Goldman CIO Marco Argenti on the Warp-Speed Improvements in AI
Digest
This podcast explores the transformative impact of AI on business software, developer workflows, and organizational strategies. It begins by highlighting the challenges of complex traditional software like CRMs, introducing Pipedrive as a simpler alternative. The discussion then delves into the accelerated timeline of AI adoption, driven by advancements like ChatGPT, and the critical need to evaluate AI's return on investment. Personal AI usage has evolved from experimentation to practical applications, with AI now capable of complex reasoning and mission-critical tasks. Goldman Sachs exemplifies pervasive AI integration with its GSAI assistant, used by thousands daily, which enhances client experience by handling complex queries and analyzing data. Data quality is emphasized as crucial for AI effectiveness, facilitated by platforms like Legend AI. Agentic AI is fundamentally changing developer roles, shifting them towards product management and planning, leading to increased output and capacity. The conversation also touches upon the potential disruption of legacy software providers by AI and the accelerating cycle of software renewal. AI has altered the "buy vs. build" equation, making in-house development more feasible for smaller needs. The concept of "forward deployed engineers" is introduced as key to rapid AI adoption. Integration remains vital, especially for systems of record, with AI enabling outward-reaching automation. Novel AI applications include rapid cloud migration and intelligent travel assistants. The podcast addresses the practicalities of managing AI, including "token budgets," compute allocation, and the importance of centralizing access for monitoring and optimization to shield users from "token anxiety." Security protocols at Goldman Sachs prevent unapproved software installations, but AI characteristics are integrated into their systems. The discussion differentiates between speed and velocity in AI adoption, prioritizing security and scalability. Regulatory discussions in banking are framed by existing experience with AI, focusing on risk management. AI-generated code requires rigorous human approval and security checks. While AI can provide significant insights, banks believe clients still value human expertise and proprietary data. AI is also reshaping talent requirements, emphasizing ideation, delegation, and supervision, leading to a metamorphosis of the developer role. Despite potential concerns about burnout from repetitive tasks, AI is also bringing renewed joy and efficiency to the profession, driven by peer influence and automation of tedious work. Key takeaways include regulators focusing on risk management, the engineering challenge of token budget optimization, and the ongoing evolution of software development towards AI-assisted coding.
Outlines
Business Software Challenges and Pipedrive CRM
Businesses struggle with complex software like CRMs, leading to inefficiency. Pipedrive is presented as a user-friendly CRM solution for small to medium businesses, designed to streamline sales processes and improve customer information management for faster deal closure.
The Accelerated Impact and ROI of AI
The rapid advancement of AI, particularly since late 2022, has drastically accelerated timelines, moving from experimentation to potential disruption of legacy software. The focus is shifting to evaluating AI's return on investment, with potential for internal coders to replace third-party SaaS contractors.
Evolution of Personal AI Usage and Goldman Sachs' AI Integration
Initial AI usage, like with ChatGPT, involved experimentation. Now, expanded capabilities, especially in areas like Cloud Code, enable practical applications. Goldman Sachs has integrated AI extensively, with their GSAI assistant used daily by 47,000 employees, impacting client experience and internal operations.
Data Quality, Agentic AI, and Developer Role Transformation
Data quality is crucial for effective AI. Goldman Sachs emphasizes curated data for better AI outcomes. Agentic AI, including tools like Devin, is fundamentally changing developer roles, shifting them towards product management, planning, and idea generation, leading to increased output.
AI's Disruption of Legacy Software and the Buy vs. Build Equation
AI is accelerating the software renewal cycle and potentially disrupting legacy software providers. It has significantly altered the "buy vs. build" equation, decreasing the time and cost to build simple applications, leading to more in-house development.
Forward Deployed Engineers, Integration, and Novel AI Applications
"Forward deployed engineers" accelerate AI adoption by working directly with AI model providers. Integration remains crucial, especially for systems of record. Novel AI applications include rapid cloud migration and building integrated corporate travel assistants.
Managing Token Budgets and AI Costs
Organizations are grappling with "token budgets" and compute allocation for AI. Centralizing access through a "Model Gateway" is key for monitoring, metering, and optimizing token usage, balancing quality and cost to shield users from "token anxiety."
Security, AI in Banking, and Evolving Talent Needs
Strict security protocols govern AI installation and code approval at Goldman Sachs. Regulatory discussions focus on risk management. AI is shifting talent requirements towards ideation, delegation, and supervision, fundamentally changing developer roles and emphasizing managerial-like skills.
Keywords
CRM (Customer Relationship Management)
Software systems designed to manage and analyze customer interactions and data throughout the customer lifecycle, aiming to improve business relationships, assist in customer retention, and drive sales growth.
AI (Artificial Intelligence)
The simulation of human intelligence processes by machines, especially computer systems. This includes learning, reasoning, problem-solving, perception, and language understanding.
ChatGPT
A large language model developed by OpenAI, known for its ability to generate human-like text, answer questions, and perform various natural language processing tasks.
Agentic AI
AI systems designed to act autonomously to achieve specific goals. These agents can perceive their environment, make decisions, and take actions independently.
Cloud Code
Refers to AI-powered coding assistants or tools that operate in the cloud, helping developers write, debug, and optimize code more efficiently.
Forward Deployed Engineer
A role involving direct collaboration with AI model providers to integrate and optimize AI solutions within a company, often bridging the gap between AI development and practical application.
Token Budget
A financial allocation for the usage of tokens, which represent computational units consumed by AI models, particularly large language models, for processing and generating text.
Model Risk Management
A framework within financial institutions to identify, assess, and mitigate the risks associated with the use of AI and machine learning models, ensuring compliance and operational integrity.
Information Barriers
Security protocols implemented in financial institutions to restrict access to sensitive information between different departments or systems, preventing data leakage and ensuring regulatory compliance.
Velocity (in AI adoption)
Refers to sustained, directed progress in AI adoption, emphasizing security, scalability, and long-term strategic direction over rapid, potentially unmanaged, bursts of activity.
Q&A
What are the main challenges businesses face with traditional CRM software?
Traditional CRMs are often overly complicated, packed with unused features, and have clunky interfaces. This leads to teams spending excessive time searching for basic information, hindering efficiency and productivity.
How has the timeline for AI adoption and impact accelerated?
The release of ChatGPT in late 2022 marked a significant acceleration. What was once an age of experimentation has rapidly evolved, with AI now seen as a potential disruptor of legacy software companies.
What is the role of data quality in AI implementation at Goldman Sachs?
Data quality is considered the primary determinant of AI effectiveness. Goldman Sachs emphasizes curating and making data understandable to AI, using platforms like Legend AI to ensure accurate and reliable outputs.
How is Agentic AI changing the role of software developers?
Agentic AI is shifting developers' roles from pure coding to becoming more like product managers, planners, and idea generators. The focus is on explaining requirements, delegating tasks to AI agents, and supervising their output.
How does Goldman Sachs manage the cost and usage of AI tokens?
Goldman Sachs centralizes AI access through a "Model Gateway" that intelligently routes requests to optimize for both quality and cost. This prevents "token anxiety" for users and allows the central team to manage economic viability.
What is the significance of "forward deployed engineers" in AI adoption?
Forward deployed engineers work directly with AI model providers, acting as product specialists who help integrate and optimize AI solutions. They bypass intermediaries, accelerating adoption and leveraging rapid advancements in AI technology.
How does Goldman Sachs address regulatory concerns regarding AI?
Goldman Sachs frames AI discussions with regulators similarly to previous technological advancements, focusing on risk tiering, controls, and human supervision rather than deep model explainability. They have established model risk management functions.
How is AI impacting the traditional profitability sources for banks?
While AI can provide significant insights, banks believe clients value the "extra 10%" derived from proprietary data, cross-asset correlations, and global relationships, which AI alone may not fully replicate.
What are the key skills required for employees working with AI agents?
Employees need to be able to explain their desired outcomes, delegate work effectively to multiple specialized agents, and supervise the agents' outputs. These skills are akin to managerial competencies.
How does AI contribute to reducing developer fatigue?
AI automates repetitive and tedious tasks, such as library upgrades or code refactoring. This frees up developers to focus on more strategic and engaging aspects of their work, like planning complex projects and ideation.
Show Notes
When we last spoke to Marco Argenti, chief information officer at Goldman Sachs, we were talking about how the bank was deploying AI, including the development of its own internal tools. But that was a year and a half ago and a lot has changed since then, especially with the arrival of agentic platforms like Claude Code. So what exactly is Goldman Sachs doing with AI now? And what has its experience with the new tech been like so far? On this episode, we catch up with Marco to discuss what AI deployment at the bank actually looks like at the moment — including how AI coding is changing the work of its developers and engineers — to all the data challenges and regulatory concerns that come with integrating this technology at scale.
Subscribe to the Odd Lots Newsletter
Join the conversation: discord.gg/oddlots
See omnystudio.com/listener for privacy information.
Table of contents
United States
