Information for this eWEEK Data Points article was supplied by Dr. Charla Griffy-Brown, Professor of Information Systems and Technology Management, and Associate Dean of Executive and Part-Time Programs at Pepperdine University’s Graziadio School of Business. Here she discusses five key reasons AI strategies fail and what businesses can do to avoid these pitfalls.
Data Point No. 1: Technical performance
Early work on AI solutions usually involves small subsets of data, which require smaller computing resources. When AI expands into broader production systems, performance can be impacted exponentially. Insufficient attention to performance at scale creates AI systems that appear to work well during testing but quickly become unusable by the business at large.
Solution: Businesses should be accurate in computing requirements for scaling up and test, as often as possible, in a near-production environment.
Data Point No. 2: Veracity of data and volume of data
There are fundamental issues that arise from decisions regarding data architecture. The wrong database can easily render a scaled AI working test system unusable. Furthermore, this is enhanced by data cleansing and preparation problems. For example, manual interventions by humans might be effective in preparing test data, but this typically cannot be scaled.
Solution: Make data architecture decisions based on not just growth but an understanding of the processes required for the data training required to build AI.
Data Point No. 3: Business processes and people
One of the biggest challenges facing implementation of new technology is human beings, and AI implementation will only be as strong as the training and support for the staff implementing it. AI solutions must also be developed with a mechanism for ensuring customer facing channels are fully prepared for customer reactions. For example, this could include a temporary spike in phone calls if chatbots aren’t working properly or a tsunami of emails if a phone answering service isn’t getting them where they need to go.
Solution: Realizing that AI requires human work is fundamental to thinking through AI deployment. Businesses will need to implement strategies to address challenges quickly in advance of an AI initiative, including considerations for how it will impact human staff and customers.
Data Point No. 4: Unexpected behaviors
Supporting business issues that didn’t appear in testing is very challenging to scale. Scaling AI requires production systems to allow for situations not in designs or plans. Over time, new challenges may arise because of changes in the AI system itself. Machine learning is designed to improve itself over time, and usually this improves the accuracy of an algorithm. However, it can also lead to other revelations, such as identifying new patterns of customer behavior or fraud.
Solution: An important part of scaling AI means developing and working through a variety of hypothetical scenarios. Businesses should develop technical and operational contingencies, such as asking how to “switch off” an AI solution temporarily with minimal disruption.
Data Point No. 5: Data security and governance
One of the most important problems in scaling AI for production are the security implications. Cyber risk is an element that has to be considered from all angles when deploying AI. AI introduces new vulnerabilities and represents new risks to established cybersecurity solutions.
Solution: Before deploying AI, companies should develop a risk-based approach to implementation, identifying any points of weakness and reinforcing these appropriately. They may also consider working with a third party to test cybersecurity protections ahead of time to identify points of vulnerability.
If you have a suggestion for an eWEEK Data Points article, email [email protected].
social experiment by Livio Acerbo #greengroundit #thisisnotapost #thisisart