Next-Level AI Products and Complementary Perspectives

The Data Science team was worried about overfitting, but the Product Manager wondered if some of the Machine Learning (ML) “noise” was real insight related to subgroups of users. After a Topological Data Analysis spike, two subgroups with distinct characteristics were identified and tuned recommendations for those two subgroups led to an additional $79 million in annual sales.

I'm bombarded with recommendations for AI training. Out of curiosity, I looked at a few of the trainings last week and was reminded of something that I’ve learned over the last two decades of AI Product Management –

To create great AI products, Product Managers need to be careful not to look at AI in exactly the same way that most Engineers and Data Scientists do.

In the example above, the Data Science team is worried about overfitting, while the Product Manager is worrying about missing opportunities to better serve subgroups. They’re both right. The Data Science team needs to worry about variance and bias in their models, and the Product Manager needs to worry about an end product that exceeds user experience and revenue goals. Those two perspectives work together to create great AI products. If the Product Manager has an identical perspective to the Engineering and Data Science teams, then the resulting AI products are unlikely to be differentiated or deliver optimal value, as I’ll illustrate further below.?

First, here are a few of the complementary perspectives that I’ve found yield great AI products:

Here’s an example from one of the AI courses that I took a peak at last week. It was a course by Souvik Ghosh , Director of AI at LinkedIn. In the introduction, the following statement was made:

“In the years to come, we expect AI to be able to solve problems in Agriculture, like utilizing resources on land and water.”

The rest of the course focuses on machine learning, regression, decision trees, and neural networks. What if we collected all of the historical agricultural data from around the world and used these techniques to try to make the best use of resources on land and water??

Hmmm.

There’s a fundamental problem.?

The vast majority of the world’s agriculture depletes soil nutrients, kills soil biota, and reduces the soil’s ability to hold water. In other words, the most pervasive agricultural system works against our stated goals of best “utilizing resources on land and water.”

As Product Managers, we would first have to ask “What are the agricultural systems that make the best use of land and water?"

• No-till agriculture.?
• Permaculture.?
• Food forests.?
• Alegria Farms Soxx system.?
• The similar forest-canopy-suspended system that I’ve been working on.?

Broadly these systems can be grouped together as Regenerative Agriculture.?

As good Product Managers, we’d want our core models to learn from these Regenerative Agriculture systems. Consequently, we’d insist that 99.9+% of the world’s agricultural data be excluded from core model training because that data is from agricultural systems that undermine our stated goals. Instead, we’d point at tiny and fragmented data sets that only cover a handful of crops and that only represent a miniscule fraction of the world’s growing regions. What could we do to close those data gaps? Well, the 99.9+% of agricultural data that we excluded from core model training might be useful to train Generative AI models to create synthetic data for missing crops and regions. So we wouldn't ask that the majority of historical agricultural data be thrown away. We’d just ask our Data Science and Engineering teams to use that data in a different way than they might have initially intended.?

Can you see how the Engineering/Data Science and Product Management perspectives complement each other?

By focusing on Regenerative Agriculture, we’d uncover numerous highly-differentiated AI product opportunities. For example, automating and scaling the harvesting in permaculture farms and food forests is going to require very sophisticated AI-driven machine vision and autonomous robots because of a non-uniform environment, a high crop mix, and the requirement for minimal disruption--imagine tiptoeing harvest-bots running convolutional neural networks (CNNs) for vision and recurrent neural networks (RNNs) for prediction and context building. This kind of harvesting automation makes staggered micro-harvests possible which further improves food quality and reduces waste, as long as we develop AI products to redesign and optimize food transportation and distribution. And on and on it goes. Good questions spawn more good questions which lead to revenue growth, customer delight, and competitive advantage.?

Our responsibility as AI Product Managers isn’t to be the best mathematician or statistician on the team. Our Data Science and Engineering teams can and should run circles around us. But it’s our responsibility to have the broadest view, the deepest customer empathy, the most relentless focus on delivering value, and to be outstanding systems thinkers. We need to know the strengths and shortcomings of various AI techniques and always be eager to learn about new ones.

I hope sharing a little of my experience and perspective is helpful. Please add examples and insights of your own.?Can't wait to see your amazing AI products!