Agentic AI: Unlocking autonomy through trust
Agentic AI has the potential to fundamentally change how public services work, going far beyond generative AI and the next version of a chatbot. Agents can interpret information, plan a route through a problem, interact with systems and other agents, and take action. This creates the opportunity to redesign entire workflows and organise services around citizens, rather than simply making individual tasks faster or cheaper. Whether that transformation builds or erodes trust will depend on the systems and guardrails we design around it.
Imagine agentic case working. Rather than simply summarising a case, an agent can interpret policy, define missing evidence, retrieve information and actively progress the case. It can carry out these tasks all at once, across potentially thousands of cases.
Consider a house sale – often a complicated, slow process. A series of agents could progress the land transfer: validate property and ownership information, request and complete the transfer deed, identify whether enquiries have been responded to with sufficient evidence, request missing information, and coordinate checks across authorised services and utilities. If the agents encounter conflicting information or exceptions, the case would be escalated for human review. This could massively reduce delays and frustration. However, getting it wrong has serious consequences: processing a fraudulent transaction, transferring property incorrectly, overlooking restrictions, or exposing sensitive information, each jeopardizing trust in the process and harming buyers and sellers.
A second example is to consider digital twins for public policy. AI-enabled simulations could model how policies affect different groups and places, test competing objectives, and flag when conditions move beyond the original assumptions. A director could challenge those assumptions and preserve the evidence behind significant choices. The aim is not for agents to make policy, but to give professionals a richer environment to test decisions.
This would allow policymakers to dig into potential unexpected behaviours when making high impact social policies like taxing or housing. For example, take “Eat Out to Help Out” during COVID-19. The policy was designed to stimulate demand in the hospitality sector, but research later found it created higher infection clusters. Testing these trade-offs across consumers or geographies before rollout could have helped policymakers anticipate unintended consequences, refine the policy, and improve its chances of achieving its intended outcome safely.
Finally, consider how agentic AI will change citizens’ interactions with government. As people increasingly use personal and professional agents, these will interact with organisations’ existing digital services. Eventually, accessing government may no longer involve a person navigating a website.
Returning to our housing example, the citizen moving could instruct their agent to update their address once. The agent could then identify all relevant government services and interact with authorised government agents to update the records. Suddenly, public services are organised around a person’s needs and life events, rather than expecting citizens to understand the structure and rules of government.
These are much bigger propositions than making existing processes a percentage quicker. They change the process itself, increasingly giving AI autonomy in deciding how that process unfolds. The question then becomes: what would we need to trust an agent enough to give it that autonomy?
Trusting people, process and output
There are already warning signs of what greater autonomy changes. We’ve seen news announcements about agents covertly changing their code, leaving simulated environments to steal data from government or third-party websites, or deliberately concealing lies and omissions taken to achieve their wider goals. Each of these scenarios had weak links in evaluation and design and demonstrate why testing agents and their final goals is so important.
For agentic AI, trust needs to exist across three areas: people, process, and output.
We need to trust people to provide meaningful oversight. However, there is a paradox: the more reliable an agent is, the less frequently people exercise their own judgment. Over time, this can atrophy skills, encourage automation bias, and reduce human review to merely a ‘rubber stamp’. People need the capability to recognise something is wrong, opportunities and confidence to intervene, and genuine authority to stop the system.
Second, we need to trust the process. Agents need explicit objectives and boundaries: what information they can access, what actions they can take, when they must stop or escalate, and which decisions or consequences require human approval. Clear accountability lines are essential – if an agent makes the wrong decision, responsibility for that outcome must still sit with an identifiable person. We also need a record of the route the agent took, because a successful outcome does not necessarily mean the agent got there in a successful way. Alignment is about the goal we give an agent, and whether they pursue it in a way that is consistent with our intent.
Finally, we need to trust the output. Not every agentic decision needs the same level of assurance. A reversible, low-impact decision may need little intervention, whereas one affecting fundamental rights or social services needs stronger evidence, fairness and bias checks. We need to test whether agents behave reliably and fairly under predicted and unpredictable circumstances.
This is where assurance turns trust from an assumption into evidence. We need to test whether agents succeed and how they do so – including whether they were challenged, respected boundaries, and produced reliable, fair outputs. Assurance cannot be left until the end; it must be designed alongside the agentic system from the beginning. Retrofitting assurance is costly and complicated.
The goal should not be to restrict agents until we remove the innovation that makes them valuable. It should be to earn autonomy through evidence. Start with clear boundaries, test that the agent operates safely within them, monitor what happens in practice and expand autonomy as confidence grows.
Trust and assurance are therefore not the brakes on agentic innovation. They are what give us the confidence to take our foot off the brake.
