When the Government Disappears: Christine Bellamy on Trust in the Age of AI Agents
As citizens increasingly reach public services through commercial AI assistants, the UK’s digital chief asked an unsettling question: who gets the credit, and who keeps the accountability?
The most provocative idea in this Main Stage fireside chat was not about what AI can do for government. It was about what AI might quietly take away from it.
Christine Bellamy, Interim Director General of the Digital Products Group at the UK Government Digital Service, sat down with Manuel Kilian, Author and Initiator of The Agentic State and Former Founding Managing Director Global Government Technology Centre, to talk about building citizen-centred services in an era when AI agents increasingly sit between the citizen and the state. GDS is a fourteen-year-old organization with real achievements behind it (GOV.UK, common standards, a deep user-centred design culture) but Bellamy’s framing was that it now needs to “grow up” into its next phase. And that next phase, she suggested, is probably not about websites. It may not even be about apps. It is about personalisation, reduced friction, and meeting people where they actually are.
Trust begins with whether things work
Before any of the futuristic questions, Bellamy planted a flag on first principles. Trust in government does not begin with AI, branding, or clever communication. It begins with whether services actually work, end to end. A polished digital front end means nothing if the process behind it is slow, fragmented, or opaque. She was blunt about a common failure mode: superficially digitising the front end while leaving the back office untouched, so that a citizen submits an online application and then waits weeks for a decision. That, she said, is not good enough, and governments should never confuse digital access with service success.
Tellingly, GDS measures trust as a core performance indicator, alongside uptime, usage, and cyber incidents. Trust, in this framing, is not a vague aspiration to be invoked in speeches. It is a measurable outcome to be tracked like any other.
The disintermediation problem
Then came the harder territory. Citizens, Bellamy noted, increasingly begin their journeys outside government channels, through messaging apps, general-purpose AI assistants, or commercial intermediaries, and yet still need accurate, accountable public information at the decisive moment. She offered the example of UK driving-test bookings, where scarcity and platform distortion have pushed users toward third parties and paid intermediaries, effectively turning a free public service into a marketized and less trustworthy one. The same dynamic, she warned, could spread across many services if governments do not act.
The session’s moderator pushed the scenario further: imagine a commercial AI assistant that not only books your travel but determines whether you need a visa and then applies for it on your behalf. The promise is seamlessness, the citizen states a goal, and the system handles the process. The threat is disintermediation. The citizen may never experience the government’s service, its brand, or its accountability, only the commercial interface routing the transaction. That, Bellamy argued, has implications well beyond convenience: for democratic legitimacy, for trust, and for whether citizens can even see what their government has done for them.
The right kind of friction
Her response was that governments must be deliberate about attribution and about keeping a visible, intelligible role in service delivery. She drew a sharp parallel with the media industry, which distributed its content through social platforms only to discover, too late, that the platform had captured the user’s relationship. In government, the stakes are higher. If a service works well, citizens should understand that their government delivered it.
That may mean designing in the right kind of friction not bureaucratic delay, but moments of clear consent, acknowledgment, or branded transition so that people know when they are dealing with an official government agent. She pointed to GOV.UK’s conversational agent, “Dot,” as an early attempt to maintain continuity and attribution. The principle becomes especially important in sensitive domains: taxes, earnings, benefits, anything involving rights and obligations. Her conclusion was unambiguous: governments cannot remain passive. If they do not define their place in this ecosystem, the ecosystem will define it for them.
Inclusion, handoff, and content as data
Bellamy was candid that no one yet fully understands what “good” agentic public services look like, describing internal teams as “pirates in the corner” deliberately breaking assumptions to discover future patterns. She addressed inclusion directly: GDS designs for all users, retains offline paths, and supports people through assisted channels such as post offices. Intriguingly, early experiments suggested conversational interfaces may help some people who struggle with conventional forms, letting them ask questions in a more natural, emotionally expressive way – questions, as she put it, “more about the heart than the head.”
She also flagged a challenge that will resonate with anyone running a public website: citizens increasingly get information from AI summaries rather than official pages, which means outdated content and legacy PDFs can quietly shape inaccurate AI outputs. Content, she argued, is now a form of data and must be governed by the same discipline. Her closing personal lesson, drawn from a year of navigating all this, was about organizational courage: if deep experience tells you something is right, back yourself, because too often public servants stay quiet while teams repeat the same avoidable mistakes.
Key Takeaways
- Trust starts with service performance, not branding. A digital front end that leads to a long wait erodes trust; governments must not confuse digital access with service success.
- Disintermediation is an emerging threat. As AI agents and commercial assistants mediate access to public services, governments risk losing visibility, attribution, and accountability.
- Design in the right kind of friction. Moments of clear consent, acknowledgment, and branded transition help citizens know when they are dealing with an official government service.
- Inclusion can be helped, not just hindered, by AI. Conversational interfaces may make services easier for some users who struggle with traditional forms, provided offline paths remain.
- Content is data. Outdated web pages and legacy documents can shape inaccurate AI outputs, so content governance is now a strategic concern, not an afterthought.
The GovTech 4 Impact World Congress returns in 2027. To stay connected and be the first to hear about what comes next, visit g4i-congress.com and follow us on social media. #G4I2027
