Spend enough time around enterprise AI projects, and you’ll realize it’s rarely the model that teams complain about. Usually, the conversation turns to information that’s difficult to locate.
It’s trapped inside disconnected systems or buried in documents that were never designed for software to interpret. Organizations may be investing heavily in AI, but many are discovering that accessing their own knowledge remains a stubborn obstacle.
That challenge becomes more apparent as AI moves from experimentation into day-to-day operations. Employee access to AI tools increased significantly last year. It has increased pressure on technology leaders to turn pilots into production deployments.
The information enterprises depend on was never created for AI
Every large organization sits on a vast archive of institutional knowledge. Insurance carriers maintain decades of claims records. Manufacturers store engineering drawings, maintenance logs, and technical specifications. Utilities manage invoices, infrastructure documentation, and regulatory filings. Healthcare providers work with clinical records and compliance documents.
Most of these assets share something in common: they were created for people. Humans can look at a document and immediately understand how information fits together. We instinctively recognize the relationship between a table and the paragraph that follows it.
We understand that a note in the corner of a drawing may carry more significance than a block of text elsewhere on the page. Humans easily distinguish between supporting information and the information that drives a decision. Software, however, has historically struggled with these distinctions.
Many enterprise AI initiatives assume organizations can index and retrieve information easily. That assumption fails when AI cannot consume the data.
Organizations encounter thousands or millions of documents that contain hard-to-read data for AI. They contain diagrams, annotations, forms, charts, handwritten notes, and visual structures that carry meaning beyond the words themselves.
There is no shortage of information inside organizations. The challenge is how to turn that information into something AI can reliably use.
Solving the problem is complex
There’s a tendency to treat document understanding as a solved problem. After all, optical character recognition has existed for decades. Modern language models can summarize reports, answer questions, and generate content in seconds. From a distance, it’s easy to assume that enterprise document processing has largely been figured out.
This is where many easy solutions begin to struggle. Extracting text from a document and understanding the significance of that information are very different tasks. Enterprise workflows depend on context, relationships, and traceability. Those requirements can be invisible in a product demonstration and impossible to ignore once a system is deployed.
This is one reason many organizations find themselves revisiting the data layer after investing in AI. In fact, data quality, data processing, and data governance continue to rank among the most common barriers preventing organizations from scaling AI initiatives.
The conversation often begins with models and ends with information.
The cost of getting document understanding wrong
As enterprises push AI into more operational environments, the consequences of misunderstanding information become more significant.
Consider an insurance claim. A single case may include photographs, witness statements, medical records, invoices, and supporting documentation collected over months or even years. The challenge is rarely locating a document; it’s understanding how those pieces fit together and preserving the context that allows an investigator to make an informed decision.
That challenge is becoming more complicated as fraud techniques evolve. Earlier this year, Aviva reported detecting more than £233 million worth of fraudulent claims during 2025. There is a growing number of cases that involve AI-generated documents and manipulated evidence.
When AI supports investigations, teams must understand document information, but that is only part of the equation. Understanding how that information relates to everything around it is equally important.
For organizations operating in regulated industries, confidence in an AI-generated answer depends on understanding where that answer came from. Security and risk concerns are now among the most frequently cited obstacles to scaling agentic AI initiatives. That concern highlights a broader need for traceability, governance, and confidence in the underlying information.
Organisations also exercise greater caution regarding their data’s location and usage. Some are revisiting private infrastructure and on-premises deployments after years of cloud-first thinking. Organizations must take this step when handling sensitive information. It also supports data sovereignty initiatives.
AI needs accurate data, not just more models
Ultimately, the future of enterprise AI depends on more than increasingly capable models.
Organizations have spent decades building repositories of knowledge. These sit inside contracts, drawings, reports, records, and operational documentation. Much of that knowledge remains difficult for AI systems to access, interpret, and use with confidence.
Many organizations are discovering that the path to better outcomes begins much further upstream. The starting point for them is a clearer understanding of the information they already possess.
For CIOs, the questions now are: can you access all of your key data? If not, what are you doing about it?
Valantor is a focused acquirer and operator of AI businesses built on defensible technology, differentiated data, and deep vertical expertise. Its inaugural acquisition, EyeLevel, brings with it GroundX, an AI visual intelligence platform that helps enterprises understand complex business documents while keeping sensitive data secure through private and on-premises deployments.
Built by veterans of IBM Research and IBM Watson, the EyeLevel team developed one of the earliest large-scale conversational AI systems, serving more than two million users a day, before founding the company in 2019. After completing UC Berkeley’s SkyDeck accelerator, the team developed GroundX to help organizations unlock information trapped within contracts, reports, technical drawings, forms, and other document-heavy workflows.
Today, Valantor delivers secure, scalable AI to Fortune 1000 enterprises, with customers including ADP and Arcadia and strategic partnerships with Red Hat and BigBear.ai. The company recently introduced GroundX Studio, combining AI coding agents with a no-code interface to help enterprises build document-intensive AI applications more quickly while maintaining the security and governance required for production environments.
