LLMs don’t need to be perfect to power real-world workflows. But they do need to be predictable. 80–90% of enterprise data is unstructured: PDFs, scans, images, freeform text. This type of messy data used to block automation because brittle rules, fixed templates, and early ML couldn’t handle the variability. But as @anantpbj explains on the AI + a16z podcast with @appenz, the emergence of: – Layout-aware transformer models – Structured validators + fallback review – Compile-time agent workflows is powering real-world use cases including Indian banks approving loans via WhatsApp.