Cause of AI Success and Failure
CEO and CDO secrets to successfully use AI in their company
Through different interviews that I have made while discussing with Chief Executive Officers and Chief Data Officers, it is possible to identify several factors that have allowed the success or failure of their ambitions. As a result, some companies consider these criteria as necessary to start development because of their impact on the viability of their projects.
1. Explanation of the Challenges of AI to the Business
The first essential factor is understanding the business. This is achieved by using active listening and short iterations to validate the need throughout the project. Moreover, the success of projects also depends on explaining the issues related to AI to the various people who do not have knowledge in this field. This allows better communication and facilitates the management and the development of the project.
Moreover, these explanations will allow demystifying for many the functioning and the possibilities linked to this technology. Indeed, the term intelligence is confusing, because it is not totally founded.
2. Reflection on the Quality, Access and Value of Data
The data has a great impact on the outcome of the project. The failure of the latter may be due to lack of data, poor quality data, and non-existent added value. In addition, it is difficult to really define how much data is needed and what value it contains. These ideas are quite abstract and research papers are totally dedicated as these topics are so complex. Nevertheless, an adapted infrastructure is necessary to be able to manage and fully use this data.
The majority of companies surveyed advise using Business Intelligence before Artificial Intelligence. Big Data can already bring a lot of value to the company and allows for future investment in the integration of tools using Artificial Intelligence. In addition, the use of BI allows for a better understanding of the data, which is essential to best perform ML algorithms. Finally, all these approaches depend on the technological maturity of the company.
3. Clear Definition of Sponsorship
Clear sponsorship is the third decisive factor for project success. For the majority of respondents, this element is so important that it is one of the key elements to validate their implementation. Without the support of an individual controlling the financing of their activity, it is sometimes difficult to prioritize this development which sometimes has no direct financial impact.
There is no specific financing method, it depends on the company. Whether it is an expense charged to the IT department, the business, or a mix of both, it must be defined at the beginning to ensure that the development reaches its conclusion and that the project remains a priority.
4. Definition of KPIs from the Beginning
For many, defining business KPIs from the beginning is a necessity. These indicators provide the goal and condition for the success of the project. As with sponsorship, for some, the definition of this metric is imperative to validate the start of the development.
Without knowing the success of the algorithm, it is difficult to define when the project has met expectations and to determine when to stop the iterations. Without this indicator, the project can be endless, as it has no quantifiable goal.
5. Explanation of How the Algorithm Works
To promote the use of applications using AI, it is necessary to explain to the various people using it how it works. For some, it is difficult to trust an algorithm that predicts results that are not understood.
To ensure the success of its integration, it is necessary to communicate how it works and justify its choices. The only problem with this approach comes from the comprehensibility of AI. Indeed, in some sectors of activity, it is necessary to have algorithms that are explicable and interpretable.
For example, this is the case in banking for fraud and money laundering analysis. As explained by several interviewees, it would be extremely unpleasant for a customer to have all his financial transactions blocked due to the use of an uncontrolled AI algorithm.
Today, succeeding in creating interpretable AI algorithms is a real challenge for companies. Many researchers are working to solve this problem, as it limits the use of technologies such as neural networks due to their level of abstraction.
6. Precise Problem Solving
Many indicate that focusing on solving a very specific problem is a good approach to maximize the chances of success of projects. Working on a very broad topic can lead to a very imprecise algorithm that is ultimately not good at anything.
On the contrary, focusing on solving a specific problem is a better approach, because it will avoid the integration of special cases from the beginning. Afterward, it will always be possible to extend the scope of the algorithm if it can be used in other domains.
7. Encourage Constructive User Feedback
Constructive discussion between end-users and developers should be encouraged. Some of the companies interviewed had implemented user feedback systems, but the content exchanged did not add any value.
The goal of these exchanges is to understand how the AI really performs to know how to modify the algorithm to improve its usability and results. Among other things, unconstructive deadlines can also cause the project to fail. Without this communication, it is impossible to perform iterations to make the algorithm fit the user’s needs.
8. Focus on Meeting the Need
Some developers may be tempted to be very focused on using complex technology. As a developer, it is very understandable that they want to work with challenges and progress by learning new models.
However, the desire of the data scientist is sometimes not aligned with the goals of the company. To help the project succeed, it is necessary to use the approach that will best meet the business needs by bringing the most value while being aligned with the pre-project thinking. In fact, a lot of value can already be provided with simple Machine Learning models.
IV. Conclusion
In the scientific literature, we can easily find the different problems that companies have to use AI. However, it is more difficult to find the approaches that facilitate their realization. On the one hand, we find in common all the data-related problems presented by L’Heureux et al. (2017) and Zhou et al. (2017).
On the other hand, the lack of funding and understanding of AI is expressed by Ransbotham et al. (2017). The same is true for the lack of confidence related to the applicability of models presented by Manlhiot (2018). In addition, the problem of bias induced by indirect communication is also raised by Megler (2019).
Nonetheless, it is clear that communication and team training are necessary for project success, as it alleviates the problems identified by Ransbotham et al. (2017).
L’Heureux, A., Grolinger, K., Elyamany, H. F., & Capretz, M. A. M. (2017). Machine Learning With Big Data : Challenges and Approaches. IEEE Access, 5, 7776‑7797. https://doi.org/10.1109/ACCESS.2017.2696365
Manlhiot, C. (2018). Machine learning for predictive analytics in medicine : Real opportunity or overblown hype? European Heart Journal — Cardiovascular Imaging, 19(7), 727‑728. https://doi.org/10.1093/ehjci/jey041
Megler, V. M. (2019). Managing Machine Learning Projects Balance Potential with the Need for Guardrails. Amazon White Paper, February.
Ransbotham, S., Kiron, D., Gerbert, Ph., & Reeves, M. (2017). Reshaping Business With Artificial Intelligence : Closing the Gap Between Ambition and Action. MIT Sloan Mangement Review and The Boston Consulting Group, 59(1), 1‑17.
Zhou, L., Pan, S., Wang, J., & Vasilakos, A. V. (2017). Machine learning on big data : Opportunities and challenges. Neurocomputing, 237, 350‑361. https://doi.org/10.1016/j.neucom.2017.01.026
