Goldman Sachs launched its GS AI Assistant to all employees in June 2025, following a pilot programme that had reached approximately 10,000 knowledge workers since January of that year [1]. The announcement received the standard coverage: a Wall Street giant embracing generative AI, productivity gains anticipated, a quote from the Chief Information Officer comparing the assistant to a new hire who would absorb Goldman culture over time [2]. What that coverage largely missed was the architectural detail that makes the announcement more interesting than it appears: the GS AI Assistant is explicitly designed to route tasks to different underlying models depending on the nature of the work, switching between OpenAI's GPT-4o, Google's Gemini 2.0 Flash, Anthropic's Claude 3.7 Sonnet, and Meta's open-source Llama variants as the task demands [2].
That design choice is not a convenience feature. It is a structural position. Goldman's leadership has made a bet, visible in the architecture if not in the press releases, that the capability gap between frontier foundation models will narrow to the point where no single provider commands a decisive advantage. In that world, the institution that wins is not the one that has the best model. It is the one that has the best integration layer, the best data governance, and the deepest encoding of institutional knowledge into a form that any competent model can retrieve and apply. Goldman is building for that world.
The Architecture of Model Agnosticism
The GS AI Platform, which sits underneath the assistant interface visible to employees, is not a wrapper around a single external API. It is a routing layer: a system that receives a request, classifies it by task type, selects the model most suited to that task from the available inventory, and returns a result within Goldman's audited security perimeter [3]. Code requests go to coding-optimised models. Financial document summarisation routes to language models fine-tuned on structured financial text. Research translation tasks take a different path still. The selection logic is proprietary, but the architecture is explicitly documented in the firm's technology disclosures and in statements by CIO Marco Argenti [2].
This design has a direct implication for vendor negotiating power. A bank that commits its entire AI workflow to a single provider is, in practice, locked in: switching costs are high, data dependencies accumulate, and the provider's pricing leverage increases with each passing quarter. Goldman's model-agnostic architecture preserves optionality. If OpenAI's pricing rises or a competitor's model materially outperforms on a specific task class, the routing layer can redirect traffic. The infrastructure investment is in the layer that Goldman controls, not in the layer that its vendors control.
The Legend Data Layer
The more consequential piece of Goldman's infrastructure rebuild is less visible in the press coverage, because it predates the AI wave by several years. Legend, Goldman's data modelling and management platform, was developed internally and open-sourced through FINOS (the Fintech Open Source Foundation) starting in 2020 [4]. Its integration with Databricks, announced in a partnership formalised in 2023, extended it into a full lakehouse architecture built on Apache Iceberg's open table format [5]. The combination provides what Goldman's engineers describe as a governed, queryable layer over the firm's structured data: consistent definitions, enforced access controls, and the ability to run transformations via Legend functions that integrate directly with Databricks User Defined Functions [5].
The relevance to the AI stack is direct. A large language model is only as useful as the data it can access and the accuracy of that data. An institution that routes queries through a model agnostic layer but has no reliable mechanism for ensuring the underlying data is consistent, current, and correctly permissioned is building on sand. The Legend/Databricks infrastructure addresses precisely this problem. Goldman is not the only bank investing in data governance, but the open-source route, which has attracted contributions from other financial institutions and allowed Goldman to share infrastructure costs while retaining the proprietary data layer, is a structurally different approach from the closed, proprietary data stacks that most of its competitors maintain.
Goldman hired more than 500 AI engineers in 2024. The investment in headcount tells you what no press release will: this is not a project. It is a rebuild.
What the Headcount Data Says
Goldman hired more than 500 AI engineers in 2024 [6]. That figure, drawn from the firm's own disclosures and confirmed by multiple industry surveys of technology hiring in financial services, deserves to be read in context. Goldman's total technology headcount across all functions is roughly 10,000. Adding 500 AI specialists in a single year, while also restructuring the data infrastructure and rebuilding the model integration layer, is not an experiment. It is a commitment of a scale that takes years to reverse. The investment in headcount, combined with a $200 million strategic stake in AI chipmaker Cerebras [7], tells you what no press release will: the build phase Goldman described in 2023 was exactly that, and 2025 is the deployment phase.
The gap between Goldman's public AI narrative and its actual infrastructure investment is not as wide as it is at most of its peers, but it still exists. The public narrative emphasises productivity: AI assistants that help analysts draft documents faster, summarise earnings calls, translate research into client-preferred languages. These are real and useful applications. The more consequential application is the one that receives less coverage: the deployment of AI reasoning over Goldman's proprietary structured data, the thirty years of transaction history, pricing models, risk parameters, and counterparty records that sit in the Legend layer. That is the asset a commoditising model market will be unable to replicate. Goldman is not building a moat around its models. It is building a moat around its data.
- Goldman Sachs. "Goldman Sachs Expands Availability of AI Assistant Across Firm." Goldman Sachs Press Release via PYMNTS. June 2025. pymnts.com
- Arash Massoudi and Joshua Franklin. "Goldman Sachs Rolls Out an AI Assistant for Its Employees." CNBC. 21 January 2025. cnbc.com
- Nanonets. "Goldman Sachs Deployed Its AI Platform." Nanonets Blog. 2025. nanonets.com
- FINOS. "Goldman Sachs Open Sources Its Data Modelling Platform Through FINOS." FINOS Press Release. 2020. finos.org
- Databricks. "FINOS' Legend to Integrate with Databricks Lakehouse." Databricks Blog. 2023. databricks.com
- DigitalDefynd. "5 Ways Goldman Sachs Is Using AI." DigitalDefynd. 2025. digitaldefynd.com
- Klover.ai. "Goldman Sachs AI Strategy: Analysis of AI Dominance in Financial Technology." Klover.ai. 2025. klover.ai
- Goldman Sachs. "2024 Annual Report." Goldman Sachs Investor Relations. 2025. goldmansachs.com