Enterprise AI and Content Debt: A Barrier to Productivity

Content Debt: An Underestimated Barrier for Enterprise AI

The enthusiasm surrounding artificial intelligence (AI) promised a revolution in business productivity. However, tangible results have yet to materialize. In 2024, a study conducted by the Boston Consulting Group (BCG) revealed that only 4% of companies manage to generate substantial value from AI. The following year, Gartner estimated that by 2026, 60% of AI projects would be abandoned. This disillusionment is partly attributed to a problem often overlooked: content debt.

When the Document Becomes the Problem

AI models are evolving at an impressive pace, but companies struggle to translate these advancements into concrete successes. This gap deserves particular attention. The obstacle may not lie in the algorithms themselves, but in the data they utilize. When an enterprise AI generates a response, it does not rely on a universal knowledge base. It draws from the documents and information that the organization provides. If this data is confusing, contradictory, or outdated, the AI cannot produce reliable results. Even the most sophisticated AI model is dependent on the quality of the data, often compared to essential fuel.

This problem is not new. Before the era of AI, companies were already wasting time searching for accurate information across a multitude of documents. The question of the correct version of a process or procedure was already a headache. How many times has information been searched for, verified, and ultimately found in a different document than the one initially consulted?

AI Does Not Create Disorder. It Reveals It.

Office documents such as Word, PowerPoint, and PDF have become standard tools for recording knowledge. While effective for writing and sharing information, they present significant limitations. By encapsulating content, they make it difficult to retrieve and reuse. This encapsulation leads to practices of copying, adapting, and duplicating, which ultimately distort the original information. The proliferation of versions of the same document can lead to contradictions, making it challenging to identify a single source of truth.

In practice, this translates to procedures being updated in one document but not in another, or different versions circulating via email and shared spaces. This duplication undermines the integrity and value of the information. For example, a procedure may be updated in a shared file but not in the one sent by email, creating discrepancies.

What Constitutes a Viable Knowledge Base

A poorly fed generative AI struggles to create coherence from inconsistent data. While anti-hallucination rules limit obvious fabrications, they cannot resolve internal contradictions. The AI may then produce plausible responses that are not necessarily true, seeking statistical coherence rather than factual proof.

To succeed in an AI project, it is crucial to focus on the quality of the knowledge provided. An exploitable knowledge base relies on four essential principles:

• Identify a single source of truth, where critical information is centralized.
• Define clear governance for the creation, access, and validation of content.
• Enrich content with metadata to determine its nature, status, and relationships.
• Ensure the sustainability of this set through regular checks and audits.

Although these principles may seem obvious, their implementation often encounters the complexity of the documents themselves. Companies must ensure that each document is up to date and that all versions are aligned.

Moving Towards Structuring and the Knowledge Graph

Traditional documents group varied information without clear distinction. This poses a problem, especially in regulated sectors like aerospace or defense, which have adopted structured content for certain technical documentation.

This approach involves organizing knowledge into unique components, each with its own lifecycle and metadata. Thus, any publication results from a dynamic assembly of validated content, illustrated by the DITA XML standard. The benefits include multichannel publishing, task parallelization, better traceability, and reduced translation costs.

By continuously enriching content with metadata, companies gradually create a knowledge graph. This living mapping allows AI to respond effectively to queries by relying on a single source of truth, thereby improving the speed and reliability of responses.

Long considered a concern for technical writers, this approach is now recognized for its importance in optimizing AI usage. Companies that adopt this method see a notable improvement in the efficiency of their AI systems, as they can rely on reliable and consistent data.