In short
- Reading a document is the cheapest part of automating it: AWS lists its expense and invoice analysis at $0.01 a page for the first million pages a month.
- Most of the cost is after extraction: matching the fields to your own customers, products and prices, reviewing the documents the system is unsure about, and keeping up as senders change their layouts.
- The confidence threshold decides the saving. AWS suggests 90% or higher for processes involving financial decisions, and every document below the threshold goes to a person.
- For VAT invoices between businesses, the UK has decided that e-invoicing will be mandatory from 2029, and a PDF will not count as an e-invoice, so reading invoice PDFs has a limited shelf life.
Document processing is usually the first automation a business asks about, and for good reason. Someone is retyping purchase orders from emails, keying supplier invoices into the accounts package, or copying figures off delivery notes. The work is repetitive, the documents look alike, and the tools for reading them are cheap and good.
That last point is why the business case tends to look better than it turns out to be. The price on the tool’s website is the price of reading the page. Reading the page is not where the money goes.
Reading the page costs about a cent
The cloud extraction services are priced per page, and the numbers are small. AWS Textract lists its expense analysis, which pulls out supplier, totals and line items from invoices and receipts, at $0.01 a page for the first million pages a month in its US West region. Plain text detection is $0.0015 a page. A business receiving 1,000 invoice pages a month pays about $10 to have them read.
Language models have pushed this further. A general model can now read an order written in the body of an email, with no fixed layout at all. For almost any business, the cost of reading documents is not the constraint. The monthly bill for extraction is small next to the wage of the person doing it by hand.
If that were the whole cost, every business would have automated this years ago. It is not.
Where the cost actually sits
Extraction gives you text in fields: a customer name, a product description, a quantity, a price. None of it is usable until it has been matched to your own records, and that is where the build effort goes.
- Matching. “Bramley apples 10kg” on a customer’s order has to become your product code. “J Smith Plumbing Ltd” has to become account 40213. A wholesaler with a few thousand products and customer-specific price lists needs a lookup that copes with abbreviations, old product names and the customer’s own codes.
- Checking. A price on an order that differs from that customer’s agreed price is either a mistake or a conversation. The system has to know which prices to expect and flag the ones it does not.
- Exceptions. Every document the system is not sure about goes to a person, with the original next to the extracted fields so they can correct it in seconds rather than rekey it.
- Integration. The result has to land in the accounts package, the order system or the stock system, through an import or an API, without creating duplicates when a document arrives twice.
- Upkeep. Suppliers change their invoice layouts, customers change how they write orders, and product lists change every month. Someone has to own the system after the build.
For a typical build, the extraction call is a few lines of code. The four items after it are most of the project, and the last one never ends.
The page costs a cent to read. What you are paying for is everything that has to happen before the answer can be trusted.
The threshold decides the saving
Every extraction service returns a confidence score for each field, and the business has to decide where to draw the line. AWS’s own guidance is that archival uses might accept 50%, but that “business processes involving financial decisions might require thresholds of 90% or higher.” Everything below the line is flagged for a person.
That line is the business case. Take an illustrative wholesaler, not a measured one, receiving 120 emailed orders a day that take three minutes each to key in: six hours of someone’s day. If a strict threshold sends 40% of orders to review, and reviewing a pre-filled order takes one minute, the daily work drops from six hours to under an hour. If the orders are messier and 80% go to review, the saving is much smaller, and the business is paying to maintain a system that still needs most of a person.
Nobody can tell you the review rate before seeing the documents. It depends on how many different senders there are, how consistent they are, and how much of the matching is ambiguous. The only honest way to price the project is to run a sample of real documents through extraction before committing to the build.
When it isn’t worth it
Some cases fail the test before any code is written.
- The volume is low. Fifteen supplier invoices a week at three minutes each is 45 minutes of work. No build and maintenance budget pays that back, and an accounts package with a built-in inbox scanner will do it for the price of a subscription.
- The sender could send data instead. If most orders come from a handful of large customers, a trade ordering portal or a CSV they can export is cheaper and more accurate than reading their PDFs. Automating the reading of a document the other side could have sent as data is solving the wrong problem.
- A mistake is expensive and nothing catches it. If a misread figure goes straight into a payment or a price change with no reconciliation after it, the review rate has to be so high that the saving disappears.
- The document is about to stop existing. The UK government has decided that all VAT invoices must be issued as e-invoices from 2029, covering invoices between businesses and to the public sector. The same response says PDF, Word and image invoices will not count as e-invoices. The technical standard is due in a roadmap at Budget 2026, but the direction is set.
The e-invoicing point deserves weight in any plan made now. A system built this year to read supplier invoice PDFs has perhaps three years of useful life before those invoices arrive as structured data. That can still pay back at high volume. At moderate volume, the better investment is often the matching and approval logic, which is needed either way, fed by whatever format the invoices arrive in.
Orders, delivery notes, remittance advice and certificates are not covered by the mandate. For a wholesaler taking orders by email, the unstructured order is not going away, and that is usually where automation earns the most.
Count before you build
Before choosing a tool, take one month of documents and sort them into three piles: how many there are of each type, how many different senders produced them, and how many could have arrived as data if you had asked. Then run a sample of a hundred or so through an extraction service and see how many would clear a 90% threshold.
Those numbers answer the question. High volume, a manageable number of layouts and a good pass rate is a project worth building properly. Low volume, or senders who could be persuaded to send data, is a cheaper problem with a cheaper answer.
Common questions
How much does it cost to extract data from an invoice with AI?
The extraction itself is cheap. AWS Textract’s expense analysis is listed at $0.01 a page for the first million pages a month. Most of the cost is in matching the results to your own records, reviewing uncertain documents and integrating with your systems.
When is document processing automation not worth it?
When volume is low, when the people sending the documents could send structured data instead, or when a wrong field is expensive and nothing downstream would catch it. At a few documents a day, typing them in is usually cheaper than building and maintaining automation.
Will UK e-invoicing replace invoice OCR?
For VAT invoices between businesses, largely yes. The government has decided that all VAT invoices must be issued as e-invoices from 2029, and PDF, Word and image invoices will not count. Orders, delivery notes and other documents are not covered.
Written by the Voxalone team. If this raised a question about your own systems, get in touch — you will get a straight answer from an engineer, with no obligation attached.