AI projects won't deliver results until you fix your data (www.businessinsider.com)

AI projects are stalling not because LLMs lack capability but because companies feed them messy, dynamic unstructured data — emails, PDFs, slide decks, manuals and contracts that make up roughly 90% of enterprise content. The piece argues this “garbage in, garbage out” problem is the primary blocker: files are often outdated, duplicate-filled, rife with ROT (redundant/obsolete/trivial content), and missing crucial context like metadata, definitions, and timestamps. Unlike structured data, unstructured content represents nuanced human thought and requires continuous quality assurance, semantic processing, and change management rather than one-off fixes. Technically, common workarounds (context engineering, RAG/graph RAG, templating, or metadata-only ETL) only optimize retrieval of bad content and accelerate incorrect outputs. The article recommends decomposing documents into metadata, concepts, and free text, validating each part, enriching with "who/what/when/where/why," and automating cleanup and semantic enhancement at scale. Real-world results cited include Accenture’s finding that ~50% of GenAI pilots can’t scale due to data readiness and vendor-reported jumps in accuracy (e.g., a coffee brand chatbot to 93%, an airline RAG to ~90%). The takeaway: fixing unstructured data is a prerequisite for scalable, reliable agentic systems and rapid ROI from enterprise AI.