AI-generated article. This article was researched and drafted using AI tools and published automatically, and its featured image was generated by AI. Facts are drawn from the sources cited in the text.
Somewhere in most small businesses there is a pile. Supplier invoices in four different layouts, a contract someone scanned at an angle, delivery notes with handwriting on them, receipts in a shoebox or a photo roll. Somebody types all of it into a system by hand, usually at the end of the week, usually resenting it. That job is the one AI document processing has quietly gotten good at.
Not good in the way software demos are good. Good in the boring, measurable sense. Analysts at Parseur describe 2026 as an inflection point for vision-based document processing, on three fronts at once: accuracy, cost and how long it takes to set up.
What changed in AI document processing
Older optical character recognition read shapes and guessed at letters. It broke on unusual layouts, merged table cells and anything handwritten. Current vision language models read a document more like a person does, taking the layout as part of the meaning. Reported extraction accuracy on modern platforms sits above 95% on ordinary business documents. Hyperscience has reported fine-tuned models reaching around 99% on invoices and identity documents in production, though that figure comes with a condition worth holding onto: it applies to workflows with a human checking the output.
The cost shift is larger than the accuracy shift. Vendors in this category report document processing costs falling somewhere between 75% and 92% against manual entry or legacy OCR systems. The setup economics moved even further. A system of this kind used to mean something like six months of integration work and a six-figure budget, which put it firmly out of reach for a business with nine employees. Current write-ups describe comparable extraction being stood up in days for a few hundred dollars. One guide aimed at small and mid-sized firms puts the saving on accounts payable processing time at around 80%.
What multimodal actually means here
The underlying change is that the leading models stopped being text-only. A multimodal model takes text, images, audio and video together and reasons across them in one pass. In practice that means a photograph of a damaged delivery, the note that came with it and the original order can be assessed as one thing rather than three. Gemini’s larger context window is repeatedly singled out for document work specifically, because a long contract fits in one go instead of being chopped into fragments that lose track of each other.
This is the same broader movement that has taken AI into the software a business already runs on, rather than sitting in a separate chat window waiting to be visited.
Where it still goes wrong
An error in a drafted email is embarrassing. An error in a payables run is money. That asymmetry is the whole argument for keeping a person in the loop, and it is why the 99% figure and the human review condition belong in the same sentence. A system that is right 95% of the time on a hundred invoices a month is wrong on five of them, and the five will not announce themselves.
The sensible pattern is narrow and checkable. Automate the extraction, keep the approval. Route anything the model flags as low confidence to a person. Spot check a sample of what it was confident about, because confident and correct are not the same thing.
A small way to find out
Take twenty documents that already went through by hand, where the correct answer is known. Run them through one tool. Compare. That takes an afternoon and produces something better than a vendor’s accuracy claim, which is a number from this business, on these documents, in these formats.
Twenty is enough to see the shape of the errors. If the failures cluster somewhere specific, one supplier’s layout, one language, handwriting, that is a fixable boundary rather than a verdict. If they are scattered randomly, that is worth knowing before anything gets connected to the accounting system. Either way the decision to build or buy gets easier once there is real data behind it.
The pile is not going anywhere on its own. But the question has shifted from whether a machine can read it to whether anyone has checked how well it reads yours.
