A top IBM Research executive outlines the 3 critical steps organizations pursuing enterprise-wide AI need to take

• Only 4 percent of executives plan to deploy enterprise-wide AI in 2020, down from an expected 20 percent in 2019, according to a new study from PricewaterhouseCoopers.
• But despite the difficulties in scaling projects from the pilot stage, it's essential if companies want to achieve significant organizational impact, according to Sriram Raghavan, vice president of IBM Research AI.
• Raghavan says companies should focus first on "core AI," or the projects that are already backed-up by robust data sets and strategic enough to serve as a proof of concept for other efforts.
• Focusing on trust next to ensure the model doesn't have any bias that could undermine results helps pave the way for a smoother expansion across the enterprise, according to Raghavan.
• Click here for more BI Prime stories.

Enthusiasm around artificial intelligence appears to be waning as companies grapple with how difficult it is to actually implement the technology across the enterprise.

Only 4 percent of executives plan to deploy AI in their operations in 2020, down from the expected 20 percent this year, according to a new survey of over 1,000 C-Suite and IT employees from professional services firm PricewaterhouseCoopers.

A key reason for that, according to PwC, is the challenge companies face in organizing their stored data to make it easily analyzed by AI models. It's also an acknowledgement of the limited return on investment that smaller AI projects provide, but the difficulty in expanding ongoing test cases across the company, according to Sriram Raghavan, vice president of IBM Research AI.

"People are recognizing that if they don't think about scaling and if they don't bring innovation to the scaling of AI, they are going to be stuck in pilot," he told Business Insider. "It's critical to scale AI because that's when the advancements are going to be made."

Those companies that do intend to pursue the tech face a litany of challenges. Cultural resistance can doom even the most earnest efforts, and the battle for tech talent means organizations could be stretched to find the amount of employees necessary to manage the initiatives. Bias algorithms can also undermine trust in the technology, making it more difficult to get buy-in from all parts of the enterprise.

That's where IBM comes in. The tech behemoth is advising clients to think about their efforts in three steps: advancing, trusting, and scaling.

The insight from Raghavan - who previously served as the director for the IBM Research Lab in India and its research center in Singapore - can help organizations successfully implement the advanced tech in their operations. As one of IBM Research's leaders in AI, Raghavan has spearheaded projects like AutoAI, a platform that automates key steps in the process to deploy the technology like data preparation and model development.

Understanding your 'core AI'

The first step for leadership is honing in on the most appropriate initial use cases, or what Raghavan refers to as "core AI."

Those are the areas that organizations already have robust data sets on and are strategic enough to serve as a proof of concept for other AI-based efforts. In the case of a fast food company like McDonald's, for example, that can be using ordering habits to better predict customer preferences.

"The core only tells you of a use case [that] the technology is ready to solve," said Raghavan. That's not to say no new technology will be required. Instead, it means pursuing projects that can be aided by existing applications offered by IBM and others.

Say an airline wants to better predict the customers that should get free seat upgrades. The company could quite easily pinpoint individuals based upon existing loyalty members and how often they fly. But it may also want to include data on how often someone calls into the customer service center. That information is likely to be less-structured - meaning it can not be easily analyzed by the company's existing technology.

Alongside tools to organize the data, the airline would also need to consider investing in platforms to monitor the corresponding AI model for potential bias. Those decisions should be made before organizations try to expand the initiatives beyond the pilot stage.

Demystifying the 'black box' technology

Trust in AI remains a hurdle to corporate adoption of the technology. That's why IBM is putting ample resources towards tackling the problem.

Among other things, it offers "AI Explainability 360," a platform that uses algorithms, demos, and other resources to provide insight into how models reach a final conclusion. But companies need to think about trust beyond just issues of racial or gender discrimination.

Other types of biases have the potential to produce inaccurate or unhelpful insights. A model, for example, might be much better at predicting the purchasing habits of East Coast customers versus those on the West Coast. Understanding those inconsistencies is paramount to fixing them, whether that be gathering better data from customers in states along the Pacific Ocean or building separate models for the different markets.

Another challenge for organizations is explaining the technology in a way that all employees can understand. AI can often be referred to as a "black box," or a platform that ingest data and produce recommendations without clear insight into how it comes to those decisions.

That's putting more pressure on companies to be able to explain their models. In some industries like financial services, it's mandated by law. In others, it's key to get employees to "believe the results of AI and use it for decision-making," according to Raghavan.

Models also have to be robust, meaning they must be able to operate in environments where the inputted data is continually changing. "Even if your data was not biased, your model may not be robust to changing conditions," Raghavan said.

The holy grail: autonomously scaling AI models across the enterprise

Once companies are able to successfully implement the smaller pilot projects, the next step is scaling AI across the enterprise.

While the jump signals the promise of reduced costs and more time for employees to tackle the less-mundane aspects of the job - like, in the case of Walmart's store associates, removing product from trucks - it's also a perilous step for leadership.

For one, companies need to pursue AI models that can be managed adequately by the existing staff, said Raghavan. Organizations should aim to "reduce the amount of time [their] precious data scientists have to create a model."

Scaling projects requires considerable manpower, including individuals that regularly monitor the models and infuse new data as it becomes available. But the hope is that, in the future, the AI application would be able to automatically evaluate and pull in data without the need for human interference.

So in the case of the airline wanting to better predict who should receive seat upgrades, an autonomous model would be able to automatically pull in data on new customer segments or adapt itself to acknowledge that travel is likely to increase around major holidays.

To assist customers in progressing towards that goal, IBM offers clients tools that monitor the models and alert companies to when it may be acting atypical. This helps to bring down the operational costs and reduce the amount of data scientists and other employees that must oversee the project.

While enterprise-wide AI is unlikely to get easier anytime soon, this checklist can serve as an initial foundation for companies starting on the journey.