Artificial intelligence (AI) isn’t a single technology; it’s a term that covers a field of computer science in which machines learn to work and act like humans. Within AI there are numerous variations, and in my mind, the greatest potential for businesses today lies in narrow AI. Quite simply because that’s where the technology is at.

I’m increasingly talking to and working with business leaders who are deciding how to dip their toes in the AI waters. And a question that almost always comes up is, “Where should I start?”

Having spent many years in the research labs, my unhesitating answer is, “In the real world, where you want to deliver value.”

First: what is narrow AI?

Frank Chen, partner at Andreessen Horowitz (which is, in full disclosure, a Drishti investor), discussed the three stages of AI at the 2017 a16z Summit. In that presentation (which is worth a full watch, when you have a chance), Frank outlines the definition of “narrow AI” (also known as “weak AI”) as designed to assist with or take over specific tasks; essentially, it’s AI that’s dedicated to a single purpose.

He contrasts that to “broad AI” (also called “strong AI” or “general AI”), which transfers knowledge across domains to function more like (or, arguably, better than) a human, and does not yet exist today.

Here’s an example: Narrow AI is used to recognize anomalies on an MRI scan, but it won’t assess the patient’s medical history and set up a course of treatment to address that anomaly – that’s where a human doctor is still critical. At the 2018 International Joint Conference on Artificial Intelligence, Facebook’s AI Chief Yann LeCun delivered a keynote in which he agreed with “current approaches in applying narrow AI to autonomous vehicles, medical image processing and translation, etc. However, such AI remains incapable of making real deductions, working as a true smart assistant, or finally achieving the holy grail of artificial general intelligence (AGI).”

When you deploy AI today, you’ll be deploying narrow AI. By recognizing that fact, you’re already taking your first step to setting realistic expectations.

How do manufacturers best benefit from narrow AI?

The success of AI implementation in a company will ultimately depend on how its executive driver thinks about deployment, value measurement and success. (This is typically the Chief Innovation Officer, Chief Artificial Intelligence Officer or even the Chief Information Officer; though others can be involved, depending on the company.) CEOs count on these innovation drivers to review the solutions available and put the best ones to work, learning what is effective and, just as importantly, what isn’t.

You’ll notice I said “put the best ones to work,” not “research the best ones in a corporate lab environment.”

That’s an important distinction: One of an Innovation Officer’s most important contributions is to bypass the lab and take AI solutions to the valuation point.

Why? Because labs don’t reflect reality. They lack the variability that makes the real world so challenging (and presents so many opportunities to create value). In fact, labs are controlled specifically to eliminate variability. And, if I were to take a contentious stand, labs slow down the progress third-party technologies often make because of the persistent influence of NIH: Not Invented Here.

The only way to discover the true value of an AI solution is to have users poke, prod, iterate and discover how it works best on the floor. Any AI solution designed for the lab will collapse in the real world.

What’s the path to value with narrow AI?

Innovation Officers and others who are keen to deploy AI need to be collaborative, tolerant of failure and, most importantly, willing to learn with an open mind. AI is still emerging; as such, naturally, many of its applications and uses are yet to be discovered.

My best advice is to plan, measure and iterate:

  • Plan the experiment: Try it in three places, with three different teams. When you deploy the technology for the first time, it’s critical that you methodically lay out the components of your experiment before you start experimenting. I’ve seen AI deployments where human emotions and organizational biases can muddy the results. Run them blind or with an open information flow, but set the experimental parameters and success metrics ahead of time.
  • Measure: While the measurable signals might not always be obvious, I’d urge you to measure any relevant signal – with a maniacal focus. Ideally, so it lines up with your experimental hypothesis. In the worst case, even if you don’t know yet what you’ll do with that information.
  • Expect improvement, but not perfection: It’s important to keep in mind that AI is not, and will likely never be, 100 percent accurate. By measuring whatever you can, you’ll not only understand how to design your processes around AI, but also where AI needs to be backstopped by other kinds of technology or tools. Remember that man–machine systems are often unbeatable!

AI is giving us more data than ever before. Drishti’s data guy, Sameer Gupta, wrote a post on how to work with massively greater volumes of data.

The potential of AI is exciting, and for those of us lucky enough to be working with the technology day in and day out, narrow AI is particularly promising. Look for more companies to embrace narrow AI in the coming years, but more importantly, keep your focus on the ones that can deliver results that impact your business.

Read how we’ve deployed narrow AI to demonstrate high value improvements in manufacturing metrics.