In 2018, a study by Forrester revealed that although AI projects abounded in companies, few entered production and diffused in the company. Three years later, the results are mixed to say the least, even though CIOs have undergone multiple transformations to develop their working methods and cultures, around the cloud and agility. Far from being completed, these transformations have a simple objective: to meet the expectations of general management and business lines to deliver the value of digital transformation initiatives.
However, the question has still not been broadened to the subject of data – AI, because companies are still developing data-labs. If these structures allow companies to test AI concepts and serve as a bridgehead in the acculturation of employees, they are unfortunately often disconnected from the business realities of companies.
This situation makes it impossible to maximize the ROI of AI initiatives. The projects are then either abandoned or decorrelated from what was done during the design phase. Because, by forcing the switch to production, the company very often transforms the project from a technical point of view, directly impacting the relevance of the project and the ability to keep it up to date. This kind of situation leads data scientists to prefer not to put the project into production for fear of failure.
Fortunately, not all projects are doomed to deadlock. Some companies succeed, hands down, in their transition to production. This is particularly the case for companies such as GE Aviation, Schlumberger or Euronext, which have made data science a business transformation project and not a peripheral project. A strategy resulting from the commitment and support of company executives but also from a desire to facilitate organizational, cultural or technical changes. Data science is no longer isolated, but is becoming a common fight for IT and businesses.
Industrializing AI projects requires setting up a central hub
In terms of methodology, the creation of a central hub in the company makes it possible to manage governance, transversality and the provision of tools, for data science or other subjects. The users of these tools are the spokes, carriers of data projects close to the field. They use technical capabilities and implement best practices. The hub is not the maker, but the guarantor of the ease of production, a role that can perfectly be provided by IT.
Because, as we have seen in the past, IT has managed subjects as varied as business intelligence dealing with training, running or providing experts to support businesses. The whole difficulty is not to fall into excessive centralization, the great temptation of IT. The hub must help without replacing the spokes.
Converging on a platform makes it possible to avoid the phenomenon of decorrelation observed in the case of AI projects. Like the majority of applications, the data platform must be seen as a central bastion of stability, a point of convergence which makes it possible to control both quality, obsolescence, and documentation in an evolutionary way.
Without this governance giving a vision of who does what, when, with which version and why, the transformation is useless. And this is all the more true as regulations are tightening in terms of data processing and AI. It therefore becomes critical to be very relevant in terms of traceability. And if the hub minimizes the risks, it only makes sense if it is agile and flexible and integrates the businesses.
Giving yourself the means to become an Everyday AI company requires setting up a common language for all players so that the data teams can hand over to those supposed to industrialize the projects. This is both a challenge in terms of skills, with the emergence of run specialists, like MLops (for machine learning and IT Operations, editor’s note), and in technical terms with platforms that avoid changes in codes and conceptualization.
It sounds simple, but the implications are numerous: having clear documentation for everyone, a common understanding of what is done, how the data is processed… This centralization of knowledge will facilitate the adoption of new incoming than pre-existing projects, while ensuring the success of the only objective that matters: delivering value to the company.