AI Exchanges: The Role of Data

Episode Summary

Executive Summary: Goldman Sachs data leader Nima Raphael argues that generative AI is both an extension of machine learning and a step change in computers’ ability to create content. Its enterprise value will depend on organizing, cleaning, linking, and contextualizing proprietary data—not just using powerful models. He sees real capabilities, especially coding agents, while enterprise-scale differentiation remains to be proven; synthetic data and untapped corporate data offer further paths forward.

Main Topics: Data’s role in enterprise AI: Raphael says model outputs depend on the data provided. Enterprises need to consolidate, clean, normalize, connect, and explain their data so models can generate reliable, business-relevant answers. Generative AI: continuity and step change: AI continues the shift from hand-coded rules to learning from examples, but generative systems add the ability to create language, images, and audio rather than only identify or predict patterns. Probabilistic systems and organizational change: Unlike deterministic software, generative AI produces probabilistic outputs that may be useful but are not always correct or repeatable. Users need to understand that outputs are extrapolations from examples, not guaranteed facts. Enterprise opportunity versus hype: Raphael considers AI real and enduring, citing coding agents as a compelling capability, but says the extent to which companies can turn proprietary data into differentiated enterprise value remains to be demonstrated. Data availability and synthetic data: Although public data may be exhausted or increasingly supplemented by model-generated data, Raphael sees extensive untapped enterprise information. The speakers also raise questions about synthetic-data quality and possible limits on fresh human creativity. Goldman’s data foundations and AI use: Raphael traces his data work to a 2008 effort to consolidate exposure information after Lehman Brothers collapsed. He describes data engineering practices and AI agents that assist with data cleansing, normalization, and linking.

Key Arguments: Generative AI builds on machine learning’s learn-by-example approach, but its capacity to generate content marks a significant expansion beyond earlier pattern prediction. Because model behavior reflects the patterns in its input data, enterprise AI quality requires well-understood business semantics, accurate information, and links among related facts. Corporate data can create enterprise-specific value because substantial information remains trapped in separate systems and repositories; public and synthetic data alone do not exhaust the opportunity. AI is not merely hype: Raphael’s experience using a coding agent to solve problems he could not previously solve demonstrated practical capability. However, enterprise-wide differentiation is still an open question. Generative AI’s probabilistic outputs require a different user mindset from conventional software, where the same input is expected to produce a repeatable, traceable result. AI agents can support the data work that enables AI itself, including cleansing, normalization, and connecting information.

Data Points: Raphael’s tenure at Goldman Sachs: 20-plus years - He says he joined after college as a software engineering analyst. Start of his data journey: About five years into his career; 2008 - The financial crisis and Lehman Brothers’ collapse prompted work to assemble firm-wide exposure and liquidity data. Period of rule-based computing discussed: The first 50–60 years of computer science - Raphael contrasts programming computers with explicit rules against machine learning from examples. Time since the newer generation of models emerged: A couple of years - Raphael uses this to emphasize how early the enterprise adoption journey remains. Raphael’s son’s age: Three and a half years - Raphael describes using AI to help answer his son’s questions during his “why” phase. Episode recording date: September 25, 2025 - Stated in the episode’s closing disclosure.

Pivotal Quotes: "We've already run out of data." — Nima Raphael: He reframes the question of future data limits, then explains that models can build on prior models and that substantial enterprise data remains untapped. "Whatever data patterns you are feeding this machine is what it's going to learn and what it's going to extrapolate from." — Nima Raphael: His central explanation for why data quality and business context determine the usefulness of enterprise AI. "The potential, I think, in the enterprise is still to be seen." — Nima Raphael: He distinguishes the demonstrated usefulness of consumer tools and coding agents from the still-unproven scale of enterprise-specific value.

Implications: Organizations seeking value from AI should invest in data engineering, governance, and business context alongside models. AI can augment employees and data operations, but probabilistic outputs require appropriate expectations, and enterprise differentiation will depend on using proprietary data effectively.