An artificial intelligence (AI) agent developed while working alongside soldiers with the Army’s 101st Airborne Division reduced a cyber workflow from 100 hours to 15, according to the Pentagon’s GenAI Task Force lead.
Capt. Ryan Hetrick said his team built the agent while working in a tent at Fort Campbell, Ky., then enlisted the cyber officers involved to help spread its use. The approach pairs people who understand the mission with people who understand the technology, and it relies on trained champions to bring the resulting capabilities to other units.
Hetrick said his task force embeds with units to build tools using platforms including GenAI.mil, Maven Smart System, and War Data Platform. He urged defense industrial base companies to work directly with military users to achieve and expand AI-enabled workflow gains.
“You have to go where the warfighter is and understand their use cases and their mission,” Hetrick said Nov. 6 at GovCIO Media & Research’s AI Summit in Tysons, Va.
His team also built a Spot Report Agent for the White House’s FIFA World Cup task force, working with personnel from the FBI, Coast Guard, and military. Hetrick said the tool reduced a reporting process from about 20 minutes to 90 seconds, giving officials more time to coordinate responses to events on the ground.
Army chatbot helps recruiters resolve policy questions
For Army recruiters, AI is helping answer a different operational question: how to move an applicant through a complicated set of requirements without spending time searching manuals or seeking help from more experienced colleagues.
To do that, the Army developed Staff Sgt. Star, a chatbot that draws on a curated knowledge base of approved Army business processes and system guidance, said Mark Rodriguez, an infrastructure IT specialist with the Army.
The tool, which was developed in less than three months, addresses gaps that training alone cannot cover. Recruiters handle applicants with varied backgrounds and can ask the chatbot which waiver an applicant needs or how to process it, he said.
Chatbot assistance reduces demands on senior recruiters and station commanders, he said. And by helping users distinguish a process question from a system malfunction, the chatbot can reduce unnecessary help desk tickets, creating time for technical teams to enhance systems.
AI can produce insights more sensitive than its source data
The ability to combine information quickly can create security risks, particularly when AI draws conclusions that do not exist in any individual source document, warned Garrett Lee, regional vice president of public sector sales for Broadcom’s Enterprise Security Group.
Lee offered a hypothetical military logistics example. An organization gives an AI system access to maintenance schedules, supply chain information, equipment readiness data, and deployment information. Each dataset is appropriately controlled and protected.
Asked which units are most likely to face readiness challenges over the next 60 days, however, the system could combine those sources into an assessment of operational vulnerabilities, Lee said.
“No single document contained that assessment. The AI created it through inference,” he said.
That possibility requires organizations to consider the sensitivity of what AI can derive, as well as the information it receives, Lee said.
“In the AI era, we must also protect the information we produce, including insights we don’t even know our data could reveal,” he said.
Brandon Dixon, co-founder and chief technology officer of endpoint security company Ent, said organizations also need to understand how people and agents use information within specific workflows. Identifying sensitive data and approving applications provides only part of that visibility, he said.
Dixon described an incident at a hospitality company in which human resources employees uploaded sensitive records through WhatsApp after the application added Meta AI. Employees were using an approved communications tool, he said, but exposed information through an AI capability the company had not approved.
Security controls need to operate close enough to the work to interrupt risky actions as they happen, he observed.