The US government’s difficulty in adopting artificial intelligence is not primarily a shortage of powerful models. The harder problem is reorganising the institutions around them. That is the argument made by Garrett Berntsen, former Deputy Chief Digital and Artificial Intelligence Officer at the State Department and now Chief AI Officer at Accenture Federal Services, in a recent ChinaTalk conversation with host Jordan Schneider.
Berntsen uses the U-2 reconnaissance aircraft as a historical comparison. The aircraft, developed through a classified government programme, required advances in high-altitude flight, sensors, cameras and pilots’ equipment. But its importance during the Cuban Missile Crisis came from more than the aircraft itself. The US also created the National Photographic Interpretation Center, or NPIC, to analyse the imagery and connect raw intelligence to senior decision-makers. “The way they bought it mattered,” Berntsen said. “The innovations around acquisition reform mattered. The way industry and government worked together mattered.”
At NPIC, analysts turned photographic intelligence into usable assessments by examining enlarged images alongside manuals and intelligence reports on Soviet weapons. They used human judgement to determine whether vehicles, missiles and other objects matched existing intelligence. The process had to operate quickly because decisions during the Cuban Missile Crisis were being made under intense time pressure.
Berntsen argues that AI presents the opposite institutional problem. The technology is advancing largely through the commercial sector and is moving faster than government agencies can redesign their systems and decision-making processes. “It has moved so far ahead of our ability to integrate it into actual systems, workflows, and decision-making processes that we’re now flat-footed in the face of this technology,” he said. “That’s why APIs alone are not enough.”
There has been progress in operational applications, particularly Combined Joint All-Domain Command and Control, or CJADC2, where AI is being used to provide battlefield situational awareness, identify objects and intelligence, and deliver information rapidly to decision-makers. Berntsen sees slower progress in core business systems such as logistics, personnel and finance. Those systems, he argues, are themselves part of military capability because operational AI depends on accurate and timely data.
A basic administrative rule can therefore become an obstacle to AI adoption. Berntsen cited systems in which supply personnel are required to update information only once a week. Changing that requirement to daily updates would affect thousands of personnel, while many of the systems remain difficult to use. “You have to invest in getting the data cleaned, organised, and ready for use at an operational tempo,” he said.
The institutional challenge extends to incentives. Unlike private companies, government agencies do not face a constant competitive test that forces rapid adaptation. Berntsen argues that government must create its own “forcing functions” through leadership, personnel policy, rewards and consequences. Innovators should be rewarded for taking risks and pursuing unconventional career paths, while those who fail to make progress should face career consequences.
He also argues that programmes and budgets cannot all be protected indefinitely. During the ChinaTalk discussion, Schneider recalled an anecdote involving technology reform advocate Jen Pahlka and a general who reportedly argued that his budget should be cut by 50 per cent because such a reduction would force change. Berntsen said government programmes need “winners and losers” and that changing what government buys can create opportunities for new approaches.
Berntsen sees an advantage in the government being a second mover in AI. Unlike the U-2, which originated in a government programme, today’s most important AI developments are emerging commercially. Government can therefore observe what works before adopting it, learn from early mistakes and use its purchasing power to negotiate better rates. That advantage, however, depends on the ability to absorb the technology once it is available.
He also wants more advanced tools placed directly in the hands of service members. Drawing on his experience as a logistics officer, Berntsen recalled relying on Excel spreadsheets when existing software could not be trusted. He argues that troops should be given modern hardware and software, including computing resources at the operational edge, while policies that prevent experimentation should be reconsidered. “As much as humanly possible, get whatever the most modern tools are out to the edge,” he said.
The same debate applies to intelligence analysis and diplomacy. Berntsen believes AI could improve intelligence analysis because the process is structured and produces large quantities of historical data that can be evaluated for quality. Modelling high-level political decisions is harder because crucial information, including private communications and deliberations, is unavailable.
Whether AI could eventually conduct negotiations without humans remains unresolved. Schneider suggested that within five years some business negotiations could involve two models reaching an agreement automatically. Berntsen accepted that highly transactional exchanges may move in that direction but argued that major diplomatic and political negotiations would be different. “People want to talk to one another and negotiate,” he said.
Cybersecurity presents a more immediate problem. Berntsen said AI is accelerating the discovery of existing vulnerabilities rather than necessarily creating entirely new ones. The response, he argued, must involve using AI defensively to identify, prioritise and fix weaknesses. Organisations may also need to abandon slow routines such as monthly patching when threats can develop much faster.
The procurement system remains another obstacle. Berntsen supports moving away from time-and-materials contracts that reward hours worked rather than clearly defined outcomes. He said there were signs of change during the previous 12 to 18 months, but most government spending remained based on hourly billing.
For Berntsen, one practical route into government AI adoption is simply demonstrating what the technology can do. At Accenture Federal Services, government leaders are shown a chat application that asks questions, gathers requirements and can produce software from a description of a problem. “When you show customers that kind of thing, they say, ‘Whoa, okay, can you do this in my environment?’” he said.
