AI Inside the Enterprise

What It Covers

Steven Sinofsky, Aaron Levy, and Martin Casado examine the widening gap between AI capabilities in Silicon Valley and actual enterprise deployment. They analyze why top-down AI mandates fail, how integration bottlenecks stall transformation, why agents function more like new employees than software, and what the realistic productivity timeline looks like for large organizations.

Key Questions Answered

• •Top-Down AI Mandates Fail: When boards pressure CEOs to "add AI," the typical response is hiring consultants to run centralized projects that lack operational alignment. These initiatives consistently fail because they bypass the people doing actual work. Enterprises should instead identify where individual employees are already using AI effectively and scale those organic workflows outward, rather than imposing centralized programs disconnected from daily operations.
• •Integration Is the Real Bottleneck: Any organization with over 1,000 employees or more than ten years of history carries accumulated legacy systems that AI cannot automatically connect. Agents hitting access control walls cannot improvise workarounds the way humans do — they cannot "ask Sally" for a file or "call Bob" for a number. Enterprises must audit and modernize data permissions and system access before deploying agents into consequential workflows.
• •Treat Agents Like New Employees, Not Software: Rather than building complex API integrations, enterprises should provision agents with their own identity, email address, and role-based access permissions — mirroring human onboarding. This approach drafts on forty years of existing access control infrastructure designed for human users. Agents given human-equivalent permissions inherit established governance frameworks instead of requiring entirely new technical architectures.
• •Architecture Paralysis Slows Enterprise Adoption: Enterprise AI teams are stalled debating agent orchestration paradigms — whether to run agents in-cloud or locally, which model provider to commit to, and how to handle tool access. Organizations burned by deprecated AI investments three to four years ago are reluctant to commit again. Practical mitigation: start with read-only, information-retrieval agents that carry lower architectural risk before building agents that take consequential actions.
• •AI Expands Complexity, Which Sustains Engineering Demand: The premise that AI-generated code reduces the need for engineers inverts the actual dynamic. More code means more complex systems, which generates more upgrade cycles, security incidents, and downtime events requiring human expertise. Historical precedent supports this: computerized accounting created more accountants, not fewer. Engineers at non-tech companies — John Deere, Caterpillar, Eli Lilly — represent the next large wave of software engineering job growth.
• •Productivity Gains Are Real but Constrained at 2–3x: Box reports AI contributes roughly 80–90% of new feature code, but release velocity remains gated by mandatory security reviews and code review processes. The realistic enterprise productivity gain is approximately 2–3x, not the 5–10x figures circulating in Silicon Valley. The rate-limiting factor shifts from writing code to reviewing, validating, and safely deploying it — meaning human oversight capacity becomes the new constraint to optimize.