What AI Actually Feels Like Inside Government Work

Picture two public-sector workers on the same Tuesday morning. Both have just sat down with AI-generated notes from a client meeting they conducted the day before.

The first is a telecare service manager at a small Welsh council. She asked the AI to give her a structure for a marketing strategy — a job she'd taught herself through YouTube because there's no marketing department. She read the draft, filled in the human detail, and sent it. Her team saw what she was doing and started experimenting themselves. Morale improved.

The second is a social worker in England, one of 39 interviewed by researchers at the Ada Lovelace Institute. She opened her AI-generated case notes and found the tool had substituted nonsense words for what a child said about parents fighting — "fishfingers or flies or trees," in one colleague's description. Another worker in the same study discovered that her AI summary had incorrectly indicated suicidal ideation that was never expressed. She was now responsible for correcting a record that might reach a family court.

Same category of technology. Radically different outcome. The difference isn't the tool. It's everything built around it.

An Insider Says the Quiet Part Out Loud

Before diving into what makes the difference, it's worth acknowledging the doubt most public-sector workers are already carrying — because a senior official who runs AI programs for a living is carrying it too.

David Black (a pseudonym used in Civil Service World) has spent years running AI programs across multiple UK government departments. He puts it bluntly: "AI is to some extent a solution looking for a problem. A lot of it is driven by tech companies that need something to upsell." He's specifically skeptical of Microsoft Copilot, bundled into existing Office subscriptions at roughly a 6% price increase for UK government customers. "We're lacking good evidence that the productivity benefits justify the cost."

Black's skepticism isn't anti-technology. It's a useful filter. He argues AI tends to work best in specific, bounded applications rather than as a generic productivity layer spread thinly across government. His preferred examples are tools that do something humans genuinely cannot do at scale — like DWP's Whitemail tool, which scans thousands of incoming letters to flag people at risk of harm. "The cost of a false positive is low," he says. "And the benefit of catching someone in crisis is very high." By that measure, Justice Transcribe — a secure Ministry of Justice tool that replaced manual note-taking in probation meetings — summarized more than 150,000 meetings and saved more than 25,000 staff hours between October 2025 and February 2026.

That's Black's framework in practice: bounded task, clear error cost, measurable result. For the Concerned Worker who has been told they need to "use AI" without being told how, his question is the right starting place. What is the specific task? What does an error cost? Who catches it?

AI is to some extent a solution looking for a problem. A lot of it is driven by tech companies that need something to upsell.
— David Black, Senior UK Government AI Programme Official

Those three questions apply whether you're a caseworker, a planning officer, a communications manager, or an HR analyst. And the adoption numbers make them urgent. In Q4 2025, 43% of public-sector employees used AI at least a few times a year — up from just 17% in Q2 2023. You are already inside this shift. The question is which conditions put you there.

What It Actually Looks Like When It Works

Knowing when AI is genuinely useful is one thing. Knowing what it feels like to build it into your daily work — in a resource-constrained public-sector environment — is something else entirely.

Rachel Prothero manages telecare services at the Vale of Glamorgan Council, a small local authority in South Wales. The service supports residents over 65 at risk of falling in their homes. When she moved into the role, she had no marketing department, no specialist support, and a team competing against private providers with large budgets. Every pound she spent was public money.

Before Copilot, she taught herself through books and YouTube. Research hours turned into research hours. After Copilot, those hours turned into seconds. She describes the shift precisely: she asked the tool for a marketing strategy framework, then filled in every field herself with content drawn from her professional knowledge of the service and the people using it. The AI provided structure. The human content was entirely hers.

"So ask it for a structure and make the content human," she says. "Let it take away the fear of the blank page."

Her team saw what she was doing and adopted the same approach. They became enthusiastic users. Morale improved. Customer numbers grew — though she's careful about attribution, calling AI "one of a number of contributing factors" rather than the cause. That caution is part of what makes her account credible.

Her description of AI as a team member is worth holding onto: "It only speaks when it's spoken to. It's not going to help you unless you ask for help." That's the whole lesson. The tool has no agency. The professional does.

So ask it for a structure and make the content human. Let it take away the fear of the blank page.
— Rachel Prothero, Telecare Service Manager, Vale of Glamorgan Council

Prothero's approach is repeatable anywhere in government. A housing officer scaffolding a resident-engagement strategy, a procurement analyst building a tender evaluation matrix, an HR manager drafting interview-question frameworks — all of it follows the same logic. Structure from AI, substance from the professional.

It's also worth noting that she learned on her own, without formal training. That's the norm, not the exception. Some 60% of state and local government employees received no AI training from their employer at all.

When the Conditions Slip

Prothero's story is the good version — the one where the worker controls the process. The harder question is what happens when that control slips, when AI outputs enter records or public-facing communications without adequate review.

In government, a plausible-looking AI output carries official authority the moment it enters a record. That's what makes the social-work research from the Ada Lovelace Institute alarming. Among 39 social workers across 17 local authorities, researchers found wildly inconsistent approaches to checking AI transcripts. Some workers spent up to an hour verifying the output. Others spent two minutes. One described her process as "literally just quickly screened it and then cut and paste it on to the system."

The consequences weren't theoretical. One AI summary incorrectly flagged suicidal ideation that was never expressed. Another rendered a child's account of domestic violence as nonsense words. These errors entered statutory records. The British Association of Social Workers reports disciplinary action against workers for missing obvious mistakes in AI-generated notes. The AI didn't carry the professional consequence. The worker did.

This isn't unique to social work. A planning officer whose AI-generated zoning summary contains an error that reaches a resident. A benefits processor whose AI-flagged recommendation is approved under caseload pressure. The mechanism is identical: plausible output, official channel, professional accountability, inadequate review time.

That accountability gap is measurable. A 2025 MissionSquare survey of 2,000 state and local government employees found that 28% said AI-generated content was rarely or never reviewed. If you're in that group, you're not just cutting corners — you're signing your name to whatever the model produced.

HHS federal guidance on public benefits makes the official posture explicit: "It is important to carefully design these human-in-the-loop processes to ensure that human oversight provides the intended validation of AI outputs and does not result in human actions becoming a rubber stamp without the expected scrutiny." The official expectation is that you, the worker, are responsible — whether or not you had the time to be.

Three questions determine whether you're genuinely protected or merely compliant on paper. Do you have enough time to disagree with the output? Do you have the domain knowledge to recognize an error? Do you have institutional standing to override or escalate? If the answer to any of these is no, the "human in the loop" is you in name only.

More than 75% of workers in a Roosevelt Institute-cited survey said AI had made aspects of their job harder or added to their workload in at least one way — often because the work that was supposed to be performed by computers still requires human correction and verification.

What You Can Actually Do About It This Week

Rachel Prothero described the ideal AI colleague as one who "only speaks when it's spoken to." That's not a limitation. In government work, where your name goes on the output and a real person's housing, benefits, care, or safety may depend on its accuracy, that is exactly the right job description.

The public-sector workers navigating this best are not the most enthusiastic adopters. They are the most deliberate ones — the people who have identified the narrow, low-stakes task where AI saves real time and the error cost is low, and who have held firm on everything else. That deliberateness is not resistance to change. It is professional judgment applied to a new tool, which is what public servants are paid to exercise.

This week: take ten minutes to map your recurring tasks onto two columns. Column one is structure-driven work — drafts, summaries, research scaffolding, meeting notes, document formatting — where AI produces a first pass that you verify and rewrite with your own professional knowledge. Column two is judgment-required work — eligibility, safeguarding, legal interpretation, anything that enters a statutory record or a public communication without your substantive review. Start experimenting only in column one. Keep column two human until you have both the time and the domain knowledge to catch the errors, not just the good fortune to have missed them so far.

The best AI colleague in a government office only speaks when asked. The best public servant still knows when to overrule the answer.

Recommended Tools & Resources