Most important take away

Goldman Sachs has moved decisively past the AI experimentation phase: 47,000 employees now use their internal AI assistant (GS Assist) daily, generating over a million prompts per month, and the firm has already terminated contracts with third-party software vendors in favor of internally built AI-powered replacements. The key metric Goldman tracks is not headcount reduction but increased output — projects finishing ahead of schedule, higher quality, and the ability to tackle work that previously sat below the priority cutoff line.

Summary

Actionable Insights and Investment Considerations:

• Token costs per unit will decline, but total token spend will rise significantly. Argenti predicts token volume growth will outpace per-unit cost reductions, making total AI compute costs a major line item for enterprises. This is bullish for GPU makers and cloud infrastructure providers.

• The buy-vs-build equation has shifted toward build for small/medium applications. Goldman has already terminated some third-party software contracts. SaaS companies tied to processes that AI is fundamentally changing (developer tooling, simple UX-layer apps, survey/expense tools) face disruption risk. SaaS companies tied to regulated, stable processes (accounting, general ledger, systems of record) are more defensible.

• Legacy software disruption is not uniform — evaluate by process stability. Argenti’s framework: ask whether the underlying business process a software supports will change in five years. If the process is stable and heavily regulated (e.g., closing books), the software is safe. If the process itself is being transformed by AI (e.g., software development lifecycle), incumbent vendors are at risk.

• Centralized model gateway infrastructure is critical. Goldman built a platform that intelligently routes queries to the optimal model on the cost-quality Pareto frontier. Companies that solve token routing and optimization at scale have a structural advantage. This is a growing enterprise need.

• Forward-deployed engineers from model providers (Anthropic, etc.) are the preferred integration model over traditional systems integrators or middlemen. During rapid AI evolution, going directly to model makers reduces lag. This disadvantages traditional IT consulting/integration firms.

• Goldman’s data moat matters for the “last 10%.” While retail investors can now get 90% of the way to institutional-quality analysis with AI tools, Goldman’s edge comes from proprietary data, cross-asset-class visibility, global on-the-ground relationships, and complex multi-instrument portfolio management. The competitive advantage in finance is narrowing but still significant for sophisticated clients.

• AI is not reducing developer headcount at Goldman. The backlog of unbuilt projects is large enough that increased productivity is being channeled into more output rather than fewer people. Argenti frames it as optionality: capacity increased to 120-130%, and leadership can choose to do more or reduce — but they are choosing to do more.

• Key stocks/sectors implicated: GPU manufacturers benefit from rising token volumes. Cloud hyperscalers face near-term losses on power users but long-term optimization gains. SaaS companies should be evaluated individually by process-disruption risk rather than as a category. Systems-of-record vendors (CRM, ERP, accounting) are better positioned than workflow/UX-layer tools.

Chapter Summaries

Introduction: AI Is No Longer an Experiment (Tracy and Joe) The hosts reflect on the pace of AI change since ChatGPT launched in 2022, noting we have moved past the experimentation phase. Joe describes his own usage going through a trough after initial excitement, then resurging with tools like Claude Code. They set up the interview by asking what concrete ROI looks like now.

Goldman’s AI Deployment at Scale (Marco Argenti) Argenti explains that AI at Goldman is “not a drill” — 47,000 employees use GS Assist daily for complex research queries spanning geopolitical events, asset volatility, and portfolio strategy. The firm has wired up hundreds of data sources and built Legend AI, a lakehouse tool that connects data to MCP servers in a few clicks. Data quality is the key differentiator for AI output quality.

Developer Productivity and the Changing Nature of Software Engineering Every Goldman developer has access to agentic AI tools including Devyn, Claude Code, and GitHub Copilot Agent. The real impact is projects finishing ahead of schedule, not headcount reduction. Developers now spend more time as planners and product managers, with AI handling mechanical coding tasks. The backlog of unbuilt projects means increased capacity goes to more output.

The Buy-vs-Build Shift and SaaS Disruption The cost of building simple applications has dropped dramatically — people build working apps over a weekend. Goldman has already terminated some third-party software contracts. Argenti offers a framework: evaluate software by whether its underlying process will change. Regulated, stable processes (accounting) are safe; rapidly evolving processes (developer tools, simple UX apps) are vulnerable.

Forward-Deployed Engineers and Working Directly with Model Providers Argenti explains that forward-deployed engineers are product builders from model companies (like Anthropic), not traditional support staff. During rapid AI evolution, cutting out intermediaries and going directly to model providers accelerates adoption and creates productive culture clashes within teams.

Token Economics and the CFO’s Sticker Shock Goldman centralizes model access through a model gateway that routes queries to the optimal cost-quality tradeoff. Argenti’s philosophy is to isolate users from “token anxiety” and let the central platform team handle optimization. Per-unit token costs will decline, but total spend will rise as reasoning and agentic workflows consume far more tokens.

Integration, Systems of Record, and AI Security Integration remains critical, especially around regulated systems of record. Goldman enforces information barriers through a centralized identity system where every AI agent gets a badge with the same access controls as human employees. The GS AI platform took nearly two years to build because of cybersecurity, info barriers, and compliance requirements.

Goldman’s Competitive Moat: The Last 10% While retail investors can now get 90% of the way to institutional analysis, Goldman’s edge is proprietary data, cross-asset correlation visibility, global presence, and expertise in complex multi-instrument portfolios. In the “Formula One” analogy, the difference between first and last is one second per lap, but that gap determines everything.

The Future of AI Talent and Work-Life Balance AI is turning every knowledge worker into a manager who must explain, delegate, and supervise. Argenti sees renewed excitement among engineers rather than burnout, comparing the current moment to past paradigm shifts (Excel, Python, mobile). Repetitive toil is being removed, letting developers focus on higher-level planning, which he views as a net positive for job satisfaction.

Hosts’ Closing Discussion Tracy and Joe highlight the regulatory framing (black-box models are not new to finance), the engineering challenge of token budget optimization across a firm, and the incentive structures around model routing as particularly interesting topics for future episodes.