The promise was straightforward: artificial intelligence would help governments do more with less, cutting through bureaucratic backlog and saving public money. What is emerging instead is a more complicated picture. AI is not only being deployed by institutions to streamline services; it is being used by the public to engage with those institutions at a scale and speed that the systems were never designed to absorb.
The Flood No One Planned For
The analogy that explains this best is email. When electronic mail arrived, it was framed as a tool that would eliminate unnecessary meetings and reduce administrative friction. What it actually did was lower the cost of sending messages to near zero, which meant people sent far more of them, and the time saved on one end was spent managing the deluge on the other.
AI is following the same pattern, only faster and at greater scale. Former civil servant Sam Freedman has observed that governments around the world appear to be experiencing unusual surges in complaints, requests, and responses to regulatory and official bodies since the widespread adoption of tools like ChatGPT. The friction that once filtered out low-effort communication has been removed. What remains is volume.
This is not a fringe phenomenon. Employment tribunal cases in England, Scotland and Wales reportedly rose by 39% in a single year, with AI-assisted claim drafting cited as a contributing factor. Courts are now handling submissions from individuals who cannot afford legal representation but can use AI to construct arguments, sometimes citing legal precedents that do not exist. Each of those submissions still requires a human to read, assess, and disentangle, because a flawed argument may still contain a legitimate grievance underneath it.
Where the Backlog Is Actually Building
The pressure is not limited to courts. Whitehall special advisers have described government departments overwhelmed by AI-generated freedom of information requests. Members of Parliament report that their offices are receiving AI-generated constituent letters in volumes that require human review, because ignoring any one of them risks missing something genuinely important.
Public consultations, which are meant to measure what citizens actually think about proposed changes to the law, face a version of the same problem. If the responses flooding in are machine-generated rather than human-authored, the consultation loses its function as a democratic signal. The line between making civic participation more accessible and making it so frictionless that it becomes meaningless is proving thinner than anticipated.
The job market offers another illustration. Employers who adopted AI tools to screen applications at scale found that applicants responded by automating their own side of the process, sending large numbers of applications rather than targeted ones. The result is that neither party can efficiently identify a genuine match. A shortcut adopted by one side prompted a counter-shortcut on the other, and the net effect is more noise for everyone.
What This Reveals About How Efficiency Actually Works
Here is what most coverage of AI in government tends to miss: efficiency gains are not permanent properties of a technology. They are relative to context, and context changes when adoption spreads.
The article draws a useful comparison to urban traffic. Cars dramatically reduced travel time compared to horseback. But as car ownership became widespread, cities at peak hours slowed to a crawl. By 2018, daytime traffic in central London was moving at just over 7 miles per hour, roughly the speed of a trotting pony. The technology that was supposed to accelerate movement ended up, under conditions of mass adoption, reproducing the same constraints it was meant to solve.
Agentic AI, the next wave of tools designed to act autonomously on a user’s behalf, carries the same risk. Early adopters may genuinely save time and reduce friction. But if adoption reaches the scale that major technology companies are projecting, those shortcuts become the new baseline. Everyone is using them, which means no one gains a relative advantage, and the systems on the receiving end face a volume problem they were not built to handle.
This matters for how institutions, businesses, and individuals think about AI investment. Time saved in one part of a system has a tendency to reappear as cost somewhere else. A government department that uses AI to process requests faster may find that the number of requests has grown faster still. The efficiency is real but partial, and the full accounting requires looking at the whole system, not just the node that was optimized.
Historically, societies have developed mechanisms to reintroduce friction where it was removed too aggressively: spam filters for email, congestion charges for city centers. The question for AI is what the equivalent interventions will look like, and how long it will take to develop them.
In Short
AI is not simply a tool that institutions deploy. It is a tool that everyone deploys, simultaneously, including the people those institutions are meant to serve. When both sides of any interaction automate at once, the efficiency gains cancel out and the volume problem compounds. The 39% rise in employment tribunal cases is not an anomaly. It is an early signal of what happens when friction disappears from systems that depended on it to function.
Based on reporting from The Guardian.