Definition
Computer vision turns visual input into structured information a system can act on: reading a document, counting objects, detecting a defect, verifying that a form was filled in correctly, or confirming that something is where it should be.
What computer vision does in an operating business
Most business computer vision is far less exotic than the demos suggest. It is usually one of four jobs: reading something, counting something, checking something, or confirming something is present.
The reading case is by far the most common. Invoices, delivery slips, forms, identity documents, and handwritten notes arrive as images, and someone re-keys them. Vision models turn that into extracted fields with a confidence score, and the human effort shifts from typing everything to reviewing the uncertain ones.
The checking case is the second: comparing what is in front of the camera to what should be there. A part with a visible defect, a pallet loaded incorrectly, a safety step skipped, a shelf that is empty.
Computer vision vs. OCR
They get used interchangeably, and the distinction matters when you are buying.
Traditional OCR converts pixels to characters. It works well on clean, typed, predictable layouts and degrades quickly on anything else: a photographed invoice at an angle, a form with handwriting in the margin, a layout the vendor changed last quarter.
Modern vision models read the document more like a person does: they handle imperfect images, understand that this number is the total because of where it sits and what it is labelled, and cope with layouts they have not seen. They are more capable and more expensive per page, and because they are probabilistic they need confidence thresholds and a review path rather than a pass/fail result.
For a stable, high-volume, identical form, OCR is often still the better economics. For variable documents from many sources, vision models are usually the only thing that works.
What it takes to work reliably
The gap between a promising pilot and a dependable system is almost always the input rather than the model. Consistent lighting, angle, and resolution matter more than model choice. So does knowing what the system should do when it is unsure, because a vision system without a confidence threshold and a human queue is a system that will silently guess.
And as with any AI, it needs evaluation against real examples, including the bad ones: the crumpled invoice, the glare, the form someone filled in with a different pen.
How Automathing approaches it
We scope vision projects around the review step, not the accuracy claim. A model that is 94% accurate is useful when the 6% is flagged and routed to a person, and dangerous when it is not. That means designing the confidence threshold, the exception queue, and the audit trail as part of the build rather than after it.
Frequently asked questions
What is the difference between computer vision and OCR?
OCR converts images of text into characters, and works best on clean, consistent layouts. Computer vision interprets visual content more broadly, understanding structure, context, objects, and scenes, including in documents it has not seen before. Document AI built on vision models is effectively OCR plus comprehension.
Can computer vision read handwriting?
Modern models handle clear handwriting reasonably well and struggle with genuinely messy writing, much as a person does. The practical design is to treat handwriting as lower-confidence input by default: extract it, flag it, and route it for a quick human check rather than accepting it silently.
How accurate does it need to be?
That depends entirely on what a mistake costs and whether you catch it. For invoice extraction where a person reviews flagged items, high-nineties accuracy with reliable confidence scoring is genuinely useful. For anything unreviewed and consequential, the question is not the accuracy number but what happens on the cases it gets wrong.
Do we need special cameras or hardware?
Usually not for document work, where a phone camera or an existing scanner is generally sufficient. Physical inspection is different: consistent lighting and camera placement typically matter more to the result than the model does, and that is where the setup cost tends to sit.