This is a MarketScale Software & Technology Podcast Series, hosted by Daniel Litwin. This is the first episode of a three part series titled Bringing AI to Businesses with Ben Taylor, Chief AI Officer & Co-founder of ZIFF Inc. In each episode, we’ll explore different aspects of AI’s push into business-operations ubiquity, from its most useful applications to the surprising business ethics that come with implementation. Each episode will also feature a short article from Taylor, which you can read below.

How Do You Deliver on an AI Project, Both As an Executive and a Data Scientist?

Many executives are naive when it comes to AI capabilities and navigating where AI might provide value to their business. The data scientists aren't helping either, where most struggle to communicate value to the business representatives. Most data scientists also lack urgency; they have no pressure to last-mile AI into production. Funding science projects will accomplish one thing: covering the tuition of your data science team so they can land better jobs at Facebook or Google. So, once the trigger is pulled and you have a team prepped and passionate about bringing AI to your business...how do you make sure everyone delivers so you avoid wasted time and money?

Avoiding The Science Project Landmines

If you are an internal advocate for AI, do everything you can to constrain the timeline. Ask yourself: Is there anyway to do an internal proof of concept in 60 days instead of 12 months? What can I do to reduce internal budget? What can I do to reduce the number of people required? The more you reduce these variables, the more likely you are to get buy-in from the internal business units.

I've always been a fan of leveraging outside hardware companies, consulting groups, or AI platforms to shorten do-ability tests.

Crawl, Walk, Run

Some AI projects fail because they are too ambitious. They don't have a short-term proof-point, and the complexity comes tumbling down like a house of cards, revealing a project that had no clear goals, tangible value or structure. This flaw can come from inexperienced data science teams that are too "academic," where they are more interested in a challenging thought experiment than a Bayesian method in production. If you can carve your project up into bite size milestones, your chances of success are higher. It shouldn't be ignored that AI projects aren't a one-and-done either; you have the advantage of improving on your algorithms. Just look at the evolution between AlphaGo and AlphaGo Zero, and how they would've never achieved such a grand level of "reinforcement learning" without trying a few, more tangible methods first. Get some novice wins into production and then level up on subsequent versions.

It Is Harder Than It Looks

Getting a successful AI project to value is much harder than it looks. Most major wins for AI are behind six to 10 iterations on the same problem. We see successful companies solving the same problem multiple times, where each time they solve it they understand the data set and problem a little better. Once a project has crossed a predefined criteria for success, taking that AI project into production can create additional problems. Supporting AI in production requires quality monitoring (e.g. did your incoming features drift) to ensure models are behaving as designed. This requires an involved data science team. And yes, I said team. Collaborate, get multiple eyes on the project, and make sure everyone is on the same page before launching something into production. You don't want your AI project to end up like the Mars Climate Orbiter: dead in the air because of a unit conversion mistake. Double check, triple check, and then check again that the final product is in line with the initial vision you set up for success. Feels a lot like simple project management, huh?

Highlights from the Episode

[transcription]

For the latest news, videos, and podcasts in the Software & Electronics Industry, be sure to subscribe to our industry publication.

Follow us on social media for the latest updates