When OCR reads, generative AI understands

For several years, document recognition technologies have been transforming the way HR documents are processed. OCR (Optical Character Recognition) has made it possible to automate the reading of scanned documents, sparing HR teams from countless manual data-entry tasks. 

This first generation of technology was subsequently enhanced by Computer Vision, which can automatically identify the type of document received, and later by zonal OCR, which extracts information from predefined areas based on the recognized document type. More recently, Intelligent Document Processing (IDP) has enabled extracted data to be structured and automatically integrated into HR workflows. 

These advances have significantly improved operational efficiency. Yet they still have an important limitation: they can read, but they do not always understand, leading to errors and a continued need for manual intervention. 

The limits of traditional methods 

When documents are well structured and properly scanned, recognition rates are generally excellent. Difficulties arise, however, when documents are handwritten, poorly scanned, or contain ambiguous information. 

These Computer Vision technologies first proved their value in HR through a solution we deployed during the Covid period to automate the processing of vaccination records in the United States as employees returned to the workplace. Many vaccination cards were partially handwritten, sometimes difficult to decipher, with signatures obscuring key information.

In such situations, Computer Vision and OCR algorithms were often able to provide a reading of the document, but they could not determine whether that reading was actually coherent and reliable. As a result, significant and costly manual review remained necessary. 

The example of sick leave certificates 

Sick leave certificates provide a perfect illustration of this challenge, as they highlight the pressure placed on HR processes. Employees are required to notify their employer without delay, with common practice generally allowing 48 hours when no different rule is specified in a collective agreement or company policy. They also have 48 hours to inform their social security organization. 

In addition, daily sickness benefits continue to increase. According to the 2025 edition of Dépenses de santé en 2024 (Healthcare Expenditure in 2024) published by DREES, these benefits increased by 6.2% in 2024, reaching €21.4 billion, including €12.1 billion related to illness. France's National Health Insurance system also publishes open datasets covering sick leave benefits paid through 2024. 

These figures confirm that HR document automation addresses a large-scale need: streamlining processing, reducing turnaround times and ensuring the security of processes that directly affect employees and HR teams. 

For sick leave certificates, OCR can extract dates, the employee’s identity and the doctor’s details. Yet the real issue lies elsewhere: 

• Are the dates consistent with one another? 

• Does the extracted information match the data in the HRIS? 

• Is a poorly read value plausible given the context? For example, does the duration of the leave actually correspond to the period between the start and end dates? 

An error in this type of document can have an impact on payroll, HR administration and regulatory reporting. 

The objective is therefore not simply to read information correctly, but to assess its consistency and accuracy before it is automated. 

Adding a layer of understanding with LLMs 

This is precisely where the new generation of language models, including LLMs and some SLMs, represent a genuine breakthrough. 

Unlike traditional OCR, which recognizes characters, a language model can analyze the relationships between the extracted information and determine whether it is accurate. 

Our teams have tested and approved the use of this intelligence layer on top of our OCR and Intelligent Document Processing (IDP) systems to: 

• detect inconsistencies 

• check business rules cho

• ose between several possible interpretations of handwritten text 

• complete certain information 

• assess the confidence level before automation 

In other words, generative AI does not replace OCR. It leverages the output of the document recognition engine and adds a layer of understanding that has until now been missing. The value of this approach does not lie in blind automation. When information remains uncertain, the system can request approval from humans. 

The future of IDP: reimagining the role of documents 

The combination of OCR, Computer Vision, and generative AI marks a new stage in the digital transformation of HR services. It significantly increases the number of documents that can be processed automatically while improving the reliability of data integrated into the HRIS an essential requirement when dealing with information as sensitive as payroll, employee entitlements, and regulatory compliance. 

Some articles on AI suggest that the future of IDP is uncertain. Many predict that documents will eventually be eliminated because all systems will be connected through APIs, exchanges will occur in real time, and AI agents will consume data directly. 

At Sopra HR, we believe this analysis does not apply to HR. Documents are legal assets before they are data sources. Employees will continue to provide documents with legal value (sick leave certificates, bank details, identity documents, work permits and proof of address) to access HR services. Industry analysts also anticipate an increase in fraud related to these processes.  

The future of IDP is not about eliminating documents, but transforming what they can do. Rather than serving merely as data sources processed through OCR, HR documents will, through AI, become catalysts for HR decisions, controls and workflows. 

More than just a technological development, this transformation marks the transition from document-reading automation to automation built on trust in the information those documents contain.