Steven Sinofsky (board partner at A16Z), Aaron Levy (CEO of Box), and Martine Casado (general partner at A16Z) discuss the growing divide between AI capabilities in Silicon Valley and enterprise deployment realities.

The conversation explores why AI adoption succeeds in engineering teams but struggles in large organizations, examining integration challenges, architectural decisions, and the future of human-AI collaboration.

Key topics include Salesforce's headless API strategy, agent identity management, the productivity paradox of AI coding, and why historical predictions like those in The End of Work about technology eliminating jobs have proven consistently wrong.

The Silicon Valley-Enterprise AI Divide

Engineers have "insanely high" technical aptitude and can debug systems when they break, while enterprise users lack these capabilities for AI workflows

"The board goes to the CEO. What does the board say? We need more AI. And what does the CEO say? Oh, okay, I'll get like a consultant to do more AI" - centralized projects fail without operational alignment - Martine

AI model speed creates architectural paralysis as enterprises fear choosing the wrong technology stack after previous AI failures

Enterprise Integration Reality Check

"Any enterprise of a thousand people or more, or that's older than 10 years, is just a mass of stuff that's sitting there waiting to be integrated" - Steven

Agents face the same access control limitations as humans - they get "bounced to another human" when systems don't integrate properly

Legacy environments lack authoritative access controls, requiring humans to ask "Sally" or "Bob" for information that agents cannot navigate

OpenAI's partnerships with Accenture and Deloitte reflect the massive systems integration work required for enterprise AI deployment

Agent Architecture: Human-Like vs Software-Like

"Instead of viewing AI as software, just view it as a user" - agents should use CLI tools and existing human interfaces rather than specialized APIs - Martine

Agents need individual email addresses and permissions like employees, not shared credentials, for proper security and access control

Computer use agents work better with actual browsers than headless versions because websites have anti-scraping measures that detect automation

"Agent onboarding" should mirror human processes - orientation, department pitches, and cultural training for proper organizational integration

Salesforce Headless Strategy and API Evolution

Salesforce's headless API announcement signals agents will be licensed users consuming software at 100-1000x human scale

Current APIs designed for human typing speeds must evolve for agents that "fan out and do that work in a way that I can't as a person in a browser" - Aaron

Box agents search across entire environments with multiple queries and re-ranking, removing human limitations of single query workflows

SaaS systems risk collapse from 500x usage increase unless architected for agent-scale throughput and caching

AI Productivity Paradox in Practice

Box achieves 2-3x productivity gains with AI coding but remains "rate limited by how quickly can you review this stuff and check on the work" - Aaron

"AI built probably 80 to 90% of the feature" but security reviews still constrain release timelines due to code injection risks

AI coding may create "as many problems as you are solutions" with growing entropy that organizations struggle to manage - Martine

Some companies incentivize AI usage by "counting tokens," leading employees to have "agents do useless tasks" for metrics

The Jobs Expansion Thesis

The End of Work predicted job elimination in the 1990s just before the internet boom created massive employment growth

"The more code we write, the less we would need engineers, it'd be the opposite because now your systems are even more complex than before" - Aaron

AI-native companies are "hiring like crazy" while infrastructure companies see increased demand from more software being written

Engineering jobs will expand beyond tech companies to John Deere's automated tractors, Caterpillar's AI systems, and Eli Lilly's pharmaceutical design

Lawyers now all use "computerized" workflows with internet citations and track changes, yet there are "way more lawyers today than there were 30 years ago"