At the recent AI World Government event held both in-person and virtually in Alexandria, Virginia, Isaac Faber, Chief Data Scientist at the US Army AI Integration Center, discussed the US Army’s approach to building its AI development platform. He emphasized that the AI stack framework defined by Carnegie Mellon University is central to the Army’s efforts in this area.
Faber highlighted a major challenge in moving the Army from legacy systems to digital modernization: the difficulty of abstracting differences between various applications. He explained that the most critical element of digital transformation is the middle layer—the platform that simplifies running software either on the cloud or on local computers. The goal is to enable software platforms to be transferred from one environment to another as easily as transferring contacts and histories to a new smartphone.
Ethics, Faber noted, is a concern that spans all layers of the AI application stack. This stack is structured with the planning stage at the top, followed by decision support, modeling, machine learning, massive data management, and finally the device layer or platform at the bottom.
He urged that the AI stack be viewed as core infrastructure that supports application deployment rather than as isolated silos. Faber stressed the need to create a development environment that supports a globally distributed workforce, allowing teams to collaborate effectively regardless of location.
The Army has been developing a Common Operating Environment Software (COES) platform since 2017. This platform is designed for Department of Defense work and is scalable, agile, modular, portable, and open. Faber described it as suitable for a wide range of AI projects but cautioned that the success of such efforts depends heavily on attention to detail.
To build this platform, the Army is collaborating with Carnegie Mellon University and private companies, including Visimo from Coraopolis, Pennsylvania, which provides AI development services. Faber expressed a preference for working closely with private industry rather than purchasing off-the-shelf products. He explained that relying on a single vendor’s product can limit value, especially since such products are often not designed to meet the unique challenges of Department of Defense networks.
The Army also focuses on AI workforce development across several groups. These include leadership personnel with graduate degrees, technical staff who undergo training and certification, and AI users. The technical teams have diverse roles, such as general-purpose software development, operational data science, deployment and analytics, and machine learning operations. For example, a large team might be needed to develop a computer vision system. Faber emphasized that as personnel progress through the workforce, they require a collaborative space to build, share, and work together.
Projects undertaken by these teams vary. Some are diagnostic, combining historical data streams. Others are predictive or prescriptive, recommending actions based on predictions. Faber pointed out that true AI projects come later in the process and should not be the starting point. Developers must address three interconnected challenges: data engineering, the AI development platform (which he called the “green bubble”), and the deployment platform (the “red bubble”). These areas are distinct but closely linked, and successful project teams usually include experts from each area. Faber advised against attempting to solve the AI development platform problem before there is a clear operational need.
When asked which group is the hardest to reach and train, Faber identified executives as the most difficult. He explained that executives need to understand the value that the AI ecosystem can provide, and communicating this value effectively is a major challenge.
During a panel discussion on emerging AI foundations, Curt Savoie, program director for Global Smart Cities Strategies at IDC, asked panelists about AI use cases with the most potential. Jean-Charles Lede, autonomy technology advisor for the US Air Force Office of Scientific Research, pointed to decision advantages at the edge, supporting pilots and operators, as well as decisions related to mission and resource planning.
Krista Kinnard, Chief of Emerging Technology for the Department of Labor, highlighted natural language processing as a key opportunity to expand AI use within her department. She noted that the Department of Labor deals with data about people, programs, and organizations.
The panel also discussed risks and challenges in AI implementation. Anil Chaudhry, Director of Federal AI Implementations at the General Services Administration, explained that unlike traditional software development, AI decisions can affect large groups of people and stakeholders. A simple algorithm change could delay benefits for millions or cause widespread incorrect conclusions. To mitigate this, he requires contractors to include “humans in the loop and humans on the loop” to maintain oversight.
Kinnard agreed, emphasizing that humans will remain involved to empower better decision-making. She stressed the importance of monitoring AI models after deployment because models can drift as underlying data changes. Critical thinking is necessary not only to perform tasks but also to evaluate whether the AI’s outputs remain acceptable. She added that the Department of Labor has developed use cases and partnerships across government to ensure responsible AI implementation and affirmed that algorithms will never replace people.
Lede discussed the challenge of working with limited real-world data, noting that simulations are often used to train algorithms. He warned of the “simulation to real gap,” where algorithms trained in simulated environments may not perform as expected in real-world situations.
Chaudhry underscored the need for a robust testing strategy for AI systems. He cautioned against developers becoming enamored with tools and losing sight of the project’s purpose. He recommended that development managers incorporate independent verification and validation into their plans. Leaders should have a clear idea of how to justify their investments before committing resources.
Finally, Lede spoke about the importance of explainability in AI. As a technologist, he values the ability of AI systems to provide explanations in a way that humans can understand and interact with. He described AI as a partner in dialogue rather than a black box producing unverifiable conclusions.
For more information, attendees and interested parties can learn more at AI World Government.