Andrew Ng advocates sandbox-first approach to enterprise AI development

Andrew Ng, founder of DeepLearning AI, advocates for a “sandbox first” approach to enterprise AI development that prioritizes rapid experimentation over early implementation of guardrails and safety measures. Speaking at VB Transform, Ng argued that while observability and safety controls are important, applying them too early in the development process can stifle innovation by forcing engineers to seek approval from multiple executives before testing new ideas.

The big picture: Ng’s methodology flips conventional enterprise AI wisdom by encouraging companies to build and test AI applications in isolated environments before investing in comprehensive safety infrastructure.
• “I frankly tend to put those in later because I find that one of the ways that large businesses grind to a halt is that for engineers to try anything, they have to get sign off by five vice presidents,” Ng explained.
• Sandboxes allow developer teams to “iterate really quickly with limited private information” while protecting sensitive company data.
• Only after projects prove viable should organizations add observability tools and guardrails to make them production-ready.

Why this matters: The approach addresses a critical tension between AI innovation speed and enterprise risk management, as companies struggle to balance rapid development with brand protection and data security concerns.
• Ng acknowledged that big businesses “can’t afford to have some random innovation team ship something that damages the brand or has sensitive information,” but noted this caution can also hamper innovation.
• The methodology allows organizations to invest resources only in projects that demonstrate clear value before adding expensive safety infrastructure.

Speed advantage: Development tools like coding agents have dramatically accelerated AI project timelines and reduced costs, making rapid experimentation more feasible.
• Ng cited coding agents like Windsurf and GitHub Copilot as cutting development time for “projects that used to take me three months and six engineers.”
• The reduced cost of proof-of-concept projects means companies can afford to run multiple parallel experiments: “I don’t feel like the cost of a proof of concept going so low that I’m fine to do a lot of POCs is bad.”

The talent challenge: While foundation model engineers command salaries up to $10 million, the bigger issue is finding enough experienced AI application developers for enterprise projects.
• “One of the biggest challenges for many businesses is talent,” Ng said, noting that application developers don’t command the extreme salaries of foundation model specialists.
• His solution circles back to sandbox experimentation: allowing engineers to gain AI project experience through hands-on development in safe environments.

Industry context: The sandbox approach aligns with broader enterprise AI trends, as companies like Salesforce enhance agent observability tools while innovation teams seek ways to accelerate development cycles.
• Salesforce recently updated its Agentforce 3 agent library to provide enhanced visibility into agent performance and support for interoperability standards.
• Innovation sandboxes for AI agents have become increasingly common as enterprises balance creative experimentation with operational security requirements.