AI optical character recognition turns images of text into machine-readable text with deep learning models trained on large sets of examples. Classic OCR compares each character against stored shapes or hand-made rules. AI OCR reads whole words and lines in context, so it copes far better with handwriting, poor scans and unusual fonts.
Picture the order desk at a distributor on a Monday. A scanned purchase order arrives slightly skewed, with a coffee ring over the delivery date. Another customer sends a phone photo of a handwritten order form. A third sends a crisp scan of a printed form. Classic OCR reads the crisp scan and stumbles on the other two, so someone retypes them. At Lleverage we think text recognition is now mostly solved. The work that remains is deciding what each value means and whether it can be trusted.
We build AI agents that read customer orders and supplier documents and write the result into the ERP. Recognising the characters is the first step of that job. Matching, checking and creating the sales order come after it. The order management agent shows how that runs end to end, and you can book a demo to see it read your own documents.
What is AI optical character recognition?
AI optical character recognition, also called artificial intelligence OCR, is OCR built on models that learn what text looks like from examples, where classic engines follow rules a programmer wrote. The model takes an image of a page and finds the regions that hold text. For each line it predicts the most likely sequence of characters and gives a confidence score.
Vendors use the name AI OCR for 3 kinds of technology:
- Line recognition models read a full line of text at once. The open-source Tesseract engine moved to this approach in version 4.0. That release added an engine based on LSTM networks, a type of deep learning model, with language data for more than 100 languages.
- Document OCR services also detect paragraphs, lines and handwriting. Microsoft's Document Intelligence Read model, for example, extracts printed and handwritten text from PDFs and scanned images and returns a confidence value for every word.
- Vision language models look at a page image and answer in text. They can transcribe a page and pull out fields in one pass, and they are also prone to filling gaps with plausible guesses.
Each of them turns an image into text, the same job classic OCR does. They differ in how much context the model uses to decide what a smudged character is.
How does classic OCR read a page?
Classic OCR cleans the image and splits it into lines, words and single characters. It then identifies each character on its own. Some engines compare the character's pixels against stored templates of every letter in known fonts. Others measure features such as loops, strokes and line crossings and look up the closest match.
That works well on clean, machine-printed pages in a standard font. It breaks down in predictable places. When 2 letters run together in a poor scan, the splitting step fails before recognition starts. Handwriting has no stored shape to compare against, because no 2 people write an "a" the same way. A character in an unusual font or at low resolution gets the nearest wrong template, such as a 0 read as an O. Many classic engines also read strictly left to right and top to bottom, so 2 columns or a table can come out as one jumbled stream.
Because each character is judged alone, classic OCR has no sense that "lnvoice" is almost certainly "Invoice". Many systems added dictionaries and correction rules. Those helped on common words and did little for article numbers, postcodes and order references, which are the values a back office needs most.
How does AI OCR differ from classic OCR?
AI OCR differs from classic OCR in what it treats as the unit of reading. Classic OCR recognises characters one by one against stored shapes. AI OCR reads a whole line or region at once. It uses the surrounding characters, and often the layout of the page, to decide what each character is.
For a back office the differences look like this:
| Classic OCR | AI OCR | |
|---|---|---|
| How it recognises text | Character by character, against templates or hand-made features | Whole lines or regions, with a model trained on examples |
| Handwriting | Limited, mostly block capitals written in boxes | Readable in supported languages, with a confidence score |
| Skewed scans and phone photos | Needs heavy clean-up first | Copes with tilt, shadow and low contrast far better |
| Tables and columns | Often read as one stream of text | Paragraphs and lines are detected and kept apart |
| Confidence | Per character, if at all | Per word or line, usable for routing |
| New fonts and layouts | Needs retraining or new rules | Usually read on first sight |
| Typical failure | A wrong character or obvious garbage | A plausible but wrong value |
The last row changes how you check the output. A classic OCR error usually looks broken, so a person spots it. An AI model's error often looks right. Its training pushes it towards text that is likely, so a faded 8 in a quantity field can come back as a confident 3.
Can AI OCR read handwriting and poor scans?
Yes, within limits. AI OCR reads clear handwriting in the languages its vendor supports. It also handles skewed scans, phone photos and faxes much better than classic engines. Accuracy still falls on cramped, faded or overlapping writing, and every vendor sets minimums for image quality and text size.
Microsoft's documentation for its Read model shows the kind of limits to expect. It asks for one clear photo or high-quality scan per document. Its minimum text height is 12 pixels on a 1024 by 768 pixel image, which it equates to about 8-point text at 150 dots per inch.
In a manufacturing or wholesale business, handwriting turns up on delivery notes corrected at the dock, on order forms filled in at a trade counter and on printed purchase orders where someone changed a quantity by pen. Those corrections are often the most important values on the page, because they are the ones that differ from what the system expects.
Where does AI OCR still go wrong?
AI OCR goes wrong mostly on degraded input: blurred photos, faint stamps, folded paper and values half hidden by a signature. On those pages a model can produce fluent text that is wrong, because it fills the unreadable part with what usually appears there. The safeguards are a confidence score on every value and a check against the source.
Researchers presenting at NeurIPS 2025 built what they call the first benchmark for OCR hallucination in degraded documents, using identity cards and invoices with simulated real-world damage. Their research found that multimodal models lean on language patterns and produce hallucinated content when the image does not support a precise answer. Their fix was to train the model to decline to answer when a region is unclear.
A back office using AI optical character recognition needs the same behaviour from its set-up. Each extracted field should carry a confidence level, with your team setting the line between work that goes through and work that goes to a person. Values should be checked against what the ERP already knows, so an article number that does not exist, a price far from the price list or a delivery date in the past gets flagged however confident the reading. The person reviewing should see the value and the place on the page it came from, side by side.
In Lleverage, an agent that is unsure of a reading sends it to a person instead of guessing. The review screen shows the request beside the source PDF, and hovering over a field highlights the matching value in the document. That highlighting works on searchable PDFs. On a scan that is only an image, the person compares by eye.
Is AI OCR the same as intelligent document processing?
No. AI OCR is one step inside intelligent document processing. OCR turns the image into text. Document processing then classifies the document and extracts named fields such as the order number and delivery address. It validates them against your records and sends the result to the system that needs it.
A tool with excellent text recognition can still leave your team mapping "Art.nr." to the right field and finding the customer in the ERP. Our guide to intelligent document processing covers the full pipeline. The post on what document intelligence means explains where layout models sit between the two.
What does AI OCR change in a back office?
AI OCR changes which documents can be automated at all. With classic OCR, scans, photos and handwritten orders went straight to a person. With AI optical character recognition they can enter the same flow as a clean PDF, and a person sees only the values the system is unsure of.
Text analysis is already a common way for European companies to use AI. According to Eurostat's 2025 survey of enterprises with 10 or more employees, 19.95% of EU enterprises used at least one AI technology. In the same Eurostat survey, technologies that analyse written language were used slightly more often than any other kind, by 11.75% of enterprises.
Koninklijke Dekker, a Dutch timber company, receives orders as PDFs, Excel sheets and plain emails, each a little different. Its workflow reads all of them and extracts the order lines and customer details. It checks products, availability and pricing, then creates the order in Microsoft Dynamics 365 Business Central. Only low-confidence exceptions go to a person.
"The problem is, you have a lot of Excel sheets, PDFs, or just text emails coming in with an order. This requires a lot of interpretation from our internal sales team." Mart, Continuous Improvement Team, Koninklijke Dekker
At Microtechniek, a mechanical engineering firm in Velsen-Noord, each customer purchase order used to take about 10 minutes by hand. Someone opened the PDF, identified the customer, checked the contact and keyed more than 20 fields into Ridder iQ, an on-premise ERP. In August 2026, Microtechniek had between 45 and 71 purchase orders a week written into the ERP without anyone typing them. The desk got more than 500 hours a year back.
At both companies, much of the work sat in interpretation after the text was read: which customer sent this, which article they mean, whether the order follows an earlier quote. Our post on document data extraction into the ERP walks through that step in more detail.
How do you test AI optical character recognition on your own documents?
Test AI optical character recognition on the documents that fail today. Collect 50 to 100 real pages, including scans, phone photos and handwritten corrections. Then measure how many fields come out right, how many are flagged as uncertain and how many are wrong while marked as confident.
A short test plan:
- Pick the hard pages: your worst supplier layouts, faxes, multi-page orders and anything with handwriting or stamps.
- Key the expected values for each page yourself before the test, so the comparison is fair.
- Count the fields that are correct and confident, the ones flagged for review, and the ones that are wrong but confident, which cost the most because nobody checks them.
- Look hard at the line items. Quantities, units and article numbers on line 14 of a 3-page order separate the tools far more than header fields do.
- Follow each value into the ERP and check that it lands in the right field on the right record.
If a tool cannot tell you when it is unsure, keep a person on every document, whatever its accuracy figure says. Lleverage's AI agents score their own work, and your team decides where the line sits.
Frequently asked questions
What is the difference between OCR and AI OCR?
Classic OCR recognises characters one at a time by comparing their shapes against stored templates or hand-made features. AI OCR uses deep learning models trained on examples to read whole words and lines in context. That makes AI OCR better at handwriting, poor scans and unusual fonts, though it can return a plausible but wrong value.
Is AI optical character recognition accurate enough to skip human review?
On clean printed documents, often yes for most fields. On scans, photos and handwriting it depends on image quality and on what the value is worth. A safe set-up routes every low-confidence value, and every value that conflicts with your ERP data, to a person. The rest goes through on its own.
Can AI OCR read handwritten order forms?
AI OCR can read clear handwriting in the languages its vendor supports, and some services mark handwritten lines with a confidence score. Cramped, faded or overlapping writing still causes errors. For order forms, test with real examples from your trade counter or warehouse before you trust the readings on quantities and article numbers.
Does AI OCR replace intelligent document processing?
No. AI OCR is the text recognition step inside intelligent document processing. Processing adds classification, field extraction, validation against your records and the hand-off to your ERP. Better OCR gives those steps cleaner input, and a team still needs them to turn a page of text into a correct sales order.
Start with the documents your current OCR gets wrong
If scans, phone photos or handwritten corrections still go to a person on your order or purchasing desk, start with that pile. Bring 20 of those pages to a demo and we will show what the agent reads, what it flags for a person and how the result reaches your ERP. Book a demo to set it up.
