ICO AI Guidance Meets the UK's Growing Complaint Burden

The Information Commissioner's Office (ICO) is defending access rights despite growing pressure from AI-written requests. Its ICO AI guidance addresses a difficult conflict: helping people communicate with government can create additional work for institutions responsible for answering them. Faster drafting does not automatically produce faster service.

The Guardian's November 8 reporting describes public bodies struggling with chatbot-generated correspondence. The policy question extends beyond whether an individual letter used AI. Institutions need to distinguish useful assistance from unnecessary complexity without treating polished writing as evidence that a complaint lacks merit.

The central tension pits easier public participation against limited institutional capacity. ChatGPT, Claude, Gemini, and Copilot belong in that discussion as alternative writing assistants, rather than competing explanations for the same administrative problem. The relevant contest is between the speed of generating a submission and the work required to resolve it responsibly.

ICO AI guidance confronts a problem beyond the inbox

A drafting tool changes the economics of submitting a request before it changes the economics of answering one.

The ICO published dedicated guidance on May 6, 2026. Its accompanying announcement described practitioner concerns about increased request volumes, complexity, and clarification work. The regulator presented its AI-generated request guidance as help applying existing principles, rather than a change in law.

That timing matters. The current debate follows a regulatory response already underway, rather than introducing a problem authorities have only just discovered. It also provides an alternative to assuming every institution needs an entirely new category of AI-specific restrictions.

Freedom of information, or FOI, provides access to recorded information held by public authorities. A service complaint asks an institution to address dissatisfaction. A statutory appeal challenges a decision through a particular procedure. Those categories sometimes overlap in correspondence, but they should remain distinct in analysis.

Calling everything a complaint obscures what employees must actually do. A request for records requires different work from an allegation of misconduct. A demand to reverse a decision raises another question entirely. An authoritative tone does not establish which process applies.

Consider an illustrative council email containing a service grievance, a request for documents, and a demand for compensation. The employee receiving it must identify those different elements. A longer version of the same email might provide useful evidence, or simply make those distinctions harder to find.

This example is hypothetical, but the operational question is concrete. Does the writing help the recipient identify an actionable issue? That is a better test than whether the text sounds professional or contains terminology usually associated with lawyers.

AI can help a person articulate a problem they previously struggled to describe. It can also expand a narrow issue into a broader argument. Those outcomes have different implications, even when both begin with the same request to improve a draft.

The distinction changes how productivity should be assessed. Time saved by the sender belongs in the calculation, but so does time spent by the recipient. An institution cannot evaluate the full effect by measuring how quickly text appears on a screen.

Nor should a public body assume that additional participation represents failure. Some previously absent requests might concern legitimate problems. The challenge is separating newly accessible demand from avoidable processing work, rather than treating every increase as unwanted traffic.

That makes the current development more consequential than an inbox nuisance. It exposes a mismatch between consumer software that accelerates expression and public systems that must interpret expression carefully. The pressure lands where language becomes an obligation to investigate, retrieve information, or explain a decision.

The pressure falls on people who must resolve the request

The bottleneck is accountable resolution, which requires more than producing a plausible response.

A public employee handling correspondence must determine what happened, which records matter, and who has authority to act. Drafting an acknowledgment is only a small part of that work. An eloquent reply cannot substitute for finding the relevant evidence.

This creates an uneven distribution of effort. A sender can generate alternative versions of an argument without investigating each added assertion. A recipient deciding whether those assertions matter must connect them to the actual case.

The resulting burden does not depend entirely on message volume. A stable number of submissions can still require more staff attention when each contains additional claims. Conversely, more submissions might remain manageable if they are focused and complete.

For administrators, that distinction should shape the questions they ask. Are employees spending longer identifying the issue? Are they checking irrelevant references? Are cases returning because an initial response addressed the wording while missing the underlying need?

These are proposed diagnostic questions, rather than findings established across UK public bodies. They matter because a count of incoming messages cannot explain where capacity disappears. Without that explanation, an institution risks purchasing a solution for the wrong stage of work.

Frontline staff face a particularly difficult judgment. They must take a potentially serious concern seriously without accepting every statement attached to it. Skepticism about an assertion should remain separate from skepticism about the person seeking assistance.

Managers face a related allocation problem. Giving one elaborate submission additional attention can delay another person's straightforward request. Responding proportionately requires understanding substance, rather than letting length determine priority.

The pressure also reaches the public. A resident who never uses a chatbot still shares the same processing queue. Any avoidable work generated elsewhere can affect their experience, even though they made no decision about AI.

There is a countervailing benefit. Someone who finds institutional language intimidating might finally submit a clear account after receiving drafting assistance. Treating that new participation as administrative waste would overlook the purpose of public access.

The appropriate objective therefore needs careful wording. Reducing correspondence is not sufficient. A successful response should reduce unnecessary handling while preserving the institution's ability to identify and resolve legitimate concerns.

That objective changes what buyers should request from software vendors. A demonstration that produces a concise reply is incomplete. The institution also needs to know whether the system helps employees reach the right conclusion with fewer corrections.

It changes staffing discussions as well. If the difficult work involves evidence and judgment, faster writing alone will not remove the constraint. Any promised savings need to survive contact with the full case, including clarification, review, and subsequent challenge.

Easier participation shifts work across the counter

An individual productivity gain can become a shared processing cost when another organization must evaluate the output.

Researchers Chris Schmitz, Lewis Hammond, and Alan Chan describe related demand surges as agentic flooding. Here, the term means AI-assisted activity that increases pressure on government services. It should not imply that every affected message comes from an autonomous agent.

Their August 2026 preprint, government service flooding, collected 84 potential cases across 11 jurisdictions. The authors describe evidence suggesting a wider phenomenon, while identifying tensions between controlling demand and preserving equitable access. Potential cases do not establish a representative measure of national prevalence.

The paper also highlights services combining complexity with financial incentives as particularly exposed in the near term. That framing helps distinguish an attractive opportunity for automated assistance from a generic increase in online communication.

The broader analytical point concerns incentives. A person submitting an appeal benefits if assistance makes their case clearer. A developer benefits if the product helps that person complete the task. Neither benefit automatically accounts for the recipient's processing costs.

This does not make the sender's objective illegitimate. It means the relevant system includes both sides of the interaction. Evaluating only the drafting experience leaves an important part of the transaction outside the frame.

The same reasoning applies across competing assistants. ChatGPT versus Claude is a meaningful product comparison for some tasks. For complaint handling, a more relevant comparison is between outputs that clarify the issue and outputs that expand it unnecessarily.

Gemini and Copilot belong under the same standard. Brand recognition does not establish whether a particular draft contains reliable facts. Neither does the length of the response or the confidence of its tone.

A restrained assistant might remove an unsupported allegation. A more accommodating one might rewrite that allegation persuasively. The latter output can feel more helpful to the sender while creating a worse document for everyone involved.

These are possible product behaviors, rather than a verified ranking of named systems. The available evidence does not justify declaring a particular vendor responsible for the UK problem. Different prompts and application designs also complicate comparisons.

For developers, this suggests a different definition of completion. A complaint-writing task is not complete merely because the application generated a formal letter. The output should preserve the user's actual concern without inventing facts or implying evidence that does not exist.

That standard is relevant to enterprise buyers too. A system designed to maximize the amount of generated text can conflict with a workflow designed to minimize ambiguity. Procurement should examine the resulting document, rather than assuming more detailed output means better assistance.

The implications extend to internal knowledge work. A well-organized AI knowledge base serves a different function from free-form drafting: it helps make supporting material retrievable. That distinction matters whenever a decision requires evidence.

The complaint itself might be short while its supporting record remains extensive. Combining everything into a single persuasive narrative can make the case harder to inspect. Keeping claims connected to their source material gives a reviewer something more useful than rhetorical polish.

The competitive opportunity therefore lies in reducing total effort across the interaction. An assistant that helps a user submit the right request might create less visible output. Its value would emerge later, through fewer misunderstandings and more direct resolution.

Suspected AI authorship cannot decide a case

Institutions need to evaluate the substance of a submission without turning writing style into an eligibility test.

The ICO's FOI and AI guidance says AI assistance does not automatically invalidate a request. Inaccurate statements also do not necessarily invalidate it. Authorities should apply the relevant requirements and assess the request itself.

The guidance identifies existing routes for clarification, excessive costs, and vexatious or repeated requests. Their application depends on the circumstances. These are not interchangeable labels for correspondence that an employee finds difficult or suspects a chatbot helped produce.

That position leaves room to challenge an incorrect assertion without dismissing the underlying information need. It also limits a tempting shortcut: treating suspected AI use as enough reason to refuse engagement.

A useful distinction separates authorship, accuracy, and merit. A human-written submission can contain an error. An AI-assisted submission can accurately describe a real problem. Knowing the drafting method does not settle whether the requested action is justified.

Formatting offers even less certainty. Headings, formal phrasing, and organized lists can appear in many kinds of writing. An institution should be cautious about turning such features into a judgment about how much attention someone deserves.

The skeptical angle also applies to claims about the scale of the problem. Staff experiences can reveal emerging difficulties, but they cannot alone establish how much additional work AI caused. Service failures, policy disputes, and changing public expectations remain possible contributors.

A crowded inbox therefore needs investigation before explanation. If an authority attributes an increase to AI, the useful follow-up concerns its evidence. What changed in the submissions, and how did that change affect handling?

The distinction between correlation and cause matters for remedies. If a service problem generates complaints, restricting drafting assistance will not fix the service. If unnecessary elaboration causes delays, clearer submission design might help without reducing access.

The same caution applies to the term flooding. It can describe a capacity problem, but it also carries associations with deliberate disruption. Legitimate individuals seeking assistance should not automatically inherit that implication.

There is a measurement challenge as well. Counting suspected AI messages can produce a confident-looking metric with uncertain foundations. An authority might learn more from measuring clarification, corrections, and handling effort than from assigning speculative authorship labels.

Those proposed measures have limitations too. Complex cases naturally require more work, whether AI appears anywhere in the process. Evaluation should account for case differences instead of rewarding teams for prioritizing the easiest submissions.

Fairness and operational efficiency converge on the same requirement: understand what the person is asking. The quickest route to that understanding will sometimes involve clarification. It will sometimes involve reading the original supporting material rather than relying on a summary.

ICO AI guidance provides a boundary for that work, but it does not establish that every local implementation is effective. The quality of institutional practice remains a separate question, requiring evidence about outcomes and the experience of people using the service.

Automated replies need evidence before expansion

Using AI on the receiving side creates another decision system whose benefits must be demonstrated.

The UK's government AI playbook, published in February 2025, emphasizes meaningful human control. It calls for validation of high-risk decisions influenced by AI, testing, live monitoring, and routes for users to trigger human review.

That provides a governance reference for proposed complaint-processing systems. It does not establish that summarization or response drafting will save time in every setting. Those claims still require evaluation within the actual workflow.

A summary can shorten a document while losing the detail that determines its significance. An employee reviewing only the shortened version might never see what disappeared. The risk concerns the relationship between the original record and the decision, rather than summary length alone.

An illustrative example makes the distinction clear. Suppose an email describes repeated delays but also includes a deadline connected to the sender's circumstances. A summary emphasizing dissatisfaction while omitting that deadline could change how the case is prioritized.

This is a hypothetical failure mode, not a documented incident from the current reporting. It shows why an apparently readable output needs a different evaluation from an accurate representation of the case.

Response drafting creates another dependency. A fluent answer can make incomplete investigation look finished. If the draft arrives before an employee establishes what happened, its structure might shape the subsequent review around an unsupported explanation.

Institutional buyers should therefore ask where the proposed saving occurs. Does the tool remove repetitive formatting? Does it help retrieve relevant material? Or does the business case assume employees will spend less time checking substantive claims?

Those categories deserve different scrutiny. Saving time on routine presentation does not require the same confidence as delegating a judgment about urgency or merit. A single productivity percentage can conceal those differences.

The comparison also needs the full process. A shorter initial handling time means little if corrections and reopened cases increase later. Equally, an implementation that preserves quality while reducing repetitive work deserves credit even if staffing levels remain unchanged.

The local government playbook, published in October 2026, adds a sector-specific governance reference. Developed through a council-led collaboration supported by the Local Government Association, it addresses responsible adoption and risk management. Its publication is evidence of preparation, rather than proof of successful complaint automation.

That distinction should shape expectations over the coming months. An announcement that a council adopted AI establishes a deployment decision. It does not establish better access, fewer mistakes, or shorter waits.

Developers building receiving-side tools face the same end-to-end test as developers helping senders. Reducing visible effort in one interface is insufficient if the product shifts work elsewhere. The value needs to survive review, correction, and escalation.

Meaningful human review also requires practical capacity. An employee cannot reliably challenge a summary if the workflow discourages opening the underlying material. Governance commitments need to connect with how the system allocates time and presents evidence.

The strongest deployment case would show where assistance helps, where it fails, and how users obtain correction. That evidence would support expansion more convincingly than another demonstration of how quickly a model can generate a professional response.

Watch outcomes, access rules, and the quality of assistance

The next test is whether institutions reduce avoidable work while keeping legitimate routes to assistance open.

The first signal to watch is operational reporting. Over the coming months, useful disclosures would connect incoming demand with handling effort, clarification, and completed cases. A rise in suspected AI writing, by itself, would provide a much weaker explanation.

If processing improves while outcomes remain sound, the argument for carefully designed assistance strengthens. If faster acknowledgments accompany more unresolved cases, the apparent productivity gain deserves closer examination. The distinction concerns results, rather than output speed.

The second signal is how public bodies change access arrangements. New forms, submission limits, or escalation policies should be examined for their actual effect. Making a queue smaller does not necessarily mean the service works better.

A focused form might help people explain what happened and what response they seek. A restrictive process might instead prevent them from supplying necessary context. Both can reduce text volume, but they represent very different outcomes.

If institutions preserve routes for clarification and individual review, that supports the possibility of managing demand without excluding legitimate users. If restrictions become the main response, the access side of the tradeoff becomes harder to dismiss.

The third signal is product behavior. Complaint-writing assistants and institutional processing tools should increasingly be assessed on factual restraint, document clarity, and connection to supporting evidence. More elaborate prose would offer little reassurance.

If products help users narrow unsupported claims and identify the actual request, they would address part of the processing problem at its source. If they mainly accelerate escalation, receiving institutions will remain under pressure to absorb the additional work.

These are observation priorities, not scheduled announcements or guaranteed developments. They provide a way to evaluate the next wave of claims without assuming every implementation succeeds or every additional request creates harm.

They also give knowledge workers a practical lens for their own organizations. The relevant question is where an AI-assisted document creates work after it leaves the author's screen. A faster draft can still be useful, but its benefit should be assessed alongside what happens next.

For public bodies, ICO AI guidance keeps the immediate boundary clear: drafting assistance alone does not determine how a request deserves to be handled. The harder task is building processes that preserve that distinction under sustained pressure.

For readers following the story, look beyond claims about overflowing inboxes or hours saved by automated writing. Ask whether institutions can show that people receive accurate decisions with less unnecessary effort. Ask whether access remains available to someone who needs help expressing a legitimate concern. Ask whether new tools make the underlying evidence easier to inspect. Those questions connect the sender's experience with the recipient's responsibilities. Before treating the next AI deployment as a productivity success, examine what happens after the message arrives: who reads it, who checks it, who can correct a mistake, and whether the person seeking help gets a useful resolution.