Document Management in the Age of AI
In 2025, it’s rare that you run into someone who uses a flip phone as their primary mode of contact. Typically, when you see someone still utilizing outdated technology you wonder where they’ve been for the past 15 years. This mode of thought extends beyond personal technology. Working in this industry for decades, your professional and personal interests center around technology. It’s easy to fall into the belief that others have the same understanding of new technologies as you do, but just like there are people hesitant to move on from their flip phones, there are businesses hesitant to stray away from the way they’ve done business for years. It’s understandable, technology evolves quickly, and investing in developments for your business is far more complex than investing in the newest smartphone. While many progressive leaders are eager to use and consume it, many end users and organizations are still incredibly low-tech.
Since 1996, there have been ongoing discussions regarding the idea of the “paperless office.” What exactly does a “paperless office” entail? It centers around the reduction of printing and copying as well as eliminating paper business processes. Nearly 30 years later, many industries and processes still heavily rely on paper. Processes that have long retention schedules, outdated state libraries which set retention schedules, and industries like K-12 and municipal government all still utilize paper the same way they have for decades.
The rise of ECM
Before artificial intelligence and cloud-based solutions dominated the technology conversation, enterprise content management (ECM), also called document management, was the hot new solution on the market. Demoing ECM back in the early 2000s is comparable to demoing AI now. It provided easy searching, distribution of documents, and scanning capabilities—all revolutionary at the time. End user customers’ and prospects’ mouths would drop when they saw the ease of finding important documents thanks to ECM.
Fast forward to 2025, when despite the shifting of the general population’s use of technology, most companies still want to control paper, manually perform light workflow, and handle their own basic file storage. There is a large segment of the US market that isn’t getting on the new tech bandwagon, and these are the organizations who are the ideal target for document management. Many of them have legacy ERP or CRM systems, dated policies and procedures, would rather throw labor at the problem, and likely has staff or management that cannot pivot with the times. For these organizations, ECM isn’t just a “nice to have,” it’s a necessity in order to keep up with the ever-evolving business landscape.
Riding the tech waves
The term “tech waves” has been a long-standing way to track the maturation of technology. As it relates to the business focus of document workflow automation, we see the first wave circa 1995 in the form of basic document management, structured zonal optical character recognition (OCR) capture and advanced searching on keywords or metadata.
Wave 2, around the turn of the millennium, now sees the advent of business process automation using rules-based workflows, electronic forms, and technology designed to be used on smart phones and tablets. Suddenly, paper forms and processes can be automated using web-based forms, tablets, and wet signatures. It was during this wave where Dotloop, DocuSign, and countless niche mortgage closing applications cannibalized the ECM gap within the mortgage industry. Once a paper factory with large closing files and minimal cloud storage options, ECM filled a gap that no longer existed thanks to the swift introduction of these solutions to the market.
Wave 3 came about circa 2010, when the iPad and smartphones began to proliferate in the consumer and business market. During this wave organizations begin to invest in promoting efficient processes using smart phones as the intake/onramp. Businesses realize that accessibility of document search and capture from the phone is a feature worth focusing on. Evolving from toolkits that allowed for screen scraping and industrial strength macro, robotic process automation (RPA) tools started to enter the market. Common repeatable labor tasks were now able to be automated using rules-based logic. There’s a growing prevalence of machine learning tools that can take prebuilt document knowledgebases and achieve operational efficiency within a specific horizontal use case like AP invoices.
AI enters the picture
Around 2020, AI and neural networks started to infiltrate hypervisors. Azure, Google, and AWS release their own tools like Azure Form Recognizer (later rebranded Document Intelligences), Google DocumentAI, and AWS Textract. Now, developers can send documents to hypervisor OCR engines to automatically extract relative data without templates or telling the engine where the data is located on the document. Intelligent document processing (IDP) now parses the JSON payload returned by the engine, but the extraction is mostly focused on key-value pair. Completely hands-off learning and extraction doesn’t hit the market until circa 2023, when ChatGPT and OpenAI start commercializing the large language model (LLM) and generative AI tools for wide adoption.
The next frontier
Now, in 2025, there’s a trend of deprecation of neural and prebuilt models for LLMs — for example, the consolidation of Azure AI Document Intelligence into Azure AI Foundry. The document extraction and classification limitations of LLM’s of the past, ability to read checkboxes, tabular data, and understand the nuances in visual pagination have been solved. Hallucinations where LLMs return inaccurate data is fixed by using retrieval-augmented generation (RAG).
The gap in interoperability between LLMs and external/internal data sources is fixed with model context protocol (MCP) via standardization. Visual language models (VLM) have started to replace LLMs, allowing more context to be given to visual aspects of a document. RAG has started to take a broader role within generative AI, keeping models honest and accurate with validating results using external or internal data sources. Agentic RAG now helps models bounce results against internal ERP/CRM data sources or third-party web services to validate information such as whether the address is a healthcare facility, whether the FIN/EIN is a valid business, and other key data points. Just like natural language processing (NLP) helps generative AI understand the context of words, so does RAG when it comes to validating the context of a prompt response. As developers continue to refine their focus on further developing this technology, there will likely be expansions regarding exception management, external validation, and making document extraction datasets able to be prompted through chatbots. In that specific case, prompts will be the accepted input source method, and the output will be much broader than internet only responses. Combining the available information on the internet with external/internal data sources means that LLM’s with have grounded truth and end users can simply ask the chat bot assistant specific questions without having to navigate data buried in a manual report.
Marketplaces, models, and monetization
Overall, within the last five years technology related to document workflow automation has drastically escalated. It’s almost a biweekly event to see changes in document AI, not only from the more than 450 IDP providers, but also from the hypervisors leading the way, such as Azure. Users can now consume non-Azure models from open-source players like Hugging Face, Qwen, Mistral, DeepSeek, Meta Llama, and the like, or even other third-party paid models like Open AI and Anthropic’s Claude. Hypervisors are now LLM/generative AI marketplaces, charging tokens for input and output characters, and driving computational cycles on their cloud.
Benchmarking is now standard within these hypervisor marketplaces, where a person can understand and see rankings of the best models for performance, groundedness, and cost. What was once the Wild West of AI models, with users asking, “where and how do I use them?” has transformed into revenue generating profit centers for these hypervisors. For instance, one can run open source or free models on Azure and pay for cloud cycles without Azure having to pay royalties for use of the model. It can be expected that open source commercial models with self-hosting will crop up in the future.
Budgets based on token pricing are challenging to figure out, and while token-based pricing is generally less expensive than per page pricing, rogue prompts that return too many output characters can quickly get out of control. With so many models in the market now, the pricing pressures will make models a commodity and it will be about the value-add software that wraps around these models that contributes to industry growth.
Brent Wesler
Brent Wesler is VP of Strategic Technologies & Digital Automation at VRC Companies.