The U.S. Army admits it lacks personnel trained for the AI it's already buying from Palantir and Anduril

By Zendoric · September 15, 2026.

Mario Roberts, chief AI officer (CAIO) of U.S. Army Intelligence (G-2), has acknowledged in an interview with FedScoop that troops still do not know how to use AI effectively, in the middle of the Army's accelerated adoption phase. Roberts describes the current moment as a "turbulent" and "a little chaotic" transition. He will expand on it in a panel at the Intel Summit 2026, on September 24.

The problem, according to Roberts, is not one of attitude but of a very specific craft: managing data centers, fine-tuning (retraining a model for a specific task) and training models, and deploying and scaling them across the organization. That profile, he says, simply does not yet exist within the Army's workforce. And he warns that reskilling people with quick courses is no substitute for experience: "it takes time to cultivate and develop a workforce to the point of really understanding how AI operates and works today," he said. IT, cybersecurity and networking skills do not disappear, he clarifies, but they are overshadowed by the demands of AI.

Roberts himself admitted that the Army recently deployed AI with an approach that, in hindsight, proved unsuited to that particular model. His conclusion is not to slow down, but the opposite: "without moving fast, without employing AI models, without scaling it to the units, we're not going to learn or grow," he argued. It is an explicit bet on learning as you go, accepting the cost of mistakes.

Roberts also draws a technical boundary worth understanding: deploying AI in the administrative rear (finance, management, policy drafting) is not the same as taking it to the tactical edge, where it operates alongside the soldier on the ground. In the corporate cloud, he says, scaling a model is fast and the return on investment arrives much sooner. At the edge —where AI has to learn to interpret images, translate and other critical capabilities— what works in the cloud cannot be deployed as is.

That tactical edge is precisely where the Army is already investing seriously. The TITAN (Tactical Intelligence Targeting Access Node) program, a ground station for intelligence, surveillance and reconnaissance (ISR) that uses AI and machine learning to process sensor data and support targeting, recently entered production. The Army awarded Palantir Technologies and Anduril Industries orders worth $192 million for eight initial systems (TITAN Advanced and Basic variants), to be delivered over the next 18 months, according to data compiled by ExecutiveGov.

This is not the first time Roberts has dealt with this kind of gap. As the Army's cybersecurity lead for the G-2, he already managed the shortage of trained personnel during the migration to the cloud: in a 2018 interview with Defense One he acknowledged that the Army did not even have a military occupational specialty (MOS) for cloud computing, and had to send its IT staff to train and get certified on platforms such as Microsoft Azure.

Our take: the case illustrates, with more honesty than usual, something almost nobody admits in public: the bottleneck for AI is not the model, it is the organization around it. The industry sells the fantasy of taking it "out of the box" and simply deploying it; Roberts himself confirms that this only works for office tasks, and that in the field —where mistakes cost lives— the technology demands a craft that no crash certification can replace.

This fits with what we have been observing sector by sector: AI does not eliminate technical skills, it shifts them toward new and scarcer profiles, and those who know how to operate and govern AI infrastructure gain weight over those who only applied traditional IT. In the Army's case, the talent gap coexists with a shift in spending toward AI-native defense contractors such as Palantir and Anduril, which are gaining ground on the big traditional contractors in programs like TITAN. The institution is buying capability faster than it can train the people who will operate and oversee it, and that asymmetry —buying before knowing how to use— is the real short-term risk, not a hypothetical military AI out of control.

In the long run, the promise of AI applied to military intelligence —compressing decision cycles, reducing human error under pressure, freeing analysts from the most repetitive tasks— remains consistent with the underlying thesis: more capability, better governed, so that people can devote themselves to what truly requires judgment. But that horizon is only reached if institutions like the Army invest in training their people at the same pace as they invest in buying the technology. For now, according to its own AI chief, they are behind.