Agentic AI is moving rapidly from research labs to operational use, reshaping how militaries plan, decide, and fight. As open-source language models proliferate and battlefield conditions evolve, defense leaders are weighing how to field these systems at speed while maintaining human judgment where risks are highest.
Garrett Berntsen, Chief AI Officer at Accenture Federal Services and former Deputy Chief Data and Artificial Intelligence Officer for Scaled Capabilities, outlined the stakes and the path forward for an agile, software-driven force prepared for 2035.
Agentic AI urgency and open-source momentum
According to Berntsen, two forces are accelerating adoption: the spread of powerful open-source models and the operational demands of modern conflict. Advanced capabilities that once sat behind classified walls now emerge publicly within months, giving both allies and adversaries tools for rapid cyber operations and automated decision-making.
He said adversaries are not only processing information faster, they are acting faster, which places a premium on hardening systems with the same tools. He pointed to the war in Ukraine as evidence of the shift toward low-cost systems trending to full autonomy. Human-in-the-loop chains cannot keep pace with swarming attacks or fast-moving electronic warfare, he said.
To preserve speed advantages, forces must prepare to delegate selected workflows to machines under defined controls.
Establishing rules for agentic AI
Berntsen argued for a dynamic sliding scale of autonomy comparable to military rules of engagement. During his Army service in Afghanistan, he said ROE varied with the threat environment.
He believes AI autonomy should follow the same principle. In restrictive contexts, autonomous functions require close human oversight. In a near-peer conflict where adversaries employ full autonomy, ROE must adapt to protect American and allied lives.
There is no one-size-fits-all policy, he added, and human command must align with mission risk.
Command authority without bottlenecks
Control should not be treated as a binary switch, Berntsen said. In low-risk, operational support functions, high autonomy with after-action human auditing can be appropriate, such as in logistics, contract analysis, or route planning.
In high-risk tactical settings involving kinetic effects or critical targeting, systems should operate with human-on-the-loop or strict human-in-the-loop oversight, where agents present verified options and commanders retain authorization.
He emphasized that the aim of agentic AI is to reduce cognitive overload so human decisions are grounded in synthesized, real-time information rather than unfiltered data streams.
Redesigning processes for speed
When machine speed outstrips human capacity, Berntsen said the answer is to re-engineer workflows rather than force people to work faster. Organizations should avoid using AI to marginally optimize legacy steps and instead determine whether those steps are still necessary.
If agents can instantly merge intelligence or generate reports, humans can focus on negotiation, strategy, and partnership-building.
Avoiding digital replicas of outdated processes
Berntsen warned that simply automating yesterday’s processes misses the point. If a 12-step paper approval is replaced with 12 AI agents, that is digitization, not modernization.
True transformation starts by designing from scratch around machine capabilities, often compressing many steps into one automated flow with a single human checkpoint.
From pilots to scale: government-industry alignment
To move beyond prototypes, Berntsen said agencies should push industry to deliver secure, edge-ready solutions that work across classified networks. He urged a shift from rigid, multi-year feature lists toward outcome-based, rapid-iteration contracting.
He called for an ecosystem that combines commercial frontier models with defense-grade security and domain knowledge, embedding engineers with operators so feedback loops are continuous and solutions evolve in the field.
Agentic AI and the 2035 force
Looking ahead, Berntsen predicted three major shifts. First, a sharper boundary will form between what is automated and what remains uniquely human. High-stakes tactical decisions will stay with people, while operational support such as logistics, administrative tasks, and routine intelligence sorting will be heavily automated by software agents.
Second, he said defense software economics will change as agentic coding reduces build costs, creating a world of software abundance. Capabilities will be created for specific mission needs and retired when they no longer add value, rather than maintained indefinitely.
Third, he expects software expertise to move to the edge. If software is central to mission success, those skills should be embedded with units.
He described a future where a platoon includes a software noncommissioned officer or warrant officer who can build a digital tool on the spot for a unique tactical problem, execute the mission, and then dismantle the tool when it is no longer needed. That distributed, agile capability, he said, is what a modern force looks like.