report on what to fix and in what order.
Frequently asked questions
Can AI read handwritten delivery dockets?
Sometimes. Neatly printed numbers and signatures are usually picked up; rushed handwriting, smudges and photos taken at an angle are not reliable. Treat handwritten dockets as "AI suggests, a person confirms", or move drivers to an electronic POD where you can.
Is AI invoice capture accurate enough to pay suppliers automatically?
Not without a check. Capture is good at reading typed invoices, but it can misread figures or match to the wrong order. Use it to prepare and match the bill, and keep a person approving payment, especially for new suppliers or anything that doesn't match the PO.
What's the difference between bill capture in my accounting system and AI document processing?
Bill capture is built to post supplier bills into the accounts. AI document processing connected to your other systems can also handle customer orders, dockets, PODs, timesheets and compliance documents, and send the information to your stock, job or payroll systems as well as the accounts.
Do we need to clean up our data before using AI to read documents?
Usually, yes. If suppliers, customers, products or job numbers are duplicated or inconsistent, the AI can read a document correctly and still fail to match it. Tidying those core lists is often the most valuable first step.
If you're spending your week typing dockets, orders and invoices into more than one system and want to talk through what's worth automating, book a free initial chat.