How this editorial can be challenged
When generating a request becomes cheap, how do we protect both the right to ask and the capacity to answer?
A sender can produce more material at little extra cost while the recipient still has to check its meaning, evidence and consequences. If that work is not reduced or funded, private convenience can become a shared queue.
Complaint volumes can rise because people previously excluded from public services finally have a usable way to exercise their rights. Treating that demand as waste would protect institutional comfort at the public's expense.
That is a reason to distinguish valuable new access from avoidable processing work, not to dismiss the cost. Systems should help people make clear, supported requests and preserve a human route to challenge an answer.
Practitioner accounts establish a problem worth investigating, not the national share of workload caused by AI. Research on agentic flooding uses an exploratory case collection and cannot establish representative prevalence or causal effect sizes. The whole-process productivity test is our proposed framework, not a measured finding.
Evidence showing more people obtaining correct, timely resolutions with lower combined effort by requesters and public staff would demonstrate genuine productivity gains—even if the number of requests rose. Evidence that concise-form requirements exclude complex or disadvantaged cases would require redesign.
The send button is not the finish line
I understand the appeal of turning a difficult letter into something clear and confident. A confusing decision, an unanswered question or a service that has let you down can already take more energy than you have. Help with the wording can be genuinely useful. Nobody should need a professional vocabulary just to be heard.
But the moment we press send, the task changes hands. Someone has to identify the question, find records, check the references and decide what an answer requires. Saving time at the keyboard is a real benefit. It is not, on its own, proof that the whole job became easier.
Count the resolution, not the pages
The ICO's guidance recognizes that AI-assisted information requests may create extra clarification work. It also makes clear that AI use does not by itself invalidate a request. Those two points belong together: a burden can be real without the person asking being in the wrong.
My proposed measure is simple: follow a case from the first question to a usable answer. Include the requester's effort, the staff time, the waiting and the corrections. If a tool makes the opening letter effortless but leaves everybody else untangling it, we have changed where work sits. We have not yet shown that we removed it.
More demand can be a sign of progress
There is a serious objection to this argument. Institutions have sometimes confused a quiet queue with a satisfied public. A system that was hard to navigate may have looked efficient because people abandoned valid questions before anyone had to answer them. Making those questions possible is an achievement, even when it costs money.
Research on agentic flooding explicitly recognizes the access benefit and the risk that quick fixes such as fees can undermine it. That research is exploratory, not a national workload audit. Still, it identifies the choice we should resist: protecting the institution from demand by making the public's rights harder to use.
Reduce the unnecessary part, not the person
The better target is work that serves neither side. A request padded with irrelevant material is harder to answer. A reply that never addresses the question may provoke another round. A website that hides basic records makes people ask for information that could already be available. These are design problems on both sides of the exchange.
A useful assistant could help someone specify the decision they are challenging, separate evidence from suspicion and check that cited material actually supports the point. A useful public service could explain what it holds and what it needs to locate it. Neither should require a long performance of legal confidence before a human problem becomes legible.
The test must include people who are easy to overlook
I would be persuaded by a pilot that records resolution quality and waiting time alongside staff effort, then checks who benefited. Did people with limited English obtain better answers? Did complicated cases get flattened into a form that could not accommodate them? Did a shorter reply resolve the issue, or merely close the ticket? Those are proposed evaluation questions, not results we already have.
We should also accept evidence that challenges the premise. If AI helps more people obtain accurate answers while reducing combined effort, call that progress. If a new concise-request rule suppresses valid claims, change the rule. The point is to measure the service people receive rather than defend a preferred position about the tool.
An empty inbox can mean two very different things
This is the paradox: a more accessible institution may initially look less efficient because more people can finally reach it. A supposedly efficient institution may look calm because the people who most need an answer have stopped trying. Volume alone cannot tell us which one we have built.
I want the kind of AI that makes it easier to ask and easier to answer. That requires caring about the person at both ends of the message. The most convincing productivity story will not end with how quickly a letter was generated. It will end with somebody receiving an answer they can understand and use.
Read the reporting
Opinion is ours. The factual record is linked below.
ICO — FOI and artificial intelligence; access rights and handling guidance Characterizing Agentic Flooding of Government Services — exploratory study, methods and limitations