While your company debates whether to monitor every developer's AI prompt, Goldman just built a playbook for what actually scales.
The Summary
- Goldman Sachs rolled out AI tools to its 12,000 engineers and built an in-house ChatGPT, but CIO Marco Argenti refuses to track individual usage metrics
- JPMorgan monitors dashboards of AI user activity and Meta tracks keystrokes, but Argenti says this approach "misses the forest for the trees"
- Goldman measures team velocity: how fast engineers move from idea to production-ready code
- Argenti describes AI-enabled development as "3D printing" software, generating prototypes in real time
The Signal
The AI adoption measurement war is splitting corporate America into two camps. One side is counting keystrokes. Meta is installing software on US employee computers to track mouse movements and keystrokes to train its AI. JPMorgan built dashboards showing tens of thousands of users' AI activities so employees can compare themselves with peers. The other side, led by Goldman's tech chief, is asking a different question entirely: are we shipping faster?
Marco Argenti oversees 12,000 engineers at Goldman Sachs. He has access to the same granular usage data his peers are obsessing over. He just doesn't care about it.
"Tracking individual AI usage misses the forest for the trees."
Instead, Argenti is focused on team velocity. How long does it take Goldman's engineering teams to go from an innovative idea to production-ready software? That's the only metric that matters in his framework. Not how many prompts an engineer writes. Not whether they're using Copilot more than their desk neighbor. Whether the code ships.
This isn't just philosophical preference. Goldman built its own in-house version of ChatGPT and deployed AI tooling across its entire engineering organization. They have the full dataset. Argenti's choice to ignore individual metrics is deliberate. He's betting that the unit of measurement for AI productivity isn't the developer. It's the team.
Key differences in approach:
- JPMorgan: individual dashboards, peer comparison, activity tracking
- Meta: keystroke and mouse movement monitoring for AI training data
- Goldman: team velocity from concept to production deployment
Argenti describes the shift as moving to "3D print" software. Engineers generate prototypes in real time. The metaphor is precise. 3D printing didn't make individual CAD designers 10% faster at drawing lines. It collapsed the entire path from design to physical object. That's what AI tooling is doing to software development, and measuring it requires different instruments.
The contrast with Meta's approach is stark. One firm is monitoring keystrokes to generate training data for future AI models. The other is measuring cycle time to evaluate whether current AI tools are working. One is building the future product. The other is measuring the present impact.
The Implication
If you're a tech leader trying to justify AI spend to the CFO, you have two paths. You can present a dashboard showing 87% of developers used Copilot last quarter. Or you can show that your teams are shipping features in half the time they did a year ago. One number is a usage stat. The other is a business outcome.
Watch what Goldman does next with velocity metrics. If they're seeing meaningful compression in development cycles, they'll scale AI tooling aggressively. If they're not, they'll kill it regardless of adoption rates. That's the actual test. The keystroke counters are measuring activity. Argenti is measuring results.