Episode #11
Dan Maloney on Visual AI, Agentic Document Extraction & Building Trust in Enterprise AI | Episode 11
OCR has been around for more than 40 years. So why do the world's biggest banks still have thousands of people reading documents by hand? When Dan Maloney became CEO of Landing AI in 2024, as Andrew Ng stepped back from the day-to-day, he was returning to a problem he had first worked on at SAP back in 2001. When he looked closely at it again two decades later, he was struck by how little it had actually moved. Landing AI's mission is to make the world's documents computable. Instead of growing up from OCR and patching its limits with templates and heuristics, Landing AI came at the problem from visual AI, blending purpose-built models, an intelligent router, and agentic reasoning into a single system that reads a document the way a person does. Today that system extracts structured data from the messiest documents enterprises have, the scanned pages, the tables inside tables, the handwritten forms, at accuracy levels they can build on. Before every enterprise had an AI strategy... Before "agentic" became a boardroom word... Before the industry spent a year token maxing... There was a quieter, more stubborn problem: The world's data was trapped in documents, and no one could read it at scale. In this episode of the Future of Data & AI Podcast, Dan Maloney, CEO of Landing AI and a two-decade veteran of enterprise software and AI, joins Raja Iqbal for a grounded conversation about what it actually takes to move visual AI from an impressive demo into production. Dan is candid about where the hype outruns reality, why the model is the smallest part of the equation, and how a company earns the trust of a compliance team, not just an engineering one. What You'll Discover Why "good enough" accuracy can be as costly as zero accuracy. At billions of documents, a small error rate forces humans back into the loop. Dan explains why the goal is to get so close to 100% that enterprises can stop checking.Why the model is the smallest part of the system. Landing AI blends its own document pre-trained transformers with LLMs only when needed, wrapped in a harness that handles routing, parsing, and extraction. The intelligence is in the system, not the model alone.How Landing AI reads a document the way a person does. Its DPTx models take in layout, tables, figures, signatures, and marginalia, break them down to the block and word level, and hand each piece to the right model downstream.Why document processing is 80% a vision problem. Dan explains how solving the visual complexity first, the columns, the modalities, the spatial structure, turns the rest into a much simpler language problem.How you earn trust from compliance, not just engineering. Visual grounding shows exactly where every answer came from. SaaS, VPC, on-prem, and air-gapped deployments let regulated industries adopt AI without their data ever leaving the building.Why token maxing gave way to a reckoning. Companies loosened the guardrails to see what AI could do, then met the bill. Dan explains what separates spend that changes the game from spend with nothing to show for it.Why you shouldn't boil the ocean. Dan's advice for leaders getting started: pick one core, high-value use case, prove it end to end, and learn the orchestration, retries, and human-in-the-loop decisions that carry over to everything else.Where document intelligence is headed. Dan makes the case that freeing humans from decades of administrative document work is the disruption that unlocks entirely new revenue models and businesses. This Episode Is For ML and AI engineers building document ingestion and extraction pipelines Enterprise and data leaders deploying AI in regulated industries CIOs and CISOs weighing SaaS, VPC, and on-prem trade-offs Teams trying to measure real ROI on their AI spend Anyone thinking seriously about visual AI, trust, and scale
