AI Agents Podcast Building AI That Thinks Like a Human - Brian Raymond Unstructured on Agentic Software & Human-AI Collaboration | EP 128
Mar 17, 2026
Brian Raymond, founder and CEO of Unstructured, builds tooling that turns messy enterprise files into GenAI-ready formats. He discusses why structured inputs like JSON and HTML matter, the persistent challenges of tables and scanned docs, and how preprocessed data can power more reliable agentic systems. The conversation also covers practical industry use cases and what enterprise AI needs to move prototypes into production.
AI Snips
Chapters
Transcript
Episode notes
RAG Solves Frozen Knowledge And Hallucinations
- Retrieval-Augmented Generation (RAG) fixes models' frozen-in-time knowledge and reduces hallucinations by supplying external data at inference.
- The model looks up fresh data, combines it with pretrained knowledge, then generates answers with citations.
Only Feed Models What They Need
- Give models only the data they need in the highest-quality, structured formats like JSON, Markdown, or HTML.
- That reduces context rot, cost, latency and improves likelihood of correct responses for large-scale enterprise pipelines.
Tables And Layouts Are Still Hard
- Scanned PDFs and tables remain hard for models: OCR plus layout parsing are separate hard problems.
- Tables, headers, page linkage and engineering schematics still perform below human level without specialized preprocessing.