Every business has someone typing information from documents into a system. Invoices, delivery notes, CNICs, bank statements, purchase orders. It is slow, error prone and nobody enjoys it, which makes it an obvious automation target.
It is also where automation projects most often disappoint, for a reason that is easy to understand in advance.
Why accuracy claims mislead
A vendor says ninety five percent accuracy. That sounds excellent until you work out what it means.
An invoice has around twenty fields. At ninety five percent per field, the probability that an entire invoice is correct is about thirty six percent. So two thirds of documents contain at least one error somewhere, and you do not know which two thirds.
This is why systems designed to run without review fail. The correct design is not higher accuracy. It is a fast review step, plus confidence scores so the system tells you which fields it is unsure about.
The four document types worth automating
1. Supplier invoices
The highest volume repetitive document in most businesses. Extract supplier, date, invoice number, line items, tax and total, then match against the purchase order.
The value is not only typing time. Automatic three way matching between purchase order, goods received note and invoice catches overbilling that manual processing misses, and that catch frequently pays for the system.
2. Delivery notes and goods received
Particularly valuable where documents arrive as photographs from the warehouse. Extract items and quantities, match to the purchase order, flag differences immediately rather than at month end.
3. Identity documents
CNIC front and back for customer onboarding, tenancy, account opening or KYC. Structured, consistent layout, well suited to extraction.
Handle this carefully: identity documents are sensitive personal data. Decide where images are stored, who can see them, how long they are kept, and whether they need to be kept at all once the fields are extracted.
4. Bank statements
For reconciliation and for lending decisions. PDFs from different banks in different layouts, converted into transaction rows. Tedious by hand, well solved by automation, and the output is immediately checkable against a closing balance, which makes verification cheap.
What is not worth automating
- Low volume documents. Under a few hundred a month, the setup and review effort exceeds the saving.
- Highly variable documents where every one has a different structure and no two suppliers agree.
- Handwritten forms, especially in Urdu. Accuracy is poor enough that review costs as much as typing.
- Documents requiring judgement. Contracts where the question is whether a clause is acceptable. Extraction can find the clause; it cannot decide.
- Poor quality photographs. Fix the capture process first. Better lighting and a document scanning app improve results more than any model change.
The design that actually works
- Capture, ideally through an app that guides the photograph rather than accepting anything.
- Extract, with a confidence score per field.
- Auto approve documents where every field is high confidence and the totals reconcile.
- Route the rest to a review screen showing the document image beside the extracted fields, with low confidence fields highlighted.
- Post to your system, with the original image attached to the record.
Step four is the whole design. A good review screen means a person confirms a document in five to ten seconds instead of typing it in three minutes. That is where the saving is, and it survives imperfect extraction.
Step three is what improves over time: as you learn which suppliers and document types are reliably correct, more flows through without review.
Realistic expectations
- Clean printed documents, consistent layout: most fields extracted correctly, review in seconds.
- Printed documents, varied layouts: good extraction of key fields, line items less reliable.
- Photographs from phones: workable with good capture, poor with bad capture.
- Handwritten: assume it needs typing.
- Urdu text: substantially weaker than English. Test with your actual documents rather than trusting a demo.
Cost and payback
- Off the shelf document AI services: usually priced per page, in the range of a few rupees per page at volume.
- Integration into your systems: 400,000 to 1,500,000 PKR depending on how many document types and which systems it posts into.
- Running cost: per page charges plus maintenance.
The payback calculation is straightforward: documents per month, minutes per document, cost of that person’s time. A business processing two thousand invoices a month at three minutes each is spending a hundred hours a month on typing. That is the number to compare against.
For the wider question of what AI features cost to run, see this breakdown, and for where AI genuinely absorbs operational work, our sober map.
Frequently asked questions
How accurate is document extraction?
High per field on clean printed documents, which still means most multi field documents contain at least one uncertain field. Design for review rather than for perfection.
Does it work with Urdu documents?
Noticeably less well than English. Test with your own documents before committing, and expect more review.
What is the minimum volume to justify it?
Roughly a few hundred documents a month of a consistent type. Below that, the configuration and review overhead outweighs the saving.
Can it replace the data entry person entirely?
Rarely, and it should not. It converts typing into checking, which is faster and less error prone. The role shifts rather than disappears.
Ezitech builds document processing and workflow automation connected to ERP, accounting and inventory systems. See our AI solutions or tell us what you are typing by hand.