Each month, through our partnership with Harvard Business Review, we refresh our resource library with five new HBR articles we believe CIOs and IT leaders will value highly.
8 reasons AI projects fail
Why do artificial intelligence projects fail? Let’s examine some recurring issues – and expert advice on how to avoid them and increase your AI project’s chances for success
5. The black box effect
In some cases – for example, when an organization must trust AI-based systems to make a critical decision or prediction – understanding how the machine works is paramount. The U.S. Defense Advanced Research Project Agency (DARPA) is pursuing efforts to produce explainable AI solutions. For now, however, “cases where decisions need to be ‘explainable’ may also not be great for AI as many of the neural-network-based systems act as black boxes,” Vijayan says.
6. Misalignment with business priorities
“AI projects deployed to address problems that are not aligned with the business imperatives and objectives, or are addressing underwhelming business questions, mute AI’s impact and affect its adoption,” says Pace Harmon's Baritugo. Instead, IT leaders should identify meaningful business problems that have a significant upside and are backed by substantial data.
7. Architectural complexity
AI-enabled applications and networks rely on different processing architectures. That’s likely to change soon, according to ABI Research’s 54 Technology Trends to Watch. The next generation and AI and ML frameworks will be multimodal by their nature and may require heterogeneous computing resources for their operations, ABI Research analysts predict, noting the leading chipmakers will move away from proprietary software stacks and begin to adopt open Software Development Kits (SDKs) and API approaches to their tools. For now, however, it can be a stumbling block.
8. A dearth of AI talent
At the very top of LinkedIn’s top 15 emerging jobs in the U.S. for 2020 is AI specialist. Hiring for artificial intelligence pros of various titles (including AI and ML engineers) has grown 74 percent annually over the last four years, according to LinkedIn.
According to a recent Gartner survey, 56 percent of respondents cited a shortage of skills as their number-one obstacle to the deployment of AI.
“Firms need solid, multi-disciplinary talent to drive AI initiatives, especially given the expertise required to extract, clean, model, and analyze the data – not to mention evangelize AI and drive user adoption,” Pace Harmon's Baritugo says. “AI talent is still scarce, and developing AI talent organically may not be feasible.”
The move by vendors toward automating more machine-learning model creation will help alleviate this issue. In the meantime, reskilling application developers in machine-learning skills is a potential solution, as is outsourcing early AI forays.
Remember: Not all failure is bad
While unmet AI expectations are surely never the goal, it’s important to remember that not all failures are bad. “Finding how not to do something might actually be a success,” says Wayne Butterfield, director of cognitive automation and innovation at ISG. “This is really relevant in the world of AI and data, so we need to be careful in broad-brushing failures.”
[ What’s next in AI and NLP? Read 10 AI trends to watch in 2020. ]