PTAS AI · Blog

What is OCR in finance, and where does it quietly stop?

OCR turns a scanned invoice into text. That is genuinely useful, and it is also where most finance teams overestimate what they are buying. Here is what the technology reads, where it hands the hard part back to a person, and what an agent does about that.

A finance lead asked me last month whether OCR would clear her invoice backlog. Fair question. She had three people keying numbers off PDFs all day and the queue kept growing. The honest answer was that OCR would solve maybe a third of her problem and quietly create a new one.

So before you sign anything, it helps to know exactly what OCR is, what it reads, and the precise point where it hands the difficult part back to a human. That handoff is where most finance teams lose the time they thought they were saving.

What OCR actually does

OCR is short for optical character recognition. Give it a scanned page, a photo of a receipt, or a PDF that started life as paper, and it returns the text it can see. That is the whole job. It reads characters. It does not know the page is an invoice, that the figure in the bottom right is a total, or that the total should equal the sum of the lines above it.

The pipeline hiding inside one word

What gets sold as a single feature is really a short pipeline. Three steps inside it matter for finance documents.

Cleaning up the image

A scanner rarely hands you a clean page. There is skew, a shadow down one edge, a coffee ring, a staple hole punched through the date. The first step straightens and de-noises the image so the next step has a fighting chance.

Why scans fight back

Vendor mail is about the worst input you will ever feed software. I have watched the same invoice get faxed, printed, scanned, photographed at an angle on a warehouse floor, then emailed as a 90 KB JPEG. Every one of those steps throws away detail the next stage needs.

Recognising the characters

With a cleaner image, the engine matches shapes to letters and digits. Modern engines lean on models trained on huge samples, so they cope with odd fonts and even some handwriting. This is the step vendors quote their accuracy on, and it is also the step that flatters them most.

Laying the text back out

Reading characters is not enough. The engine also has to remember where they sat on the page, because in finance the relationship between a label and a number is the entire point. “Total 4,820” means nothing if the engine returns “Total” and “4,820” as two unrelated strings.

Where OCR quietly breaks in a finance workflow

Here is the part the demos skip. OCR reads the page. It does not check the page. So the moment your document is real instead of a clean sample, the gaps open up.

An invoice number reads correctly but lands in the purchase-order field. A 7 reads as a 1 on a faded thermal receipt and nobody notices until the payment run. The line items extract cleanly but do not add up to the stated total, and OCR has no opinion about that, because arithmetic was never its job. Each of those becomes a human task again, which is exactly the task you were trying to delete.

What finance teams actually need

Once you see where OCR stops, the requirement gets clearer. You do not want a faster typist. You want something that treats the document the way a good AP clerk does: read it, sanity-check it, match it against what you already know, and only ask for help when it is genuinely unsure. That is what we build as agentic document extraction, and it is the difference between text on a screen and a posted entry you can trust.

Read the document, not the pixels

An agent knows an invoice is an invoice. It finds the total even when every vendor parks it somewhere different, and it labels each field instead of returning loose text.

Check the math before you do

Line items should sum to the subtotal. Tax should match the rate. The total should reconcile. An agent runs those checks the second it reads the page, not three steps later inside your ERP.

Match it to what you know

A number is only useful in context. The agent matches the invoice to its purchase order, the vendor to your master, the GST number to the register, and flags whatever does not line up.

Escalate, do not guess

When confidence is low, a sensible agent stops and routes the document to a person with the reason attached. Silent guessing is how you pay a duplicate invoice twice.

The numbers we actually see

People want a single accuracy figure. I can give you ranges from our own projects, but only with a warning attached to them.

92–99% character accuracy on clean, printed PDFs, the figure most vendors quote
60–75% end-to-end field accuracy on real, messy vendor mail before any validation
3–6 wks typical time to a working AP pilot running on a client’s own documents
~20 sample documents we ask for to give you a real, field-level accuracy read

Read these as ranges, not benchmarks. They come from PTAS AI client projects, not a published industry study. Where your documents land on each range depends on scan quality, how consistent your vendors are, and how much validation sits behind the extraction.

The gap between the first row and the second is the whole reason agentic extraction exists. Character accuracy is easy to win and easy to brag about. Field accuracy on the mail you actually receive is the number that decides your team’s workload.

When plain OCR is still the right call

I am not anti-OCR. It is the foundation everything else sits on, and for some jobs it is all you need.

If you are digitising an archive so it becomes searchable, or pulling text out of consistent, high-quality scans where a person will read the result anyway, plain OCR is fine and cheaper. The trouble starts when the output feeds a decision or a payment with nobody in the loop. That is the line. Below it, OCR is enough. Above it, you want an agent.

If you want the longer argument, I made the case for why OCR is the wrong frame for document automation, and you can watch the agent approach play out in a real accounts payable queue. The product itself lives on our agentic document extraction page.

Common questions about OCR in finance

Is OCR the same as intelligent document processing?

No. OCR reads characters off a page. Intelligent document processing, and the agentic version we build, sits on top of OCR and adds understanding, validation, and matching. OCR is one layer of a larger stack, not the stack.

How accurate is OCR for invoices?

Character accuracy on clean invoices is high, often above 95%. Field-level accuracy on real, varied vendor mail is lower, frequently in the 60 to 75% range before validation. The second figure is the one that matters, because it decides how much work reaches your team.

Can OCR replace my accounts payable team?

On its own, no. OCR removes some keying, then pushes every exception back to a person. An agentic system removes far more of the manual work because it validates and matches, and only escalates the documents that genuinely need judgement.

Bring us your messiest invoices.

Send PTAS AI 20 of your real documents, scans, phone photos, and the faxed-then-scanned ones included. We’ll run them through DocPro and send back field-level accuracy with confidence scores, so you can see exactly where OCR stops and where an agent takes over.