Peter Beckett / Oct 2026
Image: Shutterstock
Could AI agents overwhelm the bureaucracies that make exercising our rights so difficult?
I’ve been doing a fair few access to documents requests over the last few months, mainly trying to get at papers on EU nicotine policy for my website Clearing the Air. Anyone else who’s been through that system a few times will tell you that it’s a nightmare.
The Commission often ignores you, sometimes for months or years. When they do get back to you, there’s almost always an excuse to deny the good stuff, even where it’s probably not legal for them to do so. Appealing takes another month and all too often leads to having to make a complaint to the EU Ombudsman, a whole new institution.
The web interface you’re able to submit your requests and appeals through, like much of the Commission’s IT infrastructure, barely functions. At least they’ll accept a request via email: the Ombudsman won’t even talk to you unless you use their clunky online form, which requires another user account.
One journalist I spoke to calls the system a form of “weaponized incompetence”, requiring more time and patience to navigate than most professionals, let alone citizens, can muster. All too often, the result is that documents that should be subject to public scrutiny stay hidden.
Fighting back with robots
Last week, I decided to build a robot to do all this for me.
The agent will, when it’s finished, file requests by email to the Commission and the Council, count the deadline in working days against the institution's holiday calendar, and classify each reply as a grant, partial grant, refusal, "not held" or silence. A refusal, a partial grant or a missed deadline starts a draft confirmatory application, and later a draft Ombudsman complaint. It runs on open-weight models hosted in the EU, and a human (this human) approves everything it sends. It’s taken a week or so of mostly passive time to build the foundations and cost about €50. Not bad.
Following an access to documents request from start to finish manually takes a couple of days of active time, but less than an hour when automated. Initial applications to the Commission fell to 6,434 in 2025 from 6,938 in 2024; confirmatory applications, the step after a refusal, rose to 697 from 641. That indicates that the process hasn’t really been subject to automation yet, and even today the Commission struggles to get responses to applicants within the deadline.
But the Ombudsman's office handled 54% more complaints in 2025 than in 2024 and believes that AI tools suggesting the Ombudsman are part of the cause. Contesting a refusal is also where an articulate or professionally advised applicant has always had the advantage, because it means finding the case law and writing a legal argument within 15 working days. The Ombudsman has named the burden on citizens of finding and citing case law as a problem.
Weaponized incompetence or demand management?
The problem is that while access to documents is free, exercising that right isn’t. You need to know what to ask for, understand the legal framework, keep track of deadlines, find relevant case law and be prepared to argue with the institutions, sometimes for months. That takes time, and time is money. For most people, the cost of pursuing a request simply isn’t worth it, particularly when the chances of success seem slim.
In effect, administrative complexity has acted as a form of rationing. The institutions have little incentive to make the process easier when they know that many applicants will eventually give up. But once you’ve built an agent, the marginal cost of pursuing another request, or appealing another refusal, is close to zero. Suddenly, it makes sense to pursue cases that previously wouldn’t have been worth the effort. And while that’s legitimate demand that the system has historically managed to suppress, it’s not demand that the system is geared up to manage.
This got me thinking: do we risk overwhelming the system, not just in access to documents but across government, as these systems proliferate? What happens when the cost of making an administrative request falls much faster than the cost of processing it?
Are we seeing the rise of “agentic flooding”?
Evidence from elsewhere might give us a clue as to how this plays out. A recent paper on “agentic flooding” documents a sharp rise in demand for public services across a number of countries and topics since the launch of ChatGPT in 2022.
UK Housing Ombudsman complaints tripled from 2023-2025; administrative court litigation rose by around 40% in France during the same period; complaints on solicitors conduct in the UK rose by around 60%. And in the US, complaints to the Consumer Financial Protection Bureau rose from 330,000 in 2018 to 6.6 million in 2025, driven by credit reporting disputes, most likely to affect those on lower incomes with less access to professional advice.
The EU institutions have only really had to deal with AI-supported submissions at scale in one area: public consultations. I was recently involved in one and built a tool to help people complete a response to a Call for Evidence. It was a dismal failure (only about 50 people used it), but other similar tools helped get the total number of responses close to 90,000.
The Commission has long treated mass and template responses as less representative, particularly when they point to something they disagree with, but that won’t hold. It already requires you to have an EU Login to submit anything, but AI makes it increasingly difficult to distinguish between a thousand citizens independently expressing their views and a thousand responses generated by the same campaign.
And besides, there’s no reason to assume that someone using AI to write a response doesn’t genuinely believe what they’re submitting; not everyone has the time or the ability to commit their views on relatively niche topics to pixel. Just because someone isn’t a strong writer, should we be dismissing their legitimacy because they used AI to organise and draft their thoughts?
The Commission is always saying they want more citizen engagement. It’s somewhat disingenuous of them to get it and then say “well, that’s not the kind of citizen engagement we meant”.
Does agentic flooding require agentic flood defences?
Absent a revision of the 25 year old access to documents regulation (not happening any time soon), the Commission will probably need to determine that if it can’t beat the AI flood, it had better build an AI flood barrier. That might be a good thing: it’d certainly be faster. It would potentially be more consistent, but that doesn’t mean it would be more objective: a model tuned on past decisions inherits the refusals it learned from.
Applicants have an incentive to use AI to maximise disclosure. Institutions may have an incentive to use AI to minimise the administrative burden of disclosure. What if the result of all of this is an applicant’s agent and an institution’s agent exchanging increasingly elaborate legal arguments, with each side able to produce them at negligible marginal cost?
A refusal followed by an automatically generated confirmatory application, answered by an automatically generated confirmatory decision, followed by an automatically generated Ombudsman complaint. The volume of administrative activity would explode without any corresponding increase in meaningful human scrutiny.
That said, if the Commission’s agent were to be instructed to be as obstructive as possible within the law, that’s something we’d end up finding out. Most likely through some kind of access to documents request, although no doubt the Commission would try to argue that information was covered under some kind of exemption.
