Why AI projects get stuck in testing hell

As companies struggle to implement AI in the enterprise, one of the things that tends to happen is letting perfect get in the way of good enough. One of the problems of working with AI as an individual, team or a company is that it’s easy to go down rabbit holes and get stuck in testing hell. And you can waste the productivity gains AI is supposed to deliver when this happens.

Earlier this month at the HumanX conference in San Francisco, I hosted a panel on enterprise AI readiness. I was joined by Atlan CEO Prukalpa Sankar and Brad Menezes, CEO at Superblocks. We discussed the challenges that companies are facing as they try to implement AI.

One way to go faster and break out of the testing cycle is by having a strong foundation in place. That involves having your data house in order because the old maxim, garbage in/garbage out still applies. It also means having the right governance structure in place to make sure people are building applications in a safe and responsible way.

Sankar, whose company has built a data governance platform that helps ensure there are rules and guidelines around how the data is used, says that it’s easy for companies to get bogged down in the perfection cycle. One of her customers spent 12 months trying to get an AI analyst to give the same answer every time to the same question and got lost in what she calls “testing hell.”

“There's no north star. I don't know what 100% accuracy is. And so then you're just stuck in this loop,” she said. So what does success look like if that’s the case? “So the best teams we work with have AI use case teams that are focused on shipping and getting to value, but they're also doing it with the right foundation,” she said.

“So when I ship this, I make sure my context is portable. I make sure that I'm shipping with the right tooling infrastructure in the background so that the chances of something crazy happening are lower.”

Sankar uses the classic salary data example to show the importance of restricting access to certain data in an organization. "I'm OK with my HR team shipping an app on payroll data that only they can use. But I don't want anyone else in my company to use that because it could spit out my CEO's salary to a random person," Sankar said.

Menezes agrees that a firm foundation is critical. Otherwise, IT is going to step in and shut things down to prevent something like this from happening. "It always starts with the excitement and the exuberance of how great the technology is, but enterprises have a different problem, which is, how do they do this without taking risks? And I think that's where the conversation is today," he said.

That requires some structure for the folks building AI applications. Superblocks is a startup focused on building governed applications using AI. "If you want to go fast, you need a seat belt. And so every organization out there is trying to figure out what's the right seat belt for their use case," he said. "Once they have that in place, they can go 100 miles an hour down the freeway. Otherwise, there's a stop sign at every [corner]."

The faster you can move, the faster you’ll get to value, but it takes upfront work to make that happen. Otherwise, you get slowed down by the mechanisms every organization has in place to protect themselves, or you waste a lot of time trying to build them. As Sankar says, you need speed to gain speed and you can’t rely on the old way of doing things.