Podcast: You Have an Idea for a New AI Tool. Now What?

“Just saying, like, ‘Oh, like, we can use this to generate more content, more information,’ maybe that makes sense,” says Rahman, an associate professor of management and organizations. “But for a lot of organizations, you actually need less generation of emails and less overall information.”

Like any other product, Rahman and Gerber say that AI tools aren’t ready until they’ve gone through the traditional product-development process. In this episode of The Insightful Leader, we linger on its key parts.

Podcast Transcript

Laura PAVIN: You’re listening to The Insightful Leader. I’m Laura Pavin. We have reached a place with artificial intelligence that has created this sort of corporate existential dread. Like, if you don’t get with it right now, you’ll be out of the job or get left behind.

Liz Gerber has been helping companies think about innovation and product development for years. And the energy around generative AI is giving her some déjà vu. It reminds her of what happened in the early days of the internet.

Liz GERBER: Everybody wanted to get into online advertising. And so people put up these really big, gaudy ads. The flashing, the rainbows, et cetera. People didn’t really know what they were doing or what they were trying to do. They just knew they wanted to be in it.

PAVIN: It was weird, directionless, and a big ol’ waste of money. For many companies, this moment feels like that. A rush to nowhere.

Gerber, a professor at Northwestern, spoke to us at an Insightful Leader Live event about how leaders can make AI actually work for their businesses. She spoke alongside Kellogg professor Hatim Rahman. Both Gerber and Rahman research AI and its impact on work.

And they told us that the key to making AI do something useful for you is to lean on knowledge you definitely have already—because it’s just good, old-fashioned product development! You know, where you ask questions like: What do people really want to use this for? Who is this going to affect? And how are you going to troubleshoot problems, not if, but when they come up?

So think of today’s show as more of a refresh on the product-development process—with an AI bent. We’ll linger on parts of the process most likely to trip you up, so you can avoid them.

Alright? Okay, we’ll dive in next.

...

PAVIN: Before I do anything else, I want to first acknowledge that I get it. I get why some businesses are launching themselves a little too quickly into AI. No one wants to be the last to adapt or capitalize on something so powerful.

But it’s worth slowing down to make sure you aren’t needlessly pouring resources into something you haven’t fully conceptualized.

So let’s start by backing up. You’ve just come up with an AI-powered product idea. The next thing that Gerber and Rahman say you should do is ask yourself, “what problem does this thing actually solve?” If more businesses were asking themselves that, by the way, then they wouldn’t all be doing this one thing.

Hatim RAHMAN: We’ve known now for a long time that we have too many emails and too much information.

PAVIN: Kellogg professor Hatim Rahman.

RAHMAN: And so just saying, like, “Oh, like we can use this to generate more content, more information,” maybe that makes sense. But for a lot of organizations, you actually need less generation of emails and less overall information.

PAVIN: Brainstorm a bit longer, and find a problem that really needs solving.

Rahman says he was looking at a hospital system that was developing an AI-powered ultrasound probe. The goal was to improve maternal–fetal outcomes—especially in countries with sparse access to clinical care.

RAHMAN: So the idea is that they would train and develop an AI model so that anyone with a handheld ultrasound probe, they will be able to scan the fetus, and then the AI model would give you relevant information: What is the expectancy date? Are there any health complications that are being picked up by the AI model’s predictions such that you should go and take the costs and the expense to go perhaps seek more medical care?

PAVIN: In this case, Rahman said, the answer to the question, “what’s the purpose of this?” is pretty clear, right? If you’re a pregnant woman with limited resources, you want to make sure you’re only spending on the care that’s most necessary. The probe would help you make those decisions. From a product-development standpoint, the actual need for this device is pretty clear. So, we can check that off our list and move on.

The next important question you should be asking yourself about a potential new AI product is, “how does this fit into existing processes?” Specifically, “how does it fit into the humans’ processes?”

Northwestern professor Liz Gerber, again.

GERBER: I’ve been working on many projects with organizations to step back and just say, “what are tasks that people should be doing and want to be doing and are good at doing, and what are tasks that we can allocate to the AI?”

PAVIN: Don’t just upend everything for the sake of the bot, Gerber says. Weave it into your existing processes. Because AI and humans each have their own strengths and weaknesses, and you’ll want to find some harmony between the two.

But the other reason you’ll want to think hard about how AI fits into the existing equation is because it could cause some frictions that you’re going to want to work through.

Gerber gave a helpful example of this. Imagine you’ve got an AI tool that’s supposed to help hospitals get a handle on infections and how they spread. That tool needs data to work, like data about who at the hospital interacted with an infectious patient. But there’s a problem.

GERBER: Not everybody wants to be tracked. Doctors and nurses don’t want a tracker on them to say exactly where they are at all times. Sometimes many of these people already feel like they’re being intensely scrutinized already. And so then to be tracked to know when they’re going into what rooms and going out of what rooms can be can be problematic. Some are willing to give it, some don’t even realize they are producing usable data, and it can create new routines for some folks.

PAVIN: There’s already a tremendous amount of oversight, and charting, and documenting in the medical space. Adding one more layer to that isn’t a small ask. It’s a big one. Gerber says you’ll have to think about those kinds of sensitivities and find a way to get people’s buy-in because it could mean the difference between you getting the data you need and not getting it.

The point is, you need to remember that your AI tool will exist in a complex ecosystem, one that’s full of little practices or ways of working that you hadn’t accounted for. And you’re probably going to need to go back to the drawing board or tweak the way people do things to make it all work.

Which leads us to our final question we want you to ask yourself: What will you do when things go wrong?

Let’s return to Professor Rahman’s example about the hospital that was developing the AI ultrasound probe—the one that would improve maternal–fetal outcomes. To work well, the thing needed data from patient scans. But they kept running into issues. Like the quality of the scans, they weren’t always so great, which made it hard to train the AI on. There was also this behavioral thing happening with the nurses.

RAHMAN: Normal procedure, when you’re doing a scan for a patient, the way the nurses explained it was that, if you stopped to look at something interesting that the baby is doing, or a patient’s like, “Oh, I want to see the baby’s hand,” or something like that, usually that’s okay. But what they found is when using the ultrasound probe to collect the scan for an AI, to train an AI model, stopping disrupted the collection of the data.

PAVIN: To get better scans, they had to retrain nurses on how to use the probe. No more stopping to watch the baby do something cool, like wave or hiccup. And if you or a partner have ever received prenatal care before, stopping for the little stuff can be a big deal!

So if the hospital was expecting their solution to be implemented without a hitch, they were likely sorely mistaken. They had to go back to the drawing board and amend their plans, probably multiple times! It’s important for you to expect the same will happen to you. Because it will.

Gerber, for example, used to work for an organization that often deployed software.

GERBER: And when they deployed the software, they gave everybody a punching bag. Like, literally, a punching bag to put on their desk. And I thought, how interesting. They already are assuming that people are going to be frustrated with this. They know it’s not going to work. Uh, they’re recognizing it’s not going to work, and they want to have a closed feedback loop with those folks so that they can correct it as soon as possible.

PAVIN: Do you need a punching bag? Your decision. But just knowing that failure is a normal part of the process can save you—and everyone around you—a lot of frustration.

...

PAVIN: In all of this advice, the throughline is this: plan, plan, plan. Expect things to go wrong, and plan again. It’ll help you land in a place that’s more satisfying for you and your organization.

So again, we get it! AI is this thing that feels very different from any other technology before it. But by looking at the places that can be the hardest to get right in ANY product-development cycle, you will probably be pretty well-equipped to succeed with AI, too. So just remember that, before you come in hot with a spiffy new AI solution, ask yourself the questions: “What’s the point of this?” “How does this fit into existing processes?” And “how am I going to handle things when they go wrong?”

That’s our show for today!

[credits]

PAVIN: This episode of The Insightful Leader was produced and mixed by Isabel Robertson. It was produced and edited by Laura Pavin, Jessica Love, Fred Schmalz, Abraham Kim, Maja Kos, and Blake Goble. Special thanks to Hatim Rahman and Liz Gerber. Want more The Insightful Leader episodes? You can find us on iTunes, Spotify, or our website: insight.kellogg.northwestern.edu. See you next time.