‘Sandbox first’: Andrew Ng’s blueprint for accelerating enterprise AI innovation

Companies can be concerned with the impact of AI applications when they are put into production, but hindered these projects with railing at the start could slow innovation.

Andrew NG, founder of In depth IA And one of the most important figures in AI development, stressed the importance of observability and railings in the development of AI during a cat by the fire in VB transform Today. However, he added that they should not be done at the cost of innovation and growth.

NG has suggested that companies are built in sandboxes to prototyper projects quickly, find the drivers that work and start to invest in observability and railings for these applications after their opinion. This may seem countertiintable for companies that seek to implement AI.

>>See all of our Transform 2025 coverage here<

“There is an important role in observability, security and railings,” said NG. “I frankly tend to put them later because I find that one of the ways in which large companies stop is that for engineers to try everything, they must be signaled by five vice-presidents.”

He added that large companies “cannot afford to have a random innovation team to harm something that damages the brand or has sensitive information”, but this can also hinder innovation.

Instead, NG said that sandboxes offer developers’ teams “to iterate very quickly with limited private information”. The sandboxes allow the organization to invest only in operating projects, then add the technology to make them responsible, including observability tools and railings.

It is not uncommon for companies to establish innovation sand bins, especially for AI agents. Sandboxes Allow innovation Within the limits of companies without affecting any sensitive information, they do not want to be public. However, they also allow teams to be as creative as possible to test ideas.

Observability quickly becomes a key subject, because many AI applications and agents come into production. Dirty Recently updated his agent library, Agentforce 3, for Provide improved visibility In the performance of agents and additional support for interoperability standards, such as MCP.

Speed ​​and drop in pilot costs

For NG, speed and innovation go hand in hand, and companies should not be afraid.

“Imagine that we were on roller coaster, but it is a slow roller coaster. What happened last year, our roller coaster just picked up a lot of speed, and it is really exciting because it goes forward,” said NG. “I have the impression that the world is now on very fast roller coaster, and it’s great.”

NG said that a factor that has contributed to this speed is the tools now available to developers at work and ideas quicklystressing that coding agents as Windsurfing And Github co -pilot reduced the development time “projects that took me three months and six engineers”.

These coding agent platforms and other tools that help developers move faster have also meant the cost of carrying out pilot projects.

“I do not have the impression that the cost of proof of concept takes place so low that I am good to do a lot of POC (proof of concept) is bad,” he said.

A barrier

A barrier, however, could be to find talent. NG has recognized that there are IA companies recruiting foundation model engineers with salary ranges of up to $ 10 million, but the price is not so high for software engineers.

“One of the biggest challenges for many companies is talent,” he said. “The good news for companies looking for engineers capable of creating requests, the price is far from the $ 5 million range,” he said.

The problem, however, is that there is still not enough talent that are experienced in the construction of AI projects for companies. Thus, NG dates back to its first solution: let them experiment in sandboxes and acquire this experience.