Data quality, fast failures and quick wins key to AI success

The importance of AI

Enterprises need to experiment with AI development and soon have success deploying AI tools because of their transformative potential.

We've created incredible new capabilities that apply to the bulk of knowledge worker activities, but … we have work to do to really figure out how to put this massive technology to work.

Applications such as chatbots and agents can improve decision-making by drastically simplifying the previously complex process of exploring and analyzing data, both boosting the overall knowledge of workers as well as improving the quality of major strategic decisions. In addition, AI tools can improve operational efficiency by automating repetitive business processes and performing previously manual tasks, so employees can spend their time doing more meaningful work and  businesses can scale operations without having to add more staff.

AI tools can even do much of the work required to prepare the data that gives them their own intelligence, such as checking the quality of data used for AI initiatives and retrieving the relevant data for agents to autonomously perform specific work, such as detecting fraud, optimizing supply chains and personalizing outreach.

"AI can help you speed up data management," Petrie said. "AI can help with data quality, with metadata generation, documentation, preparation, integration and so forth. But you need people in control. … Probabilistic models can't get it right all the time, so the task is to figure out how to have humans deal with that ambiguity."

In addition, enterprises need to be putting AI tools that perform as intended into production because  competitors are likely doing so, and those that have success putting AI tools into production can derive significant competitive advantages over those that don't properly value AI's potential.

"Your competitors are moving, so you need to start experimenting," von Simson said. "It's not so much about getting ahead, but it's about thriving, whatever that means for your organization -- you are building a concrete sense of value generation for your activities, and that will pay for your budget for AI and your continued experimentation."

AI, when organizations have success building and deploying applications, is so potentially transformative that its effect is compared to that of electricity, the telephone and the internet.

No business could compete without such technologies. And in a short amount of time, no business will be able to compete without AI being a ubiquitous part of their operations.

"[CIOs are saying] that this is moving so fast and they need to be on top of it and understand it even better than they do," Suer said. "They are convinced that this is probably one of the biggest changes that they've ever been involved with."

Best practices

While AI is viewed as a transformative technology, its benefits remain merely potential for most organizations rather than reality.

Despite the ongoing lull as organizations struggle to develop agents and other AI tools that deliver relevant, accurate outputs at an acceptable level, the panelists noted that there are logical steps that can improve the likelihood of success.

While myriad problems contribute to the high failure rate of AI initiatives, problems with the underlying data used to inform AI applications are prominent among them. Disorganized data, data that's isolated in disparate systems and can't be integrated, and poor data quality are common barriers.

Therefore, making sure data is properly prepared for AI is critical for enterprises to increase the success rate of their development initiatives, according to Petrie, who noted that Jonas Prising, the CEO of ManpowerGroup, recently emphasized the importance of a modern data infrastructure when discussing AI development with The Wall Street Journal.

"He talked about the need for a strong data foundation and … helping what is the number one obstacle to AI success, and that's data quality," Petrie said.

In addition, BARC research found that organizations with executive-level involvement in ensuring data quality have more success building AI tools than those that don't, he continued.

"It definitely makes sense for executives to show a level of proficiency with data and data quality to make sure that you have clean inputs," Petrie said. "Those folks make stronger inroads faster and … get more into production faster. They started with a strong base for the whole company."

Another path toward success with AI is to keep experimenting -- even before data is completely AI ready -- and accept that failure is part of the development process, according to von Simson.

"It takes time and money to [prepare data], and you're not going to have time to complete that exercise before you're out of time," he said. "So, you must experiment. You must try, fail fast, and learn faster. You have to start with where you're at now and what you can try this week, next week and so forth."

In addition, alignment between an AI initiative and the enterprise's overall business goals is beneficial, von Simson continued.

"Be strategic," he said. "Do things that are core to the value of your business. That's your guiding light. You try things out and hope they work, and if they don't, you move on. But like startups, you have to try and fail before you succeed. It's the human condition."

Modest initial goals are also important when trying to move AI projects into production, according to Willis.

When an organization attempts to overhaul major business processes as an initial foray into AI development, initiatives are destined to fail. But when an organization tries to build an AI tool with a very specific intent, the likelihood of success increases. That, in turn, builds momentum for additional projects.

"You have to act responsibly and with purpose when trying to do these kinds of things," Willis said. "Those quick wins, the ones that are less risky than the bigger moonshots, are addictive."

Finally, having a CIO with a foundation in AI and an understanding of models and what they can do is beneficial, according to Suer. In addition, the CIO needs to comprehend the business so that initiatives are designed to address business problems, and the CEO needs to be brought into the planning process rather than be an outside observer.

"It is what it has been -- people, process and technology," Suer said.

Concern as a constraint

Although poor data preparation for AI is perhaps the biggest problem preventing organizations from building production-ready AI tools, there are others as well.

In particular, risk -- whether it's exposing proprietary data, misusing sensitive data, violating regulations, or AI hallucinations that lead to mistakes or humiliation -- remains a concern for many CIOs, according to Suer.

"They are worried about risk and how to contain that risk," he said. "I think that keeps them awake at night."

Deploying AI tools to inform decisions and execute certain operations represents turning a company over to technology, Suer continued. He noted that organizations were forced to rely on technology more than ever before during the COVID-19 pandemic. Empowering agents and other AI tools to make decisions and operate autonomously represents a new level.

"That's what they're worried about, missing something that could completely disrupt a business and make it irrelevant," Suer said.

Systems integration -- or lack thereof -- and lack of people skills to ask the right questions of AI tools and evaluate AI-generated insights are other concerns preventing success with AI, according to Petrie.

"People are probably the hardest element to manage, and making sure that you're giving people the appropriate bridge to help them succeed is critical," he said.

Whether concerns are legitimate or not, enterprises need to get past them and, if they haven't already, begin experimenting with AI so they can learn what works for their organization and what doesn't, according to von Simson.

"You can't take … long," he said. "You have to try something that's really short, just as a test, and if the test works, then keep going. If it doesn't work, you have to re-think. But failing fast is really important because you only learn through failure. Success is the reward at the end."

Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.