Everything Goldman Sachs Taught Me About AI (In 10 minutes)

Summary

1. 1.Verify the output. High-stakes financial firms cannot afford AI hallucinations or incorrect data entries. While large language models generate polished and convincing responses rapidly, a finished appearance does not guarantee factual correctness. According to Goldman Sachs CIO Marco Argenti, practitioners can extract value from a model's reasoning breakdown while remaining cautious about its final answer. Systems should verify underlying data inputs, require models to cite source figures, and incorporate deterministic tests, spot checks, or secondary AI review agents to validate results before deployment.
2. 2.Augment, don't replace. Organizations have relied on automated data systems and risk pipelines long before modern AI agents. Standard deterministic automations follow explicit rules consistently, making them predictable, easy to test, and simple to audit. Instead of replacing entire workflows with autonomous agents, builders should use standard automation to extract, clean, and calculate figures, while deploying AI models specifically to interpret the verified findings and produce natural language summaries.
3. 3.Understand the why. Goldman Sachs engineering operates under a build with purpose tenet, requiring technical work to tie directly to client or firm benefits. Builders frequently make the mistake of starting with an AI tool rather than an actual business problem. Effective AI implementations begin with a clear problem statement and measurable success criteria, using AI as a brainstorming partner to identify simple bottlenecks rather than overcomplicating unneeded workflows.
4. 4.Loop humans in. AI tools act like megaphones that amplify minor misunderstandings into large operational mistakes. An autonomous agent misinterpreting a task list can accidentally broadcast unauthorized discount codes to an email list of almost 200,000 recipients. To prevent compounding failures, builders should enforce human-in-the-loop checkpoints tailored to the severity of an action, configuring agents to save outputs as email drafts or proposal files rather than executing irreversible live actions autonomously.
5. 5.Transparency. Reliable enterprise systems must provide full visibility into how outputs are generated. When data affects financial reporting, risk assessment, or regulatory compliance, teams must be able to trace all data sources, intermediate assumptions, and validation checks. Keeping detailed execution logs ensures that workflows remain auditable, reproducible, and easy to maintain when system logic breaks or when team responsibilities transfer to new operators.