Slow pace for AI implementation is a prudent business strategy

AI remains experimental for most

Most business AI projects are somewhat experimental right now. Organizations are trying to give new capabilities to workers in departments like customer service, supply chain and public safety, but few are implementing broad AI projects that span their operations.

Andrews said in a presentation at the Gartner Data & Analytics Summit 2018 in Grapevine, Texas, that this is because an enterprise's first AI projects won't likely generate much business value. There's a learning curve, and getting some AI experience under the organization's belt is important -- but few early initiatives will generate new revenue.

For this reason, Andrews recommended that businesses focus on other measures of project success for their first few AI implementation projects. This may include things like the number of customer inquiries handled by chatbots.

Executives will always want to know what kind of return they can expect from a technology investment, but Andrews said AI proponents within an enterprise should focus on interim goals that the executives are likely to agree merit attention.

"Aim for soft outcomes," he said. "Your first project will probably not generate financial ROI. You should always focus on worker augmentation as your story."

Additionally, projects should be focused on scale. Andrews said a lot of the AI implementation projects that get attention in the media are moonshots -- things like self-driving cars, image recognition tools and medical diagnostic systems. But these types of large-scale initiatives would be a poor fit for most enterprises today.

"It's important to address reasonable and possible goals," he said. "People aren't impressed by your lessons learned."

Keep AI expectations grounded

Along those lines, Gartner analyst Erick Brethenoux said AI project leaders need to manage expectations within their organization of what the technology can accomplish.

He pointed to maturing realms of AI, like natural language processing, smart assistants and graph knowledge bases, and said people get excited when they see these technologies. But the business cases for some of these tools are still emerging, so expectations about their ultimate impact need to be tempered.

We tend to anthropomorphize AI, and that's a huge issue because it raises expectations.

"AI really delivers super powers. However, one of the main problems is the expectations we have out of them," Brethenoux said. "We tend to anthropomorphize AI, and that's a huge issue because it raises expectations."

One important caveat to keep in mind to help mitigate unreasonable expectations for an AI implementation is that it is not intelligence in the truest sense. Brethenoux said AI applications can simulate certain functions of the human brain, but they do not think for themselves. Ultimately, because their functionality is underpinned by statistical methods, they essentially return probabilities, not answers.

Another common pitfall that should keep AI expectations under control is the difficulty of operationalizing applications. Brethenoux pointed to Gartner research showing that only about 13% of data science projects get implemented. Simply developing an AI application that performs a function doesn't guarantee it's going to move the needle on any business problem.

Keeping an AI implementation targeted, working with the business to develop it and having a plan for putting the application into production are keys to maximizing value. "The problem is, until you put a model into production, all of it is academics," Brethenoux said.