Descriptions:
Stefano Pasquali, who led liquidity research at Bloomberg and an investment AI team at BlackRock before building his current finance-focused AI platform, delivered a candid lessons-learned talk at AI Dev 25 NYC on why enterprise AI in finance keeps stalling before it scales. His central data point: an MIT paper finding fewer than 10% of AI pilots reach full production—a failure rate he attributes primarily to governance and trust deficits rather than model capability gaps.
Pasquali’s proposed antidote is a four-pillar platform designed around finance’s non-negotiable constraints. A knowledge graph layer ingests decades of legacy internal data and thousands of proprietary APIs that no general-purpose LLM can access. A model layer spans traditional linear regression (still the basis for roughly 80% of global financial decisions), deep learning for time-series forecasting, and generative AI. A purpose-built agent layer covers everything from workflow automation to portfolio management agents. The fourth pillar—governance—is what Pasquali calls an “MRI machine” that scans every reasoning step for data quality, hallucination risk, and auditability, placed at equal importance to the other three.
The talk pushes back firmly against two popular enterprise AI patterns: centralized shared models hosted outside a firm’s security perimeter, and chatbot-as-primary-interface for mission-critical decisions. For engineers and architects working in regulated industries, Pasquali’s insistence that financial AI must “prove, not just predict”—with full auditability designed in from day one—offers a useful counterweight to move-fast approaches common in consumer AI contexts.
📺 Source: DeepLearningAI · Published December 05, 2025
🏷️ Format: Opinion Editorial
