The Data Exchange with Ben Lorica The Gap Between AI Hype and Enterprise Reality
30 snips
May 7, 2026 Barry Dauber, Databricks exec focused on customer engagement and AI adoption. Richard Garris, Databricks data and AI practitioner advising on operationalization. They discuss taking AI from demo to production. Topics include handling nondeterministic LLMs, ownership and governance gaps, retrieval-augmented generation, fine-tuning vs prompting, token costs, context management, agents and evaluation, multimodal readiness, and vendor tradeoffs.
AI Snips
Chapters
Transcript
Episode notes
Preprocess Documents With Two Pass And Synthetic Data
- Preprocess unstructured documents into structured fields and use two-pass extraction plus synthetic data to improve accuracy.
- Richard describes AI Parse Document: open-source model for bounding boxes, then a foundation model for reasoning and synthetic variations for stamps/placements.
Turn Analysts' Work Into Training And Tests
- Use human-curated examples as test cases and training data to speed quality improvements rather than hand-crafting prompts for every format.
- Richard recommends converting real analysts' manual extractions into training/test sets and augmenting with synthetic variants.
Graphs Are Hard; Tool Calls Win Practicality
- Knowledge graphs were hyped but haven't solved the core problem because they are hard to build, maintain, and can become hairballs.
- Richard suggests tool-calling (SQL, specific tools) plus LLMs as agents is often more practical than large knowledge graphs.