You know the situation: a stack of forms, half printed, half filled in by hand. Scanned contracts from the '90s, where every other page sits crooked. Handwritten notes from the field team that somehow have to make it into the system. And then someone says: "Surely OCR can handle that."
It can, too. Sometimes. For printed text in clean quality, OCR really is a mature technology. But as soon as handwriting comes into play, layouts get complex or scan quality fluctuates, classic OCR reaches its limits. And this is exactly where it gets interesting, because very few people know that besides OCR there are two more disciplines that are often decisive in practice.
Three letter combinations, three different worlds
When we talk about automatic text recognition, acronyms come up quickly. Let us sort out the three most important ones, without academic baggage, but with an eye on practice.
OCR, Optical Character Recognition
What it is: Optical Character Recognition, where "character" here does not mean a person's character, but simply the sign or letter. So we are not recognizing personalities, but writing. Even if with some handwriting you could certainly draw conclusions. Technically speaking: OCR converts pixels into machine-readable text, letter by letter, word by word.
Where it works well: Clean print, standardized forms, invoices, letters, contracts in good scan quality. For these use cases, OCR has been a commodity for years, there are countless providers and the technology is mature.
Where it reaches its limits: As soon as the quality of the input material fluctuates, crooked scans, stains, faded ink, mixed fonts, the recognition rate drops rapidly. And with handwriting, classic OCR is simply overwhelmed.
HTR, Handwritten Text Recognition
What it is: The recognition of handwriting, that is, what OCR cannot do. HTR is a discipline of its own, considerably more demanding than the recognition of print. Everyone writes differently, the variance is enormous, and context-dependent interpretation is mandatory.
Why it is so difficult: Handwriting is not standardized. An "a" can take a hundred different forms. Letters overlap, lines are crooked, sometimes parts of words are missing. While OCR essentially does pattern matching, HTR has to bring genuine language understanding: the context decides whether it says "Haus" or "Hans".
A personal confession at this point: my own handwriting is unspeakable. To recognize my own handwritten notes, it helps me enormously if I know the context. And it does happen that I can no longer decipher my own hieroglyphs when the context is not enough. In that respect I am heartily glad that nowadays we mostly write on the computer, which also makes communication between me and my clients considerably easier. But I digress.
A machine has exactly this problem too. And that is precisely why HTR is so demanding: the system must not only recognize pixels, but understand what might be meant. It needs language models, contextual knowledge and, yes, a certain tolerance for creative letter formation.
Where it is relevant: Public administration (handwritten applications and forms), insurance (handwritten claims reports), healthcare (handwritten physician notes), archives and libraries (historical documents). In all these areas, handwriting is not the exception, it is the norm.
ICR, Intelligent Character Recognition
What it is: At its core, ICR is the intelligent approach that combines the strengths of traditional OCR pipelines, HTR models and modern AI methods, and adds further steps on top. You can think of ICR as the orchestration layer that decides for each document: which areas are print? Which are handwriting? Which model is applied where? And how do we validate the result?
What distinguishes ICR from OCR: Classic OCR works with fixed rules and templates. ICR learns from data, uses ensemble methods, that is, the combination of several models that complement and correct one another, and as a post-processing step can even employ Large Language Models to check recognized text for plausibility and correct errors. The result: a pipeline that is considerably more robust than any single component.
Can ICR do everything? In principle, yes. ICR can be used anywhere OCR is used, and usually delivers better results. For clean standard documents this is overkill and classic OCR is enough. But as soon as reality strikes, and it almost always does, ICR is the approach that makes the difference. It is what companies actually mean when they say "OCR": not just recognizing printed text, but being able to handle the full range of real-world documents.
Where they overlap, where they differ
The three technologies are not competitors, they are tools for different problems that are often combined in practice.
OCR is a commodity. The technology is widely available, prices are low, the quality with clean input is good. For standardized print documents you do not need a specialist.
HTR is highly specialized. Here you need expertise in deep learning, language models and domain-specific training. The providers who can really do this well can be counted on two hands.
ICR bridges the gap, and is at the same time the discipline where the field thins out the most. There are many who call themselves "AI experts" today. As a rule, that means they can operate ChatGPT moderately well, and beyond that it quickly thins out. ICR, however, is considerably more than that: it brings the intelligence and adaptivity of AI models into text recognition and makes it robust enough for reality, mixed layouts, fluctuating quality, heterogeneous document types. That requires a genuine understanding of model architectures, training data and domain knowledge.
In modern pipelines, all three are combined: OCR for the printed text, HTR for the handwritten parts, ICR as the intelligent orchestration layer that decides which method is applied where. The real challenge is never the technology alone, but the client's specific data.
Why this is relevant for your company
Many business processes are still paper-based, or are based on documents that were once on paper. The digitization of these documents is not a technical gimmick, but the basic prerequisite for any further automation.
Public administration: Handwritten forms are legally required in many areas, or simply a reality. Building applications, citizen submissions, handwritten notes on files: all of this has to be captured digitally if administrative processes are to be modernized.
Insurance: Claims reports filled in by hand. Legacy contracts that were scanned 20 years ago. Expert opinions with handwritten additions. None of this is the exception, it is day-to-day business.
Healthcare: Handwritten physician notes, findings, prescriptions. Despite increasing digitization, handwriting in medical documentation is still ubiquitous.
Archives and cultural institutions: Historical documents, parish registers, old correspondence. HTR makes these holdings machine-readable and searchable for the first time, an enormous gain for research and accessibility.
Industry and logistics: Handwritten delivery notes, inspection protocols, maintenance reports. In many operational areas, work is still done with pen and paper, and this data has to be transferred into digital systems.
The legal framework: why administrations in the DACH region cannot avoid it
What many overlook in the debate about digitization: in all three DACH countries, administrations are legally obliged to keep the analog access channel open. Citizens have the right to submit forms on paper, filled in by hand, signed, by post or at the counter. Digitization does not replace the paper route, it adds to it. And that means: someone has to process the analog submissions.
Germany: OZG, the written-form requirement and accessibility
The Online Access Act (OZG) obliges the federal government and the states to offer administrative services digitally. But the OZG does not abolish the paper route, it adds to it. The Administrative Procedure Act (Section 3a VwVfG) continues to guarantee written access to the authority. Citizens have the right to submit applications, objections and forms on paper. On top of this comes the written-form requirement (Section 126 BGB), which demands a handwritten signature for numerous administrative acts.
The consequence: municipalities and state authorities have to accept, digitize and process paper forms, and do so efficiently. The Accessibility Strengthening Act (BFSG) reinforces this even further: not all citizens can or want to use digital forms. Older people, people with disabilities, people without internet access, all of them are entitled to the analog route. So the forms keep coming, whether you like it or not. The only question is whether you retype them manually or process them intelligently.
Austria: the E-Government Act and citizen participation
Austria committed to digitization early on with the E-Government Act (E-GovG). But the General Administrative Procedure Act (AVG, Section 13) is unambiguous: submissions, that is, applications, requests, complaints, can be filed in writing. Authorities have to accept and process this access channel. In practice this means: handwritten forms at the municipal office, building applications on paper, signed objections by post. Austria's administration is digitally ambitious, but legally obliged to keep the analog application route open.
Switzerland: EMBAG and the federalism problem
In 2023, Switzerland passed the Federal Act on the Use of Electronic Means for the Fulfillment of Government Tasks (EMBAG). But here too the Federal Act on Administrative Procedure (VwVG, Art. 21) applies: submissions to authorities can be made in writing. The authority must accept and process them. With 26 cantons at very different levels of digitization, this is no theory: what has long been submitted digitally in Zurich still arrives in many municipalities as a handwritten paper form over the counter or by post.
The pattern is the same in all three countries: the analog access channel to administrative services is legally guaranteed. Applications, objections, forms, citizens are allowed to submit them on paper, and the authority has to process them. This generates a permanent stream of handwritten documents that will not let up as long as these laws apply. The question is not whether, but how efficiently these documents are digitized.
"Digitizing administration does not mean paper disappears. It means we finally deal with it intelligently, and turn analog submissions into digital processes.", Thorsten Vellmerk
Measuring quality: CER and WER
When you talk to providers about text recognition, you will hear two metrics again and again: CER and WER. What do they mean, and what do they really tell you?
CER (Character Error Rate): The percentage of incorrectly recognized characters. If 20 characters are wrong in a text of 1,000 characters, the CER is 2%. This metric is especially relevant when every single character matters, for instance with names, addresses or numbers.
WER (Word Error Rate): The percentage of incorrectly recognized words. A word counts as faulty if even a single character within it is wrong. The WER is usually higher than the CER and gives a more practical impression of how usable the result is.
What is "good"? That depends on the context. Modern AI-based pipelines achieve a CER below 2% on printed text: that is excellent for most use cases. For handwriting recognition, the benchmark is below 5% CER, which only a few years ago was considered unattainable.
Important here: these figures are averages. The actual recognition rate always depends on the quality of the input documents. A cleanly printed form is recognized better than a crumpled, faded handwritten note. Reputable providers communicate this openly, dubious ones promise 99% accuracy without context.
Traditional OCR providers who have used the same engine for 20 years often do not reach these values, especially not with handwriting or degraded material. This is where the decisive advantage of modern, AI-based approaches lies.
The vellmerk.ai approach
Vellmerk.ai does not build generic OCR solutions. Vellmerk.ai builds AI-powered document processing pipelines tailored to the specific data and requirements of each client. That may sound like marketing, but it is the decisive difference.
"Every client project is different. The data is different, the forms are different, the requirements are different. Anyone claiming to have a one-size-fits-all solution is making false promises.", Thorsten Vellmerk
The key to everything is the data. Before Vellmerk.ai talks about models, architectures or pipelines, it looks at what you actually have: which documents come in? In what quality? How much handwriting, how much print? Which languages, which layouts? Only once we understand the data landscape can we design a solution that works in practice, not just in the lab. Whoever skips this step is building on sand. Or, as we described it in another article: data is the information foundation on which everything is built.
What that means in concrete terms:
Data first, technology second: Vellmerk.ai starts every project with an analysis of your actual documents, not with test data from the lab. We look at your worst scans, your most illegible handwriting, your most complex layouts. Because that is where it becomes clear what the solution has to deliver. The solution options follow from the data, not the other way around.
AI-based pipelines instead of off-the-shelf software: Based on this analysis, we combine the best available models for OCR, HTR and ICR into a pipeline optimized precisely for your document types. The result: better recognition rates than generic solutions, at lower cost per document.
On-premise and locally deployable: In many projects this is the decisive point. If you process personnel files, health data, citizen data or other sensitive documents, this data cannot be sent to a cloud. Our pipelines run entirely on the client's infrastructure: the data never leaves the house.
"When personal data is involved, and with forms it almost always is, then the solution has to come to the client, not the data to the cloud.", Thorsten Vellmerk
GDPR-compliant by design: No dependency on US cloud services, no data transfer to third parties, full control over the processing chain. This is not a feature: it is a basic prerequisite.
Experience across the full range: Mastering OCR, HTR and ICR under one roof is rare. Most providers can do printed text. Some can do handwriting. But the combination of all three disciplines in an integrated pipeline: that is a niche Vellmerk.ai has been working in for years, across multiple client projects and over 20 years of practice.
The key takeaways at a glance
1. OCR alone is often not enough. As soon as handwriting, complex layouts or fluctuating quality come into play, you need HTR and ICR.
2. The technology is not the problem: the data is. The best pipeline is only as good as the understanding of your company's specific documents and requirements.
3. AI-based approaches beat traditional OCR. Modern pipelines with deep learning achieve recognition rates that were unthinkable a few years ago, and do so at lower cost than brute-force approaches.
4. Data protection is not negotiable. With sensitive documents, the solution has to run locally. Anyone who tells you otherwise either does not know the legal situation or ignores it.
5. There is no one-size-fits-all solution. Every project needs a pipeline aligned with the client's real documents, processes and quality requirements.
Conclusion
Text recognition is not a solved problem, at least not for the documents that land on your desk in the real world. OCR for printed standard text? Yes, that is a commodity. But as soon as handwriting, mixed formats or degraded material come into play, you enter a field where experience, specialization and the right AI architecture make the difference.
The good news: with modern AI pipelines, recognition rates are achievable that just a few years ago were considered science fiction. The even better news: these solutions can be operated entirely on-premise, your sensitive data stays where it belongs.
Do you have documents that need to be digitized, and "simple OCR" is not enough? Get in touch, Vellmerk.ai will analyze your documents and show you what is possible with a tailored pipeline.
About Vellmerk.ai
Vellmerk.ai is an AI consultancy (Danish ApS) founded by Thorsten Vellmerk. Drawing on 20+ years of process and IT experience and several years of hands-on AI consulting, Vellmerk.ai helps SMEs and public administration adopt AI in a practical, sovereign way, from strategy to local, on-premise-ready implementation. Proven across multiple client projects. Book an initial consultation.