Governments prepare for the next wave of AI

Managing agentic AI

The biggest barrier, according to Reinhardt, is not having the right data foundations in place. For agentic AI, the problem can become more pronounced because systems need access to a wide range of information and systems.

Laura Gilbert, head of the AI for Government program at the London-based Tony Blair Institute for Global Change, agreed that the problem lies with the data under the models.

"In my experience, the technology itself is rarely what holds governments back,” she said. “The real gap is in the foundations underneath them.”

She also highlighted a “capability and confidence” gap.

“Many governments have outsourced so much of their technical expertise that they struggle to judge what is realistic, what is safe and what is being oversold to them,” she added. “Nobody is sure how to answer the agentic question: when software takes an action, who is accountable for it? Civil servants lack the incentives to take risks and innovate.”

Agencies will therefore need to put in place targeted guardrails to control and mitigate the reach of AI agents without limiting usability.

In practice, that means governments should be clear about what decisions an agent can make independently, what data and systems it can access, and when it needs to hand control back to a human.

“Governments should focus on putting the data, governance and accountability structures around AI that allow it to be scaled responsibly and deliver measurable public value,” Reinhardt said. “Besides that, of course, the legacy systems need to be modernized, especially those that are not even built to be addressed by APIs and therefore by agents.”

Where government is deploying AI agents

For now, the most suitable applications may be those where agents can automate repetitive, high-volume work without directly determining outcomes that could materially affect citizens.

Reinhardt pointed to administrative workflows, document processing, case management, internal IT processes, routine workflows and approvals as potential early applications.

Gilbert made a similar case for starting where the risk is low and the volume is high. Internal processes such as handling correspondence, triaging cases or summarizing evidence are good places to learn, she said, "because mistakes are recoverable and the time savings are easy to see."

That builds the confidence and evidence base to take on more sensitive, citizen-facing work later.

"Early on, it's important to understand the difference between a two-way door, where mistakes can be undone, and a one-way door, where they can't,” she said.

But the level of risk varies considerably among use cases, and maintaining public trust is a key requirement for governments.

“AI errors that impact citizens’ lives and wellbeing damage trust in institutions and public services -- and likely will stall further rollout of this key technology, which helps administrations keep their heads above water,” Reinhardt said.

 Guardrails, therefore, need to match risk, with different levels of oversight required depending on the impact and criticality of the use case and, importantly, the potential risk to citizens.

Gilbert specified another key priority, alongside starting small and low risk: building in-house capability.

"You don't need an army, but you do need a core team of people who can build, test and challenge, and who sit close to the decisionmakers,” she said.

Scarlett Evans is a freelance writer with a focus on robotics and emerging technologies. Previously, she was assistant editor at IoT World Today, where she specialized in robotics and smart city technologies. Scarlett also has a background in the mining and resources sector, with experience at Mine Australia, Mine Technology and Power Technology.