Accounts payable is one of the few areas where automation has an unambiguous case: the work is high volume, repetitive, rule based, and expensive in staff hours. It is also the area where projects most often underdeliver, because the pitch focuses on the wrong half of the problem.
Extracting text from an invoice is largely solved. Deciding whether that invoice should be paid is not.
What the work actually consists of
For a business receiving a few hundred supplier invoices a month, the time splits roughly like this:
- Data entry from the invoice: 20%. The part automation is sold on.
- Matching to purchase order and delivery note: 35%. The real work.
- Chasing discrepancies: 25%. Quantity received differs from quantity billed, price differs from the agreed rate, a partial delivery.
- Routing for approval and following it up: 20%. Mostly waiting for somebody to respond.
Automating only the first item removes a fifth of the effort. Automating the first, second and fourth removes most of it.
What can be automated well
Extraction
Supplier name, invoice number, date, line items, quantities, rates, tax and total, from PDFs, scans and photographs. Modern document AI handles varied layouts without a template per supplier, which is the main advance over older OCR.
Realistic accuracy on clean documents is high; on a creased photograph taken in poor light, less so. Assume some invoices will always need a human.
Three way matching
Invoice against purchase order against goods receipt. Where all three agree within tolerance, approve automatically. In most businesses a large majority of invoices match cleanly and never need a person to look at them.
This is where the return is. The rule to define carefully: what tolerance is acceptable, in both percentage and absolute terms.
Duplicate detection
The same invoice submitted twice, or a supplier statement re entered. Paying an invoice twice is common and embarrassing, and this catches it reliably.
Routing and reminders
Approval by amount, by department, by cost centre, with reminders when somebody sits on it. Unglamorous and it removes most of the delay.
What still needs a person
- Exceptions. Anything that does not match. By definition these need judgement.
- New suppliers, where bank details are being set up. This is the single highest fraud risk in accounts payable and should never be automated.
- Invoices with no purchase order. Services, utilities, professional fees. Approval depends on whether someone actually authorised the work.
- Anything above a threshold. Set one and keep it.
The fraud point, which deserves emphasis
The most common accounts payable fraud is a request to change a supplier’s bank account, arriving by email from what appears to be the supplier. Automation makes this worse if bank detail changes flow through without friction.
Rule: a bank detail change is never processed from an email or an invoice. It requires a call to a known number, verified against a record held before the request arrived. Build that into the process rather than relying on vigilance.
The Pakistan specific requirements
- Sales tax invoice formats and the details required for input tax claims. Extraction must capture the supplier’s registration number and the tax breakdown correctly.
- FBR digital invoicing where it applies to your business, which changes what arrives and in what form.
- Withholding tax calculated at the correct rate by supplier type and deducted at payment. Get this wrong and the problem surfaces at audit.
- Mixed formats. Printed invoices, handwritten ones from smaller suppliers, PDFs and photographs sent on WhatsApp. Any real deployment here has to cope with all four.
- Urdu and mixed language documents from smaller vendors.
What it costs and what it returns
- Off the shelf document AI: priced per page or per document processed. Reasonable for moderate volume.
- Integration into your accounting or ERP system: 600,000 to 2,000,000 PKR depending on the system and how much matching logic is involved.
- Custom build with matching and approval workflow: 1,500,000 to 4,000,000.
The return is straightforward to calculate: invoices per month, minutes per invoice today, cost of that time. A business processing five hundred invoices a month at fifteen minutes each is spending more than a full time salary on it.
The honest threshold: below roughly two hundred invoices a month, the payback is slow and a better structured manual process may serve you better. Related reading: document processing automation and what running an AI feature costs.
Frequently asked questions
How accurate is AI invoice extraction?
High on clean digital documents, lower on poor photographs and handwritten invoices. Plan for a review queue rather than expecting every document to pass straight through.
Can it handle handwritten invoices?
Partially, and less reliably than printed ones. For suppliers who issue handwritten invoices, expect manual entry or push them toward a printed format.
Will it integrate with our accounting software?
Usually, through an API or file import. Ask about your specific package by name, and about how it handles tax fields, which is where generic integrations most often fall short.
What is the minimum volume to justify it?
Roughly two hundred supplier invoices a month. Below that, fixing the process and the approval routing generally gives a better return than software.
Ezitech builds document automation and finance workflow systems connected to real accounting platforms. See our AI solutions or tell us your monthly invoice volume.